text
stringlengths 7
4.92M
|
|---|
Serotonin and substance P colocalization in medullary projections to the nucleus tractus solitarius: dual-colour immunohistochemistry combined with retrograde tracing.
Serotonin (5HT) and substance P (SP) are colocalized in terminals within the nucleus tractus solitarius (NTS). The purpose of the present study was to determine the origin of these terminals. 5HT- and SP-immunoreactivities (IR) were visualized using dual-colour immunofluorescence histochemistry with amino-4-methylcoumarin-3-acetic acid- and fluorescein isothiocyanate-conjugated secondary antisera, while NTS-afferent neurons were visualized by retrograde labelling with rhodamine beads. Extensive colocalization of 5HT- and SP-IR was seen in NTS-afferent neurons located in the nucleus raphe pallidus, nucleus raphe obscurus, nucleus raphe magnus, and in the parapyramidal region. Over 80 per cent of the SP-IR NTS-afferent neurons contained 5HT-IR, while 68 per cent of the 5HT-IR neurons contained SP-IR. Thus, 5HT- and SP-IR are extensively colocalized in NTS-afferent neurons in the medullary raphe nuclei and associated areas of the ventral medulla. |
There has been snow on Achill’s Slievemore, there has been frost, hail and wind, bright warm sunshine and Atlantic rain. Over the fortnight since Rescue 116 crashed off north Mayo, the relatives of two missing Irish Coast Guard airmen have endured an entire weather cycle during a long and harrowing wait.
And the wait continues for the families of winch team Paul Ormsby (53) and Ciaran Smith (38). After a painstaking operation over the past two days to partially lift the helicopter wreckage from the seabed off Blackrock island, work had to be suspended on Tuesday night.
“Hugely challenging” was how Supt Tony Healy of Belmullet Garda described the conditions for Naval Service divers who had spent the last two days attaching an airbag to the fixed cabin and rotor assembly area of the wreck at a depth of 40m.
“There’s an immense flow of water at this time of the year when there are spring tides,” said Supt Healy. “That means there’s three times the amount of water flowing through that channel than there normally would be.”
Guided by a shot line from the surface, the Naval Service divers working in pairs battled different groundswells and tidal streams which greatly affected their ability to manoeuvre, according to their colleagues.
Righting the wreck
Holland 1
However, conditions proved “insurmountable”, and the airbag inflation “didn’t achieve sufficient volume”, said Irish Coast Guard operations manager Gerard O’Flynn.
The air hose attached to the air bag was “plugged” and secured on a floating mark for another attempt, possibly Wednesday.
“The work is ongoing and the objective remains the same, to right or move the wreck to carry out a visual inspection,” Mr O’Flynn said at a briefing on Blacksod pier just before 8pm.
“There is a plan to take that up again tomorrow or when the next window is available,” Insp Gary Walsh of Mayo Garda division said, adding that the weather was “looking poor” for the next two or three days.
“Moving into the weekend the intensity of the tides will drop back and the forecast is more favourable,” Mr O’Flynn said. “But a shift of wind can make a big difference out there.”
Spring tides
The RNLI all-weather Achill and inshore Sligo and Bundoran lifeboats were at sea in recent days, with gardaí and Civil Defence teams visiting the Inishkea islands to the north – also searched by local fishermen last week.
The bodies of the winch team’s two colleagues, Capt Dara Fitzpatrick (45) and Capt Mark Duffy (51), have been recovered, as has the Sikorsky S-92’s “black box” or combined flight recorders, which are being examined for data download in Britain.
Capt Duffy’s funeral takes place in his home village of Blackrock, Co Louth, on Thursday, and his family have asked that any donations be given to the RNLI and and have said that “Mark’s wish would be for you to carry an organ donor card”. |
355 F.2d 249
NATIONAL LABOR RELATIONS BOARD, Petitioner,v.INTERNATIONAL HOD CARRIERS, BUILDING AND COMMON LABORERSUNION OF AMERICA, LOCAL 894, AFL-CIO, Respondent.
No. 16294.
United States Court of Appeals Sixth Circuit.
Jan. 19, 1966.
Anthony J. Obadal, Atty., N.L.R.B., Washington, D.C. (Arnold Ordman, Gen. Counsel, dominick L. Manoli, Associate Gen. Counsel, Marcel Mallet-Prevost, Asst. Gen. Counsel, Elliott Moore, Atty., N.L.R.B., Washington, D.C., on the brief), for petitioner.
Robert E. Shuff, Akron, Ohio, for respondent.
Before WEICK, Chief Judge, CELEBREZZE, Circuit Judge, and CECIL, Senior Circuit Judge.
PER CURIAM.
1
Pursuant to Section 10(e) of the National Labor Relations Act, as amended (29 U.S.C., Section 151, et seq.) the National Labor Relations Board Seeks enforcement of its Order against respondent, International Hod Carriers, Building and Common Laborers Union of America, Local 894, AFL-CIO, reported at 148 N.L.R.B. No. 10.
2
The Board found that the respondent Union violated Sections 8(b)(2) and 8(b) (1)(A) of the National Labor Relations Act by maintaining, pursuant to arrangements with employers, a hiring system under which preference in employment was given to members of the Union. The Board found that respondent unlawfully caused an employer not to rehire one William O. Strickland because he was not a member in good standing, although Strickland was entitled to such re-employment and that thereafter, when Strickland obtained employment with another employer, respondent unlawfully demanded and secured his discharge for the same reason.
3
From an examination of the entire record, we conclude that the findings of the Board are supported by substantial evidence.
4
The petition of the Board for enforcement of its Order is sustained.
|
Wood Buffalo National Park
Wood Buffalo National Park is the largest National Park of Canada at . It is located in northeastern Alberta and the southern Northwest Territories. Larger in area than Switzerland, it is the second-largest national park in the world. The park was established in 1922 to protect the world's largest herd of free roaming wood bison, currently estimated at more than 5,000. It is one of two known nesting sites of whooping cranes.
The park ranges in elevation from at the Little Buffalo River to in the Caribou Mountains. The park headquarters is located in Fort Smith, with a smaller satellite office in Fort Chipewyan, Alberta. The park contains one of the world's largest fresh water deltas, the Peace-Athabasca Delta, formed by the Peace, Athabasca and Birch Rivers. It is also known for its karst sinkholes in the north-eastern section of the park. Alberta's largest springs (by volume, with an estimated discharge rate of eight cubic meters per second), Neon Lake Springs, are located in the Jackfish River drainage. Wood Buffalo is located directly north of the Athabasca Oil Sands.
This area was designated a UNESCO World Heritage Site in 1983 for the biological diversity of the Peace-Athabasca Delta, one of the world's largest freshwater deltas, as well as the population of wild bison.
On June 28, 2013, the Royal Astronomical Society of Canada designated Wood Buffalo National Park as Canada's newest and the world's largest dark-sky preserve. The designation helps preserve nighttime ecology for the park's large populations of bats, night hawks and owls, as well as providing opportunities for visitors to experience the northern lights.
History
Before the park
This region has been inhabited by human cultures since the end of the last ice age. Aboriginal peoples in this region have followed variations on the subarctic lifeway, based around hunting, fishing, and gathering. Situated at the junction of three major rivers used as canoe routes for trade — the Athabasca, Peace and the Slave Rivers — the region that later became the national park was well travelled for millennia.
In recorded times, the Dane-zaa (historically called the "Beaver tribe"), the Chipewyan people, the South Slavey (Dene Tha'), and Woods Cree people are known to have inhabited, and sometimes quarrelled over, the region. The Dane-zaa, Chipewyan, and South Slavey speak (or spoke) languages from the Northern Athabaskan family which is also common in the regions to the north and west of the park, and call themselves the "Dene" collectively. The Cree, by contrast, are an Algonquian people and are thought to have migrated here from the east within the timeframe of recorded history.
Sometime after 1781 when a smallpox epidemic decimated the region, the two groups made a peace treaty at Peace Point through a ceremonial pipe ceremony. This is the origin of the name of the Peace River which flows through the region: the river became the boundary with the Dane-zaa to the North and the Cree to the South.
Explorer Peter Pond is believed to have passed through the region in 1785, likely the first European to do so, followed by Alexander Mackenzie three years later. In 1788 fur trading posts were established at Fort Chipewyan just east of the current boundaries of the park and Fort Vermilion just to the west. And the Peace River, which had long been used by the First Nations as a trade route, also now also added to the growing network of canoe routes used in the North American fur trade. From the fur trade, the Métis people emerged as another major group in the region.
Canada purchased the Hudson's Bay Company's claim to the region in 1870. Agriculture was never developed in this part of Western Canada, unlike to the south; thus hunting and trapping remained the dominant industry in this region well into the twentieth century, and are still vital to many of its inhabitants. Following the Klondike Gold Rush of 1897, however, the Canadian government was keen to extinguish Aboriginal title to the land, so that any mineral wealth found in the future could be exploited despite any objections from First Nations. This led to the signing of Treaty 8 on 21 June 1899. The land then passed into the hand of the federal government as Crown land.
As a national park
Established in 1922, the park was created on Crown land acquired the territory of Treaty 8 between Canada and the local First Nations. The park itself completely surrounds several Indian reserves such as Peace Point and ʔejëre K’elnı Kuę́ (also called Hay Camp).
Between 1925 and 1928, over 6,000 plains bison were introduced to the park, where they hybridized with the local wood bison, as well as introducing bovine tuberculosis and brucellosis diseases into the herd. Parks officials have since that time attempted to undo this damage with successive culls of diseased animals. In 1957, however, a disease-free, wood bison herd of 200 was discovered near Nyarling river in Wood Buffalo National Park. In 1965, 23 of these bison were relocated to the south side of Elk Island National Park and 300 remain there today as the most genetically pure wood bison remaining. Between 1951 and 1967, 4000 bison were killed and of meat were sold from a special abattoir built at Hay Camp. These smaller culls did not eradicate the diseases, however, and in 1990 a plan was announced to cull the entire herd and restock it with undiseased animals from Elk Island National Park. This plan was abandoned due to a negative public reaction to the announcement. Since that time, wolves, the bison's main predator, have recovered in numbers due to a reduction in control efforts (mostly poisoning), reducing the size of the herd.
In 1983, a 21-year lease was granted to Canadian Forest Products Ltd. to log a 50,000-hectare area of Wood Buffalo National Park. The Canadian Parks and Wilderness Society filed a lawsuit against Parks Canada for violating the National Parks Act. Before the trial commenced in 1992, Parks Canada acquiesced and recognized that the lease was invalid and unauthorized by the provisions of the act.
In March 2019, a new provincial park, known as Kitaskino Nuwenëné Wildland Provincial Park, was established on the borders of the Wood Buffalo National Park. The protection of this park was first proposed by the Mikisew Cree First Nation and it will protect the natural ecosystems from the expanding industrial areas north of Fort McMurray. The park was created after three oil companies – Teck Resources, Cenovus Energy, and Imperial Oil – voluntarily gave up certain oilsands and mining leases in the area, following negotiations with the Alberta government and indigenous groups. This new provincial park will be closed to forestry and new energy projects, but existing wells in the area can keep producing and traditional indigenous land uses are allowed.
Climate
In the park, summers are very short, but days are long. Temperatures range between during this season. On average, summers are characterized by warm and dry days although in some years, it can have cool and wet days. The mean high in July is while the mean low is . Fall tends to have cool, windy and dry days in which the first snowfall usually occurs in October. Winters are cold with temperatures that can drop below in January and February, the coldest months. The mean high in January is while the mean low is . In spring, temperatures gradually warm up as the days become longer.
Wildlife
Wood Buffalo National Park contains a large variety of wildlife species, such as moose, bison, great grey owls, black bears, hawks, spotted owls, timber wolves, lynxes, beavers, snowy owls, marmots, bald eagles, martens, wolverines, peregrine falcons, whooping cranes, snowshoe hares, sandhill cranes, ruffed grouses, and the world's northernmost population of red-sided garter snakes, which form communal dens within the park.
Wood Buffalo Park contains the only natural nesting habitat for the endangered whooping crane. Known as Whooping Crane Summer Range, it is classified as a Ramsar site. It was identified through the International Biological Program. The range is a complex of contiguous water bodies, primarily lakes and various wetlands, such as marshes and bogs, but also includes streams and ponds.
In 2007, the world's largest beaver dam – about in length – was discovered in the park using satellite imagery; The dam, located at , about from Fort Chipewyan, had only been sighted by satellite and fixed-wing aircraft until July 2014.
Transportation
Year-round access is available to Fort Smith by road on the Mackenzie Highway, which connects to Highway 5 near Hay River, Northwest Territories. Commercial flights are available to Fort Smith and Fort Chipewyan from Edmonton. Winter access is also available using winter and ice roads from Fort McMurray through Fort Chipewyan.
Gallery
See also
List of mountains in Alberta
Buffalo National Park
List of National Parks of Canada
List of Northwest Territories parks
List of parks in Alberta
List of trails in Alberta
List of waterfalls of Alberta
National Parks of Canada
References
External links
"Aerial photos of Wood Buffalo National Park", Canadian Geographic
Park at UNESCO World Heritage Site
Great Canadian Parks
Wood Buffalo National Park, Canada (IUCN)
Category:Protected areas established in 1922
Category:World Heritage Sites in Canada
Category:Dark-sky preserves in Canada
Category:1922 establishments in Alberta |
588 F.2d 820
Hannav.U. S.
No. 78-1150
United States Court of Appeals, Third Circuit
11/17/78
1
E.D.Pa.
AFFIRMED
|
While a cold rain fell, turning what snow was left on the ground into slush, a group of about 60 Brandeis University students and faculty on Thursday demanded the school divest from fossil fuel companies as a way to address climate change.
Rallying in the school's "peace garden" outside the college's Usdan Student Center, the protesters said Brandeis should reinvest its money in socially responsible and environmentally sustainable places, and provide greater transparency about its $1 billion portfolio.
"It is unacceptable that we are still directly profiting off of these immoral and destructive industries," said Sydney Carim with Brandeis Climate Justice. "And it upsetting that we don't even know how much of our endowment is invested in fossil fuels."
The gathering at Brandeis was part Fossil Fuel Divestment Day, a national day of action to push colleges and universities divest from oil, gas and coal companies.
"This is an issue that we feel is critical to stopping the climate crisis," Carim said. "If we are going to make changes that are actually going to have an effect on the crisis that is coming our way, then we need to stop supporting and profiting off of the fossil fuel industry."
Around the country, students at more than 50 colleges and universities registered to participate in the day of action.
"What makes today so unique is that we are joining schools all around the country to call for divestment in a really organized way that we haven't really seen before," said Caleb Schwartz, one of the leaders of the Harvard University divestment movement.
"This kind of mass coordination, not just at universities but between universities, is really something that can be impactful and send a really clear message to university administrators that we're not just one group of students at one college asking for divestment, but really the voices of students around the world fighting for our future," he said.
In Massachusetts, a variety of actions were planned at about a dozen schools. At Mt. Holyoke College, about 300 students walked out of class; at MIT, students put up a big banner and distributed educational materials; and at Harvard University, students occupied an administrative building.
Given the state's longstanding reputation as a leader in higher education, student activists say, universities here have an extra responsibility to lead the call for divestment.
"MIT has a chance to make a real important impact, and if they divest, not only will it be their money that's taken out of fossil fuel companies, but this could be a domino that leads to many other universities taking the same action," said Trevor Spreadbury of the student group MIT Divest.
According to both universities, MIT's endowment is about $17.4 billion; Harvard's endowment — the largest in the country — is calculated to be about $39 billion.
In an email, Brandeis media relations Director Julie Jette said that the university has already taken actions to limit its investments in fossil fuels.
"The university has sought to balance the needs of funding financial aid, faculty salaries, and other core educational programs from its endowment with concerns about investments in fossil fuel. Letting our investments in fossil fuel private limited partnerships run off at the end of their life cycles, and suspending new investments in such vehicles, was the balance that Brandeis struck," she wrote.
But for students at Brandeis, this commitment isn't enough.
"If Brandeis is going to be a unique institution based on social justice, then it needs to be held accountable to that standard," PhD student Aneil Tripathy said. "If our goal is social justice, that has to be what is making this university unique. And it can't just be on a flyer to recruit students; it needs to be taken seriously."
Ph.D student Aneil Tripathy holds up a banner at a Brandeis demonstration. (Robin Lubbock/WBUR)
The fossil fuel divestment movement has been around for about a decade, but has gained steam in recent years, said Alyssa Lee of Divest Ed, the Cambridge-based nonprofit group helping to organize Thursday's actions.
"We have an enormous opportunity as students to really shape our institutions to make a very powerful political statement about, not just climate change, but specifically the fossil fuel industry. And divestment is one of the most powerful statements they can make," she said.
Earlier this month, Georgetown University announced it would "[freeze] new endowment investments in companies or funds whose primary business is the exploration or extraction of fossil fuels, divest from public securities of fossil fuel companies within the next five years and divest from existing private investments in those companies over the next 10 years."
And last week, the Faculty of Arts and Sciences at Harvard University voted overwhelmingly to tell the Harvard Corporation — the school's highest government body — to direct the Harvard Management Company to divest from fossil fuels.
"Divest Harvard has been going on as a campaign for eight years and we are really seeing an extreme escalation in pressure from Harvard students, faculty and alumni," said Schwartz, of Divest Harvard.
In November, students from the divestment movements at Harvard and Yale University disrupted the annual Harvard-Yale football game. |
Gilnockie Provincial Park
Gilnockie Provincial Park is a provincial park in British Columbia, Canada. This 2842-hectare park is situated southeast of Cranbrook and just north of the U.S. border. It includes the upper portion of Gilnockie Creek. Gilnockie Provincial Park protects some of the oldest fir and larch stands in the region where bears, moose, elk, white-tail and mule deer are found.
Although Gilnockie Park has low recreation values, this steep densely wooded and small wet valley encompasses wide-ranging species and habitat diversity and provides north south connectivity for many animals and birds. No facilities are provided. Visitors should be self-sufficient and proficient in backcountry travel practices.
References
Category:Provincial Parks of British Columbia
Category:Parks in the Regional District of East Kootenay
Category:Year of establishment missing |
Strain ham drippings, and heat ham stock in a small saucepan over medium heat to boiling. Add 1/2 teaspoon dry mustard, 1/4 cup brown sugar, pinches of nutmeg and ground cloves, and cook until thickened. Let your guests douse their own ham with sauce. This is a personal thing.
Advertisement
Here's the full recipe all together.
Serves 8
For cooking ham:
1/2 cup whole-grain mustard
3 Tbsp. honey
1/4 tsp nutmeg
1/4 tsp ground cloves
1 cup brown sugar
1/4 cup bourbon
1 5-8 pound half ham, butt portion, can be sliced or not
For sauce:
1 1/2 cup ham drippings or stock
1/2 tsp dry mustard
1/4 cup brown sugar
Pinches of nutmeg
Pinches ground cloves
Score ham in a diamond pattern to 1/4-inch deep (if not sliced). Place ham on rack in roasting pan.
Blend mustard, honey, nutmeg, and cloves.
Using your hands, smear mustard mixture all over surface of ham.
Again using hands, pack brown sugar all over exterior of ham, pressing to be certain it adheres.
Put bourbon in a spray bottle, and mist brown sugar coating to barely moisten. You may not use the entire 1/4 cup. Alternatively, sprinkle your bourbon over the ham.
Bake ham, uncovered, in 300-degree oven for 20 minutes per pound, basting a few times during cooking.
Strain ham drippings, skim as much fat as you wish, and add ham stock to make 1 1/2 cups. Heat in small saucepan over medium heat to boiling. Add brown sugar and spices; cook until thickened and serve over ham at table.
Food52 is a community for people who love food and cooking. Follow them at Food52.com and on Twitter @Food52. Or, get answers to your burning food questions with our new (free!) FOOD52 Hotline iPhone app. |
Where to Buy Raspberry Ketone Plus in Kuressaare Estonia?
Who hasn’t already become aware of the new Raspberry Ketone fad that has been going on in the recent years in Kuressaare Estonia. Nevertheless, unlike several other diet tablets and wonder supplements that vanish in a year or much less, raspberry ketone is still liked and utilized in Kuressaare Estonia.
This Raspberry Ketone Plus evaluation informs you concerning among the most recognized brands of this commended diet regimen supplement.
Everyone in Kuressaare Estonia wants to burn fat safely and normally, and with raspberry ketones you can. They’re scientifically verified to assist switch on the body’s all-natural mechanisms for burning fat deposits, and many people have had a bunch of success with it.
If you watch Dr. Oz you have actually viewed him talk about it, and fitness guru Lisa Lynn can not acquire enough of discussing it. Below we’re visiting talk about exactly how raspberry ketones job, the benefits and if they’re the appropriate choice for you.
click here to buy Raspberry Ketone Plus in Kuressaare Estonia
How Does Raspberry Ketone Plus Job?
What the supplement does, is it keeps your metabolism training efficiently, thus allowing you to reduce weight. Scientists state that raspberry ketones can elevate your physical body temperature level, which consequently launches much more fats that are already held and damages them down.
Raspberry ketone boosts adiponectin expression and secretion, which aids along all the metabolic procedures, regulates sugar and burns fats. Yet Raspberry Ketone Plus is not nearly raspberry ketone, there are several other active ingredients that give your weight loss an actual boost.
The materials introduce in this kind of ketone surges the production of a protein called adiponectin in the liver, assisting trigger the physical body’s organic fat deposits burning processes.
This is normally pitched in throughout malnourishment, large bursts of exercise, whenever when the body needs energy to endure. It likewise aids with managing sugar; this is just what metabolizes fatty tissue.
So if you manage to activate these sort of metabolic procedures you’ll have the ability to drop weight without significant lifestyle or diet regimen changes.
click here to buy Raspberry Ketone Plus in Kuressaare Estonia
Raspberry Ketone Plus Contents
Raspberry Ketones are the main ingredient. They produce adiponectin, which is an organic bodily hormone that reduces body weight by tweaking how your physical body takes in glucose and breaks down fatty tissue cells.
Kelp is a seaweed rich in iodine which sustains healthy and balanced feature of the thyroid. When your thyroid is not operationing properly, you will end up being over weight.
Caffeine is understood to enhance metabolism and offer you power.
Grapefruit extract has actually been made use of for weight loss ever since the 1930s. It increases metabolic process and activates thermogenic effect in your physical body, burning extra calories normally.
Resveratrol from red and dark grapes is a terrific antioxidant which regulates blood sugar. Close to benefits it has on weight loss, it also has anti-aging homes.
As you could view from the above, all the very best natural active ingredients that improve weight loss have been included in one solitary supplement, making it more powerful compared to simple raspberry ketone capsules from several other brands.
click here to buy Raspberry Ketone Plus in Kuressaare Estonia
Exactly how Does Raspberry Ketone Plus Train?
This is not a very easy question to respond to, although that’s exactly what you appreciate one of the most. Results are extremely individual in Kuressaare Estonia, nevertheless, if you consist of healthy diet regimen and a lot of exercise with your Raspberry Ketone Plus program, you will certainly view outcomes for certain.
Baseding on reviews and comments on their official website, people have lost anywhere in between 1 and 5 lbs in a week. Besides weight loss, Raspberry Ketone Plus also keeps the skin healthy and balanced and lessens dimpled skin.
Your digestion will not trouble you any longer and you will really feel much better overall. Similar to any diet supplement you must be aiming to diet at least 2 to 3 months to see excellent results. Raspberry Ketone Plus is not a miracle product what makes you thin in 3 weeks.
We suggest you diet a minimum of 3-4 months to view excellent results with Raspberry Ketones, and with existing cost savings it would cost you just $ 69.95, and you’ll get a FREE bottle of CLA (potent antioxidant) also.
Raspberry Ketone Plus benefits:
Burns fatty tissue
100 % natural
Accelerate metabolic rate
Aids healthy and balanced nutrients and minerals soak up in the physical body
Moderates cholesterol and blood sugar level
Provides electricity, stamina and vigor
Boosts fatty tissue oxidation
Every one of that for merely $ 19 a month
click here to buy Raspberry Ketone Plus in Kuressaare Estonia
How to Take Raspberry Ketone Plus
One Raspberry Ketone Plus container has 60 pills. Take one capsule before morning meal and one more one prior to lunch. In this manner you eat healthy 200 mg of raspberry ketone daily, similar to advised by medical professionals in Kuressaare Estonia.
If you have issues sleeping, do not take Raspberry Ketone Plus in the late afternoon or evening as it has caffeine. Don’t go beyond the amount.
Also, expecting or breastfeeding women, youngsters and individuals with different medical health conditions will talk to their medical professional before taking Raspberry Ketone Plus. Nevertheless, with typical use there will certainly be no side effects.
Side Effects of Raspberry Ketone Plus?
Raspberry ketones happen in attributes and countless individuals have actually reported success without any kind of side effects. The benefits stacked up beside other kinds of weight loss devices reveal this type of ketone to be secure, organic and efficient with no side effects.
click here to buy Raspberry Ketone Plus in Kuressaare Estonia
Is it Right for You?
Raspberry Ketones provide you a means to achieve your weight loss objectives at your own speed. You could tweak your diet and include in additional activity to obtain more weight loss, or you could just slim down without stressing.
You’ll be able to try it risk free for a week and view if it works for you; with a 100 % refund satisfaction assurance you’ll have the ability to get the benefits without taking all the danger. All you need to lose is the weight, so check it out today!
Where To Buy Raspberry Ketone Plus
Raspberry Ketone Plus is available just from the Advancement Slendering warehouse store (official internet site right here) and is routed from the UK. Next day taped is distribution is merely $ 1.99 or you can decide on Second class delivery totally free.
Raspberry Ketone Plus is presently on sale and a container costs just $ 19 (as opposed to typical $ 39.95. For also larger discount rates on the item cost and shipping, buy 2 or more containers.
Raspberry Ketone Plus comes with a remarkable list of substances, created to work out even on one of the most stubborn added weight by increasing your metabolic process and urging your physical body to break down kept fats.
If you compare the item to many various other weight loss supplements, it is extremely affordable, effectively known and totally secure to take. And with really inexpensive $ 1.99 1st Class Recorded delivery Raspberry Ketone Plus ticks all our boxes. |
Pages
Saturday, March 22, 2014
Another Busy Week
Last week was another busy one, so many things to do I
wondered how I would get it all done. Then, I remembered that I could only do
what I could and leave the rest. Thankfully though, I got it all done with a
little time to spare.
I ended the week by going to the doctor for a check-up on
Friday afternoon, had to make sure the old blood pressure was still under
control. It was. Everything checked out fine. I left the doctor’s office with
another handful of prescriptions for other wellness tests.
They’re sending me out for the dreaded old mammogram,
bone density study, and a few others. I have to get these done by the time I go
back in three months. Grrrrrrr!!!
If I sound like I’m complaining, I’m not. I just dread
all the poking and prodding and such, sometimes it hurts, for real. I’m
counting my blessings though, at least I am still above the ground and able to
go get these tests done.
Today turned out to be a pretty good day even with the
overcast skies and no sunshine all day long. I’m just trying now to figure out
what I want to eat, I’m getting hungry. |
Wokingham Choral Society
Monteverdi Vespers (1610) in the Great Hall, University of Reading on March 31st, 2012.
Fauré Requiem, in St Paul's Church, Wokingham on 9th March, 2013.
Emily Vine
Soprano
Born in Surrey, Emily's formative years as a singer began as a member of Farnham Youth Choir. She went on to read music
at the University of Bristol, winning the Sir Thomas Beecham scholarship for outstanding performance, and was awarded a
place at the Royal Academy of Music, where she now studies on the Preparatory Opera course with Elizabeth Ritchie and
Iain Ledingham. Recently Emily was delighted to accept an offer to join Royal Academy Opera, commencing in September.
Emily's opera roles include Barbarina for Amersham Festival Opera with Iain Ledingham, Miles (Turn of the Screw)
with Gergely Kaposi of Hungarian State Opera in Budapest, and Susanna and Donna Anna for Bristol University Operatic Society.
Opera scenes include Calisto, Clorinda (La Cenerentola), Giannetta (L'elisir d'Amore) and Lucia
(The Rape of Lucretia). Since joining the Academy, she has been highly commended in the Isabel Jay Opera Prize
and the Michael Head Prize.
Concert performances have recently included the Brahms, Fauré and Rutter Requiems and Bach's
Christmas Oratorio in Kristiansand, Norway. This year Emily will be a soloist in the flagship Royal Academy
of Music/Kohn Foundation Bach Cantata series, and in March will sing for Laurence Cummings in the London Handel Festival. |
How To Send Anonymous Email 2017 By SSTecTutorials
Send Anonymous Email From Your Desire Email Address And Name With Attachments Using Your Kali Linux Browser Or Any Browser.
What Is Anonymous Email?
Anonymous Email Is Email In Which The Sender's Address And Personal Identifying Information Cannot Be Viewed By The Recipient.
Anonymous Emails Are Designed So That The Email Recipient Will Remain Unaware Of The Sender's IdentitySend Anonymous Email :-https://anonymousemail.me/ |
import {CUSTOM_ELEMENTS_SCHEMA} from '@angular/core';
import {async, ComponentFixture, TestBed} from '@angular/core/testing';
import {ComponentThree} from './component-three.component';
describe('Component: ComponentThree', () => {
let fixture: ComponentFixture<ComponentThree>;
let component: ComponentThree;
beforeEach(async(() => {
TestBed.configureTestingModule({
declarations: [ComponentThree],
schemas: [CUSTOM_ELEMENTS_SCHEMA]
}).compileComponents();
}));
beforeEach(() => {
fixture = TestBed.createComponent(ComponentThree);
component = fixture.componentInstance;
});
it('should create an instance', () => {
expect(component).toBeTruthy();
});
});
|
Users of indicia reading apparatuses such as bar code reading apparatuses have always rated as an important factor in determining overall satisfaction with an apparatus the “snappiness” of operation—how fast a decoded message is output after reading is initiated. The time to output a decoded message after receipt of a trigger signal can be referred to as the trigger-to-read (TTR) time.
In order achieve snappiness of operation, designers of reading apparatuses have implemented designs wherein several frames of image data are captured and subjected to processing in succession one after another over a short time period. If processing of a first frame to be subject to a decode attempt fails, another captured frame is processed, and then another until an indicia is successfully decoded. While a succession of frames are being captured and subject to decoding, a user may be moving the apparatus (which may be hand held) into a position wherein a higher quality image may be captured.
Providing an apparatus which repeatedly captures and attempts to decode images has significant advantages. However, challenges continue to be noted with presently available indicia reading apparatuses. Some of the challenges faced by designers of indicia reading apparatuses have been imposed by technological advances.
For example, with advances made in circuitry and software design, including those by the assignee Hand Held Products, Inc. reading apparatuses are now capable of reading indicia formed on substrates at increasingly long range reading distances. At longer reading distances, fewer light rays projected by an on board lighting assembly of a reading apparatus (where present) are able to reach and be reflected from a target substrate. Because of the increased depth of field available with currently available reading apparatuses such as the IT4XXX imaging module, poor illumination reading conditions are more commonly encountered. For battery conservation purposes and for cost purposes, it has been a goal of designers or reading apparatuses to decode indicia such as bar codes with little even no artificial illumination.
In addition, with respect to image sensor based reading apparatuses, image sensors continue to grow in density. Fabrication technologies exists for making high density (e.g., million plus pixel) image sensors at low cost. Such image sensors generate more image data, which consumes additional processing time.
There remains a need to read bar codes and other decodable indicia quickly in normal operating conditions and in an expanding range of operating conditions. |
Shout It Out
Shout It Out may refer to:
"Shout It Out" (Alisa Mizuki song)
"Shout It Out" (BoA song)
Shout It Out (Elli Erl album)
Shout It Out (Hanson album)
"Shout It Out" (Kingpin song)
Shout It Out (Patrice Rushen album)
"Shout It Out" (Reece Mastin song)
"Shout It Out" (Shotgun Messiah song)
"Shout It Out", a 2010 song by Marc Mysterio
"Shout It Out", the theme song from The Wendy Williams Show
See also
Shout It Out Loud (disambiguation)
Shout (disambiguation) |
// license:BSD-3-Clause
// copyright-holders:David Haywood
#ifndef MAME_VIDEO_ELAN_EU3A05VID_H
#define MAME_VIDEO_ELAN_EU3A05VID_H
#include "elan_eu3a05commonvid.h"
#include "cpu/m6502/m6502.h"
#include "machine/bankdev.h"
class elan_eu3a05vid_device : public elan_eu3a05commonvid_device, public device_memory_interface
{
public:
elan_eu3a05vid_device(const machine_config &mconfig, const char *tag, device_t *owner, uint32_t clock);
template <typename T> void set_cpu(T &&tag) { m_cpu.set_tag(std::forward<T>(tag)); }
template <typename T> void set_addrbank(T &&tag) { m_bank.set_tag(std::forward<T>(tag)); }
void map(address_map& map);
uint32_t screen_update(screen_device &screen, bitmap_ind16 &bitmap, const rectangle &cliprect);
void set_is_sudoku();
void set_is_pvmilfin();
void set_use_spritepages() { m_use_spritepages = true; };
protected:
// device-level overrides
virtual void device_start() override;
virtual void device_reset() override;
virtual space_config_vector memory_space_config() const override;
private:
required_device<m6502_device> m_cpu;
required_device<address_map_bank_device> m_bank;
const address_space_config m_space_config;
uint8_t m_vidctrl;
uint8_t m_tile_gfxbase_lo_data;
uint8_t m_tile_gfxbase_hi_data;
uint8_t m_sprite_gfxbase_lo_data;
uint8_t m_sprite_gfxbase_hi_data;
uint8_t m_tile_scroll[4*2];
uint8_t m_splitpos[2];
uint16_t get_scroll(int which);
bool get_tile_data(int base, int drawpri, int& tile, int &attr, int &unk2);
void draw_tilemaps(screen_device &screen, bitmap_ind16 &bitmap, const rectangle &cliprect, int drawpri);
void draw_sprites(screen_device &screen, bitmap_ind16 &bitmap, const rectangle &cliprect);
uint8_t read_spriteram(int offset);
uint8_t read_vram(int offset);
// VIDEO
// tile bases
void tile_gfxbase_lo_w(uint8_t data);
void tile_gfxbase_hi_w(uint8_t data);
uint8_t tile_gfxbase_lo_r();
uint8_t tile_gfxbase_hi_r();
// sprite tile bases
void sprite_gfxbase_lo_w(uint8_t data);
void sprite_gfxbase_hi_w(uint8_t data);
uint8_t sprite_gfxbase_lo_r();
uint8_t sprite_gfxbase_hi_r();
uint8_t elan_eu3a05_vidctrl_r();
void elan_eu3a05_vidctrl_w(uint8_t data);
uint8_t tile_scroll_r(offs_t offset);
void tile_scroll_w(offs_t offset, uint8_t data);
uint8_t splitpos_r(offs_t offset);
void splitpos_w(offs_t offset, uint8_t data);
uint8_t read_unmapped(offs_t offset);
void write_unmapped(offs_t offset, uint8_t data);
int m_bytes_per_tile_entry;
int m_vrambase;
int m_spritebase;
bool m_use_spritepages;
};
DECLARE_DEVICE_TYPE(ELAN_EU3A05_VID, elan_eu3a05vid_device)
#endif // MAME_VIDEO_ELAN_EU3A05VID_H
|
COPYTRACK-The Future of Global Copyright Registration
INTRODUCTION
Located in the heart of Berlin, COPYTRACK is currently a fast-growing business. Over the course of the last few years, it has become a leading platform and service provider for image search and copyright enforcement worldwide. With its cutting edge technology and processes, it is well-suited to globally address the key challenges in the industry. At the core of COPYTRACK is the creation of a Global Decentralized Copyright Register for digital content, which authenticates users and links digital intellectual property. This online registry will generate a unique ecosystem for rights-holders, thereby providing new efficient marketplaces
TECHNOLOGY AND MARKET
Competition exists within nearly all industries, and services for rights-holders is no exception. They main competitors include image search engines and image-matching providers, companies that offer post-licensing services, businesses that perform collection, and even lawyers who enforce rights. It is worth noting that these services are region dependent and not available worldwide. The following figure compares other businesses, including their services and areas of operation, to COPYTRACK. They strive to be innovative and utilize the newest technologies to they advantage. They dedicated technical team currently ensure the integrity of 8 different processes.
1. High-Performance Web-Crawler they High-Performance Web-Crawler searches millions of websites worldwide every day.
2. Unique image matching they unique image matching engine compares our customer’s images with all findings. Cropping, changes and editing will be recognized and considered.
PROCESSES
COPYTRACK has developed a unique startto-finish process for the enforcement of copyright, once the rights-holder has identified the unlicensed use of their images. The whole process is highly automated, only requiring manual assessment at 2 of the nearly 50 stages. The following diagram illustrates the process for they customers. This mechanism can be broken down into 3 simple steps: 1. Image Upload • Upload images directly or via API • Create collections and select categories • The crawling process kicks off automatically and runs constantly 2. Select Hits • Mark illegal images among the search results • Filter already licensed pictures • Option to whitelist whole domains or a single hit • 3. Lean Back • Assign a royalty fee • Submit the case and the post-licensing process starts • they take care of the rest.
Token Distribution
A total of 60% of all tokens will be available for purchase by the public / community during the COPYTRACK Initial Token Sale by the Distributor. During the Pre-Sale, the distributor will sell one-third of these tokens, i.e. 20% of the total amount, for a discounted price. The other two-thirds, or 40% of the supply, will be released in the public sale to people on the COPYTRACK Whitelist. At the end of the token sale, all remaining tokens will be burned. As outlined in they roadmap,they will initially launch an ERC-20 token on the Ethereum blockchain. In Q2 2018, they will perform a token swap of the 100.000.000 ERC-20 tokens onto they new chain. At this point, node operators will be eligible to receive compensation in CPY for securing the network using they Proof-of-Stake consensus model. |
---
abstract: 'BERT-based architectures currently give state-of-the-art performance on many NLP tasks, but little is known about the exact mechanisms that contribute to its success. In the current work, we focus on the interpretation of self-attention, which is one of the fundamental underlying components of BERT. Using a subset of GLUE tasks and a set of handcrafted features-of-interest, we propose the methodology and carry out a qualitative and quantitative analysis of the information encoded by the individual BERT’s heads. Our findings suggest that there is a limited set of attention patterns that are repeated across different heads, indicating the overall model overparametrization. While different heads consistently use the same attention patterns, they have varying impact on performance across different tasks. We show that manually disabling attention in certain heads leads to a performance improvement over the regular fine-tuned BERT models.'
author:
- |
Olga Kovaleva, Alexey Romanov, Anna Rogers, Anna Rumshisky\
Department of Computer Science\
University of Massachusetts Lowell\
Lowell, MA 01854\
[{okovaleva,arum,aromanov}@cs.uml.edu]{}
bibliography:
- 'emnlp-ijcnlp-2019.bib'
title: Revealing the Dark Secrets of BERT
---
Introduction
============
Over the past year, models based on the Transformer architecture [@vaswani2017attention] have become the de-facto standard for state-of-the-art performance on many natural language processing (NLP) tasks [@radford2018improving; @devlin2018bert]. Their key feature is the self-attention mechanism that provides an alternative to conventionally used recurrent neural networks (RNN). One of the most popular Transformer-based models is BERT, which learns text representations using a bi-directional Transformer encoder pre-trained on the language modeling task [@devlin2018bert]. BERT-based architectures have produced new state-of-the-art performance on a range of NLP tasks of different nature, domain, and complexity, including question answering, sequence tagging, sentiment analysis, and inference. State-of-the-art performance is usually obtained by fine-tuning the pre-trained model on the specific task. In particular, BERT-based models are currently dominating the leaderboards for SQuAD[^1] [@rajpurkar2016squad] and GLUE benchmarks[^2] [@wang2018glue].
However, the exact mechanisms that contribute to the BERT’s outstanding performance still remain unclear. We address this problem through selecting a set of linguistic features of interest and conducting a series of experiments that aim to provide insights about how well these features are captured by BERT. This paper makes the following contributions:
- We propose the methodology and offer the first detailed analysis of BERT’s capacity to capture different kinds of linguistic information by encoding it in its self-attention weights.
- We present the evidence of BERT’s overparametrization and suggest a counter-intuitive yet frustratingly simple way of improving its performance, showing absolute gains of up to 3.2%.
Related work
============
There have been several recent attempts to assess BERT’s ability to capture structural properties of language. @goldberg2019assessing demonstrated that BERT consistently assigns higher scores to the correct verb forms as opposed to the incorrect one in a masked language modeling task, suggesting some ability to model subject-verb agreement. @jawahar:hal-02131630 extended this work to using multiple layers and tasks, supporting the claim that BERT’s intermediate layers capture rich linguistic information. On the other hand, @tran2018importance concluded that LSTMs generalize to longer sequences better, and are more robust with respect to agreement distractors, compared to Transformers.
@liu2019linguistic investigated the transferability of contextualized word representations to a number of probing tasks requiring linguistic knowledge. Their findings suggest that (a) the middle layers of Transformer-based architectures are the most transferable to other tasks, and (b) higher layers of Transformers are not as task specific as the ones of RNNs. @tang2018self argued that models using self-attention outperform CNN- and RNN-based models on a word sense disambiguation task due to their ability to extract semantic features from text.
Our work contributes to the above discussion, but rather than examining representations extracted from different layers, we focus on the understanding of the self-attention mechanism itself, since it is the key feature of Transformer-based models.
Another research direction that is relevant to our work is neural network pruning. @frankle2018lottery showed that widely used complex architectures suffer from overparameterization, and can be significantly reduced in size without a loss in performance. @goldberg2019assessing observed that the smaller version of BERT achieves better scores on a number of syntax-testing experiments than the larger one. @adhikari2019rethinking questioned the necessity of computation-heavy neural networks, proving that a simple yet carefully tuned BiLSTM without attention achieves the best or at least competitive results compared to more complex architectures on the document classification task. @wu2019pay presented more evidence of unnecessary complexity of the self-attention mechanism, and proposed a more lightweight and scalable dynamic convolution-based architecture that outperforms the self-attention baseline. These studies suggest a potential direction for future research, and are in good accordance with our observations.
Methodology
===========
We pose the following research questions:
1. What are the common attention patterns, how do they change during fine-tuning, and how does that impact the performance on a given task? (Sec. \[sec:patterns\], \[sec:fine-tuning\])
2. What linguistic knowledge is encoded in self-attention weights of the fine-tuned models and what portion of it comes from the pre-trained BERT? (Sec. \[sec:fn\], \[sec:vert\_attention\], \[sec:cross\_attention\])
3. How different are the self-attention patterns of different heads, and how important are they for a given task? (Sec. \[sec:disabling\])
The answers to these questions come from a series of experiments with the basic pre-trained or the fine-tuned BERT models, as will be discussed below. All the experiments with the pre-trained BERT were conducted using the model provided with the PyTorch implementation of BERT (bert-base-uncased, 12-layer, 768-hidden, 12-heads, 110M parameters)[^3]. We chose this smaller version of BERT because it shows competitive, if not better, performance while having fewer layers and heads, which makes it more interpretable.
We use the following subset of GLUE tasks [@wang2018glue] for fine-tuning:
- *MRPC*: the Microsoft Research Paraphrase Corpus [@dolan2005automatically]
- *STS-B*: the Semantic Textual Similarity Benchmark [@cer2017semeval]
- *SST-2*: the Stanford Sentiment Treebank, two-way classification [@socher2013recursive]
- *QQP*: the Quora Question Pairs dataset
- *RTE*: the Recognizing Textual Entailment datasets
- *QNLI*: Question-answering NLI based on the Stanford Question Answering Dataset [@rajpurkar2016squad]
- *MNLI*: the Multi-Genre Natural Language Inference Corpus, matched section [@williams2018broad]
Please refer to the original GLUE paper for details on the QQP and RTE datasets [@wang2018glue]. We excluded two tasks: CoLa and the Winograd Schema Challenge. The latter is excluded due to the small size of the dataset. As for CoLa (the task of predicting linguistic acceptability judgments), GLUE authors report that the human performance is only 66.4, which is explained by the problems with the underlying methodology [@Schutze_1996_Empirical_Base_of_Linguistics_Grammaticality_Judgments_and_Linguistic_Methodology]. Note also that CoLa is not included in the upcoming version of GLUE [@WangPruksachatkunEtAl_2019_SuperGLUE_Stickier_Benchmark_for_General-Purpose_Language_Understanding_Systems]. All fine-tuning experiments follow the parameters reported in the original study (a batch size of 32 and 3 epochs, see ).
In all these experiments, for a given input, we extract self-attention weights for each head in every layer. This results in a 2D float array of shape $L\times L$, where $L$ is the length of an input sequence. We will refer to such arrays as *self-attention maps*. Analysis of individual self-attention maps allows us to determine which target tokens are attended to the most as the input is processed token by token. We use these experiments to analyze how BERT processes different kinds of linguistic information, including the processing of different parts of speech (nouns, pronouns, and verbs), syntactic roles (objects, subjects), semantic relations, and negation tokens.
{width="\linewidth"}
Experiments
===========
In this section, we present the experiments conducted to address the above research questions.
BERT’s self-attention patterns {#sec:patterns}
------------------------------
Manual inspection of self-attention maps for both basic pre-trained and fine-tuned BERT models suggested that there is a limited set of self-attention maps types that are repeatedly encoded across different heads. Consistently with previous observations[^4], we identified five frequently occurring patterns, examples of which are shown in :
- *Vertical*: mainly corresponds to attention to special BERT tokens *\[CLS\]* and *\[SEP\]*;
- *Diagonal*: formed by the attention to the previous/following tokens;
- *Vertical+Diagonal*: a mix of the previous two types,
- *Block*: intra-sentence attention for the tasks with two distinct sentences (such as, for example, RTE or MRPC),
- *Heterogeneous*: highly variable depending on the specific input and cannot be characterized by a distinct structure.
Whereas the attention to the special tokens is important for cross-sentence reasoning, and the attention to the previous/following token comes from language model pre-training, we hypothesize that the last of the listed types is more likely to capture interpretable linguistic features, necessary for language understanding.
To get a rough estimate of the percentage of attention heads that may capture linguistically interpretable information, we manually annotated around 400 sample self-attention maps as belonging to one of the five classes. The self-attention maps were obtained by feeding random input examples from selected tasks into the corresponding fine-tuned BERT model. This produced a somewhat unbalanced dataset, in which the “Vertical” class accounted for 30% of all samples. We then trained a convolutional neural network with 8 convolutional layers and ReLU activation functions to classify input maps into one of these classes. This model achieved the F1 score of 0.86 on the annotated dataset. We used this classifier to estimate the proportion of different self-attention patterns for the target GLUE tasks using up to 1000 examples (where available) from each validation set.
#### Results {#results .unnumbered}
shows that the self-attention map types described above are consistently repeated across different heads and tasks. While a large portion of encoded information corresponds to attention to the previous/following token, to the special tokens, or a mixture of the two (the first three classes), the estimated upper bound on all heads in the “Heterogeneous” category (i.e. the ones that *could* be informative) varies from 32% (MRPC) to 61% (QQP) depending on the task.
We would like to emphasize that this only gives the upper bound on the percentage of attention heads that could potentially capture meaningful structural information beyond adjacency and separator tokens.
![Estimated percentages of the identified self-attention classes for each of the selected GLUE tasks. []{data-label="fig:attention_by_dataset"}](images/attention_by_dataset.png){width="\linewidth"}
Relation-specific heads in BERT {#sec:fn}
-------------------------------
In this experiment, our goal was to understand whether different syntactic and semantic relations are captured by self-attention patterns. While a large number of such relations could be investigated, we chose to examine semantic role relations defined in frame semantics, since they can be viewed as being at the intersection of syntax and semantics. Specifically, we focused on whether BERT captures FrameNet’s relations between frame-evoking lexical units (predicates) and core frame elements [@baker1998berkeley], and whether the links between them produce higher attention weights in certain specific heads. We used pre-trained BERT in these experiments.
{width="0.9\linewidth"}
The data for this experiment comes from FrameNet [@baker1998berkeley], a database that contains frame annotations for example sentences for different lexical units. Frame elements correspond to semantic roles for a given frame, for example, “buyer", “seller", and “goods” for the “Commercial\_transaction" frame evoked by the words “sell” and “spend” or “topic” and “text” for the “Scrutiny” semantic frame evoked by the verb “address”. shows an example of such annotation.
We extracted sample sentences for every lexical unit in the database and identified the corresponding core frame elements. Annotated elements in FrameNet may be rather long, so we considered only the sentences with frame elements of 3 tokens or less. Since each sentences is annotated only for one frame, semantic links from other frames can exist between unmarked elements. We therefore filter out all the sentences longer than 12 tokens, since shorter sentences are less likely to evoke multiple frames. To establish whether BERT attention captures semantic relations that *do not* simply correspond to the previous/following token, we exclude sentences where the linked objects are less than two tokens apart. This leaves us with 473 annotated sentences.
![FrameNet annotation example for the “address” lexical unit with two core frame elements of different types annotated.[]{data-label="fig:framenet"}](images/framenet.pdf){width="0.8\linewidth"}
For each of these sentences, we obtain pre-trained BERT’s attention weights for each of the 144 heads. For every head, we return the maximum absolute attention weight among those token pairs that correspond to the annotated semantic link contained within a given sentence. We then average the derived scores over all the collected examples. This strategy allows us to identify the heads that prioritize the features correlated with frame-semantic relations within a sentence.
#### Results {#results-1 .unnumbered}
The heatmap of averaged attention scores over all collected examples () suggests that 2 out of 144 heads tend to attend to the parts of the sentence that FrameNet annotators identified as core elements of the same frame. shows an example of this attention pattern for these two heads. Both show high attention weight for “he” while processing “agitated” in the sentence “He was becoming agitated" (the frame “Emotion\_directed”).
{width="\linewidth"}
[max width=1]{}
-------- -------- ------------------- ----------------- --------
**normal distr.** **pre-trained**
MRPC 0/31.6 81.2/68.3 87.9/82.3 F1/Acc
STS-B 33.1 2.9 82.7 Acc
SST-2 49.1 80.5 92 Acc
QQP 0/60.9 0/63.2 65.2/78.6 F1/Acc
RTE 52.7 52.7 64.6 Acc
QNLI 52.8 49.5 84.4 Acc
MNLI-m 31.7 61.0 78.6 Acc
-------- -------- ------------------- ----------------- --------
: GLUE task performance of BERT models with different initialization. We report the scores on the validation, rather than test data, so these results differ from the original BERT paper.[]{data-label="tab:glue-results"}
Change in self-attention patterns after fine-tuning {#sec:fine-tuning}
---------------------------------------------------
Fine-tuning has a huge effect on performance, and this section attempts to find out why. To study how attention per head changes on average for each of the target GLUE tasks, we calculate cosine similarity between pre-trained and fine-tuned BERT’s flattened arrays of attention weights. We average the derived similarities over all the development set examples[^5]. To evaluate contribution of pre-trained BERT to overall performance on the tasks, we consider two configurations of weights initialization, namely, pre-trained BERT weights and weights randomly sampled from normal distribution.
{width="\linewidth"}
{width="\linewidth"}
#### Results {#results-2 .unnumbered}
shows that for all the tasks except QQP, it is the last two layers that undergo the largest changes compared to the pre-trained BERT model. At the same time, shows that fine-tuned BERT outperforms pre-trained BERT by a significant margin on all the tasks (with an average of 35.9 points of absolute difference). This leads us to conclude that the last two layers encode task-specific features that are attributed to the gain of scores, while earlier layers capture more fundamental and low-level information used in fine-tuned models. Randomly initialized BERT consistently produces lower scores than the ones achieved with pre-trained BERT. In fact, for some tasks (STS-B and QNLI), initialization with random weights gives worse performance that that of pre-trained BERT alone without fine-tuning. This suggests that pre-trained BERT does indeed contain linguistic knowledge that is helpful for solving these GLUE tasks. These results are consistent with similar studies, e.g., @yosinski2014transferable’s results on fine-tuning a convolutional neural network pre-trained on ImageNet or @romanov2018lessons’s results on transfer learning for medical natural language inference.
Attention to linguistic features {#sec:vert_attention}
--------------------------------
In this experiment, we investigate whether fine-tuning BERT for a given task creates self-attention patterns which emphasize specific linguistic features. In this case, certain kinds of tokens may get high attention weights from all the other tokens in the sentence, producing vertical stripes on the corresponding attention maps ().
We tested this hypothesis by checking whether there are vertical stripe patterns corresponding to certain linguistically interpretable features, and to what extent such features are relevant for solving a given task. In particular, we investigated attention to nouns, verbs, pronouns, subjects, objects, and negation words[^6], and special BERT tokens across the tasks.
For every head, we compute the sum of self-attention weights assigned to the token of interest from each input token. Since the weights depend on the number of tokens in the input sequence, this sum is normalized by sequence length. This allows us to aggregate the weights for this feature across different examples. If there are multiple tokens of the same type (e.g. several nouns or negations), we take the maximum value. We disregard input sentences that do not contain a given feature.
For each investigated feature, we calculate this aggregated attention score for each head in every layer and build a map in order to detect the heads potentially responsible for this feature. We then compare the obtained maps to the ones derived using the pre-trained BERT model. This comparison enables us to determine if a particular feature is important for a specific task and whether it contributes to some tasks more than to others.
#### Results {#results-3 .unnumbered}
Contrary to our initial hypothesis that the vertical attention pattern may be motivated by linguistically meaningful features, we found that it is associated predominantly, if not exclusively, with attention to *\[CLS\]* and *\[SEP\]* tokens (see Figure \[fig:special\_tokens\]. Note that the absolute *\[SEP\]* weights for the SST-2 sentiment analysis task are greater than for other tasks, which is explained by the fact that there is only one sentence in the model inputs, i.e. only one *\[SEP\]* token instead of two. There is also a clear tendency for earlier layers to pay attention to *\[CLS\]* and for later layers to *\[SEP\]*, and this trend is consistent across all the tasks. We did detect heads that paid increased (compared to the pre-trained BERT) attention to nouns and direct objects of the main predicates (on the MRPC, RTE and QQP tasks), and negation tokens (on the QNLI task), but the attention weights of such tokens were negligible compared to *\[CLS\]* and *\[SEP\]*. Therefore, we believe that the striped attention maps generally come from BERT pre-training tasks rather than from task-specific linguistic reasoning.
Token-to-token attention {#sec:cross_attention}
------------------------
To complement the experiments in Sec. \[sec:vert\_attention\] and \[sec:fn\], in this section, we investigate the attention patterns between tokens in the same sentence, i.e. whether any of the tokens are particularly important while a *given* token is being processed. We were interested specifically in the verb-subject relation and the noun-pronoun relation. Also, since BERT uses the representation of the *\[CLS\]* token in the last layer to make the prediction, we used the features from the experiment in Sec. \[sec:vert\_attention\] in order to check if they get higher attention weights while the model is processing the *\[CLS\]* token.
#### Results {#results-4 .unnumbered}
Our token-to-token attention experiments for detecting heads that prioritize noun-pronoun and verb-subject links resulted in a set of potential head candidates that coincided with diagonally structured attention maps. We believe that this happened due to the inherent property of English syntax where the dependent elements frequently appear close to each other, so it is difficult to distinguish such relations from the previous/following token attention coming from language model pre-training.
Our investigation of attention distribution for the *\[CLS\]* token in the output layer suggests that for most tasks, with the exception of STS-B, RTE and QNLI, the *\[SEP\]* gets attended the most, as shown in . Based on manual inspection, for the mentioned remaining tasks, the greatest attention weights correspond to the punctuation tokens, which are in a sense similar to *\[SEP\]*.
Disabling self-attention heads {#sec:disabling}
------------------------------
Since there does seem to be a certain degree of specialization for different heads, we investigated the effects of disabling different heads in BERT and the resulting effects on task performance. Since BERT relies heavily on the learned attention weights, we define disabling a head as modifying the attention values of a head to be constant $a = \frac{1}{L}$ for every token in the input sentence, where $L$ is the length of the sentence. Thus, every token receives the same attention, effectively disabling the learned attention patterns while maintaining the information flow of the original model. Note that by using this framework, we can disable an arbitrary number of heads, ranging from a single head per model to the whole layer or multiple layers.
#### Results {#results-5 .unnumbered}
Our experiments suggest that certain heads have a detrimental effect on the overall performance of BERT, and this trend holds for all the chosen tasks. Unexpectedly, disabling some heads leads *not* to a drop in accuracy, as one would expect, but to an increase in performance. This is effect is different across tasks and datasets. While disabling some heads improves the results, disabling the others hurts the results. However, it is important to note that across all tasks and datasets, disabling some heads leads to an increase in performance. The gain from disabling a single head is different for different tasks, ranging from the minimum absolute gain of 0.1% for STS-B, to the maximum of 1.2% for MRPC (see ). In fact, for some tasks, such as MRPC and RTE, disabling a *random* head gives, on average, *an increase* in performance. Furthermore, disabling a whole layer, that is, all 12 heads in a given layer, also improves the results. shows the resulting model performance on the target GLUE tasks when different layers are disabled. Notably, disabling the first layer in the RTE task gives a significant boost, resulting in an absolute performance gain of 3.2%. However, effects of this operation vary across tasks, and for QNLI and MNLI, it produces a performance drop of up to -0.2%.
Discussion {#sec:discussion}
==========
In general, our results suggest that even the smaller base BERT model is significantly overparametrized. This is supported by the discovery of repeated self-attention patterns in different heads, as well as the fact that disabling both single and multiple heads is not detrimental to model performance and in some cases even improves it.
We found no evidence that attention patterns that are mappable onto core frame-semantic relations actually improve BERT’s performance. 2 out of 144 heads that seem to be “responsible" for these relations (see Section \[sec:fn\]) do not appear to be important in any of the GLUE tasks: disabling of either one does not lead to a drop of accuracy. This implies that fine-tuned BERT does not rely on this piece of semantic information and prioritizes other features instead. For instance, we noticed that both STS-B and RTE fine-tuned models rely on attention in the same pair of heads (head 1 in the fourth layer, and head 12 in the second layer), as shown in Figure \[fig:disable\_heads\_all\]. We manually checked the attention maps in those heads for a set of random inputs, and established that both of them have high weights for words that appear in both sentences of the input examples. This most likely means that word-by-word comparison of the two sentences provides a solid strategy of making a classification prediction for STS-B and RTE. Unfortunately, we were not able to provide a conceptually similar interpretation of heads important for other tasks.
Conclusion
==========
In this work, we proposed a set of methods for analyzing self-attention mechanisms of BERT, comparing attention patterns for the pre-trained and fine-tuned versions of BERT.
Our most surprising finding is that, although attention is the key BERT’s underlying mechanism, the model can benefit from attention “disabling”. Moreover, we demonstrated that there is redundancy in the information encoded by different heads and the same patterns get consistently repeated regardless of the target task. We believe that these two findings together suggest a further direction for research on BERT interpretation, namely, model pruning and finding an optimal sub-architecture reducing data repetition.
Another direction for future work is to study self-attention patterns in a different language. We think that it would allow to disentangle attention maps potentially encoding linguistic information and heads that use simple heuristics like attending to the following/previous tokens.
\[sec:supplemental\]
[^1]: <https://rajpurkar.github.io/SQuAD-explorer/>
[^2]: <https://gluebenchmark.com/leaderboard>
[^3]: <https://github.com/huggingface/pytorch-pretrained-BERT>
[^4]: [https://towardsdatascience.com/deconstructing-bert-distilling-6-patterns-\\from-100-million-parameters-b49113672f77](https://towardsdatascience.com/deconstructing-bert-distilling-6-patterns-\from-100-million-parameters-b49113672f77)
[^5]: If the number of development data examples for a given task exceeded 1000 (QQP, QNLI, MNLI, STS-B), we randomly sampled 1000 examples.
[^6]: Our manually constructed list of negation words consisted of the following words *neither, nor, not, never, none, don’t, won’t, didn’t, hadn’t, haven’t, can’t, isn’t, wasn’t, shouldn’t, couldn’t, nothing, nowhere.*
|
Rykrof Enloe - Homeworld - Naboo. Former Republic Commander, who is now an enemy of the Empire. He has a history of entanglements with the Badoo Corba terrorist organization, dating back prior to the Clone Wars. He is a former close friend of Imperial Baron Tylin Gere and was been forced to embark on a desperate mission to acquire an ancient Sith artifact for Tylin - which ended in failure. Since then, he has inadvertently led Imperial forces to a secret Rebel base on Banyss, which was destroyed by Darth Vader's forces. Rykrof has now been taken captive by the Empire and resides in the Imperial penal colony on Tartaaris. Most of those dear to him believe he is now deceased.
Atracion - Homeworld - Mon Calamari. This former slave trader was imprisoned after conspiring to steal a shipment of Coaxium for the crime organization known as Crimson Dawn. He quickly gained the respect of other inmates and became the leader of the largest prison gang in captivity on Tartaaris.
K3-95 - Homeworld - Tartaaris. Pieced together from scrap and given life by an unknown prisoner resident on Tartaaris. He has befriended Rykrof following the recent death of his friend, Garlin Nomad.
The Maker - Homeworld - Unknown. Very little is know about this myserious individual, other than he is the current warden of the Imperial penal colony on Tartaaris. For years, he has crafted custom hunter droids from the countless junk piles on Tartaaris and programmed them to terrorize the prisoner population. Rarely seen by the local garrison, the Maker has recently been stirring due to rumors of a pending prisoner revolt.
Baron Tylin Gere - Homeworld - Naboo. Former close ally of Rykrof Enloe, the two served as Republic Peace Officers prior to the Clone Wars and later as officers in the Grand Army of the Republic. Tylin thirsts for power and has become obsessed with acquiring an ancient Sith artifact, which was recently acquired by Darth Vader, who he secretly sees as a rival. He has taken custody of Alyssa and Caldin Enloe, and ordered the death of Rykrof's father. He believes Rykrof Enloe is dead, once and for all.
Emperor Palpatine - Homeword - Naboo. Dark Lord of the Sith and the undisputed leader of the Galactic Empire. He attempted to mold Rykrof Enloe as the ideal Imperial officer but failed. He takes great interest in Baron Tylin Gere's personal quests for power and admires his ambition.
Darth Vader - Homeworld - Tatooine. Dark Lord of the Sith and follows the instructions of his master, Emperor Palpatine. During the Clone Wars, he was once friends with Rykrof Enloe. He considers Rykrof a traitor to the Empire but also is fully aware that Rykrof once helped protect Padme Amidala, and rather than killing him, he has spared Rykrof's life by sending him to a secret penal colony on Tartaaris. However, his loyalties to Emperor Palpatine are secure.
Deep in the Outer Rim, Tartaaris servers as an undisclosed penal colony for former Separatists and other undesirables.
Many innocent beings have suffered immeasurably in the Imperial compound...
...leading one band of prisoners to rise against their oppressors.
Having already taken out a pair of Stormtroopers, the inmates know it won't be long before the Imperial authorities respond with force.
"How much further?" Atracion growls impatiently."Where is this supposed secret entrance into the complex?"
"I'm not quite sure," K3-95 admits. "Garlin never provided me with the exact coordinates, but based on my estimates, it should be just up ahead."
"We have to be getting close," Rykrof offers."Keep your eyes open for signs of a secret doorway... hatch... anything."
"You told us you knew how to get in," Lits gnarls.
"Well, it does look like we are nearing something important," the droid replies, observing the Imperial styled structures.
"Could it be under us?" Rykrof wonders aloud."Maybe the winds shifted the sands and covered a doorway?"
"I don't like this," Calo says, changing the subject."It's too quiet... we haven't seen any droids or activity since we took out those troops."
"He's right," Sarlo nods."What if we're walking into a trap? And where's the flank team?"
"I know," Rykrof agrees."Something's not right."
"K3, I think you might need this."
"Why thank you," the droid replies, gripping the blaster."It has been quite some time since I've handled one of these."
"Just don't give us any reason to regret giving you one, droid."
"We need to thin out... we're too exposed," Rykrof says to Atracion.
"Drop your weapons," a metallic voice suddenly orders the group.
"Damn..."
"Sir, I would not advise doing anything these mechanical abominations tell us to do."
For a brief moment, the droids remain frozen...
"We are here to see the Maker," Rykrof shouts to the sentinel droid.
"Grant us a ship to get off this planet, and we won't be of any trouble."
"Terminate them."
"Destroy... Destroy... Destroy..."
"Here they come!"
Rykrof ignites his lightsaber as the droids race toward the inmates!
"FLESH!"
"They've got us surrounded!"
A laser blast then rips through Parto's chest...
...as a pair of Viper droids fire at the group!
"Those probes are going to tear us apart!"
"I'm on it!" Velsa shouts...
...she then moves her aged walker into position...
...blasting the lead droid to pieces!
"Velsa! Behind you!"
But it's too late as the walker takes a heavy blast...
...crashing to the ground!
The prisoners continue to fight for their lives...
...but the sentinel droid penetrates their position!
Sarlo fires a desperate blast at the menace...
...as a laser blast rips into K3-95!
"Rykrof... I'm... sor...."
"Dammit!"
A laser blast then strikes Calo!
Rykrof then turns to face the sentinel droid...
...narrowly avoiding a deadly blow!
In desperation, Velsa sprints from the walker...
...but is instantly torn down!
The sentinel begins to lunge for Rykrof...
...but is met with a deep slash across its torso!
Rykrof then spins to face the remaining attackers...
...and finds himself facing the Maker!
"You're next... WHATEVER you are."
"You are beaten, Enloe," the Maker smiles.
"When you first arrived, my orders were to grant you leniency."
"Surrender your weapons, and only your friends will suffer..."
"...for I will show you ...mercy."
Without warning, a blast of energy slams through the Maker's cranial dome!
His body then bursts into an electrical convulsion!
Astonished, Rykrof looks into the distance...
...spotting survivors of Atracion's flank team...
...Halri Kidell and Helmo Iteris!
"Come on!" Rykrof shouts with renewed hope!
"The Maker's DEAD!"
Helmo then fires another blast...
...destroying the second Viper droid!
The survivors quickly gather at the scene of the massacre.
"This was a trap," Atracion growls."We could of all been killed."
"I'm sorry... but if we don't stand against them, we're as good as dead," Rykrof firmly replies.
"No time to argue," Halri says."It won't be long before the Imperials send the garrison out to finish us off."
"Chief Rolem will make us suffer."
"Right," Rykrof agrees."We need to carry on before they have time to react."
"Enloe..." calls the visage of the Maker, now in the body of a different droid!
"What the hell's going on?"
"HA HA HA HA HA HA HA!!!"
Another visage then emerges, charging directly at the group!
It quickly leaps onto Lits!
"We have to get out of here!"
Lits is then torn to shreds by the abomination!
Rykrof instinctively turns and runs for his life!
"Get out of here!" Helmo shouts.
With the Maker distracted, Rykrof and the others sprint to the distance...
...as the prisoner sacrifices himself for the group!
Rykrof and the others continue their desperate escape...
"We have to get back to base," Halri gasps.
"That's exactly where the Stormtroopers will be waiting for us," Rykrof argues.
"We go in there."
"The sewers?" Atracion retorts."Are you crazy?"
"The DWELLERS live in there," Halri warns."They're worse than the Maker's droids!"
Another great episode! As a long-time reader, I love seeing how your diorama and photography skills have progressed. Speaking of dioramas, I'm looking at all of the misc. vehicle parts (TIE fighter cockpits & wings, AT-AT legs, etc.) thinking, "My goodness, I wonder how much money is tied into this?" I know like most Star Wars collectors, as Hasbro has improved the sculpting of vehicles (big-wing TIE,BAT-AT) you probably have older (POTF2) versions left over for customizing & fodder.
I love the intro text to this much better than the crawl. It has sort of a "Solo" movie homage to it. By the way, I'm not sure where you got the image of the planet in the first shot, but that is a cool looking planet.
The story is well paced. The lighting to the pictures gives it that "other worldly" look. As usual, the action jumps off the screen. Absolutely LOVE this scrap yard looking diorama. There is so much going on in the junk and fodder in this place: Is that a Galoob Micro Machine Bespin Cloud City? And you even threw in a chainsaw chain in a few pics! How cool is that!
Another great episode! As a long-time reader, I love seeing how your diorama and photography skills have progressed. Speaking of dioramas, I'm looking at all of the misc. vehicle parts (TIE fighter cockpits & wings, AT-AT legs, etc.) thinking, "My goodness, I wonder how much money is tied into this?" I know like most Star Wars collectors, as Hasbro has improved the sculpting of vehicles (big-wing TIE,BAT-AT) you probably have older (POTF2) versions left over for customizing & fodder.
Thanks Jason - it's so cool that you've been keeping up with these, that means a lot of me and I know it can be hard reading through long chapters like this in this medium.
Yeah, the dioramas over time have gotten better I think, the photography is something that I really struggle with still. Just not my area of expertise... I need to eventually invest in the right camera and lighting techniques - I have such a crappy work space to photograph all of this too so it's kind of difficult.
I do have older versions of vehicles used for fodder but also have at times found some good deals at garage sales and clearance items.
jlw515 wrote:I love the intro text to this much better than the crawl. It has sort of a "Solo" movie homage to it. By the way, I'm not sure where you got the image of the planet in the first shot, but that is a cool looking planet.
Yes - I changed the text based on feedback from you and others. I plan to go back and update that for the older stories in "Trials" as well in a bit. I think the planet was a found image on google, then I altered it - been using that image for a few chapters now.
jlw515 wrote:The story is well paced. The lighting to the pictures gives it that "other worldly" look. As usual, the action jumps off the screen. Absolutely LOVE this scrap yard looking diorama. There is so much going on in the junk and fodder in this place: Is that a Galoob Micro Machine Bespin Cloud City? And you even threw in a chainsaw chain in a few pics! How cool is that!
Pacing can be hard at times to make it feel right, I try to make these to where they are not boring (but I'm sure some people would find these extremely boring!)... one of the keys I think is to try and let the photo do most of the talking... place the text for each picture directly below the image, and keep it short and sweet - don't make the reader have to think too much. I am not an elegant writer, and I also find myself easily distracted if text becomes cumbersome when reading something - so I try and keep that in mind as much as possible.
Yes - you did see a Galoob micro Cloud City! I ended up with a few of those somehow and thought they could work for some sort of droid or structures in the 3.75" scale. And the chain - lol! Good eye.
As soon as they teased that walker a while back, I wanted it for my photonovels, ha ha. I figured this one looked Star Warsy enough to not need to be customized, at least in my opinion. I did blur out the words on the front of the walker in the pics though.
This chapter was awesome! The only issue is that it is not long enough. I liked the references to solo that you put into atracion's backstory, and I am intrigued about both the maker and the dwellers. I still have to go through and look at all of the dioramas, but overall it was a great chapter.
ImperialOfficer wrote:This chapter was awesome! The only issue is that it is not long enough. I liked the references to solo that you put into atracion's backstory, and I am intrigued about both the maker and the dwellers. I still have to go through and look at all of the dioramas, but overall it was a great chapter.
Thanks for reading and the feedback!
As for the not long enough part, I agree some with that - I could have made this longer (removed maybe a dozen pics). I removed them because I felt it was getting too drawn out, and I was also wary of adding another scene because as it stands, the chapter was a bit over 100 pics and I'm afraid people will stop reading once it gets to a certain point... even more than 20 or so frames can scare people off, I'm afraid...?
More to come on the Maker and the Dwellers - I have some new customs and more dioramas to work on.
I'll also snap a pic soon of all the various diorama pieces for this chapter in a pic to give some perspective on what a mess I create with these.
Scenery--from the red clay sand to the junk greeblies strewn across the landscape to the excellent building facades the background allows the story to come alive. Special praise is garnered for the sewer; down to the greenish liquid on the lip of the opening. Wonderful use of the acid rain mech too.
Figures: Again the unique look of individual prisoners in orange is really nice. The battle robots using iron man parts, Doctor who, assorted star wars and droids I can't make out are really nice. Would love more recipes on those pieces.
Plot: Bold. Be sure to read even the character descriptions as they have been updated. Good mix of action and battle scenes with strong dialogue. Vesla came to the rescue and I thought she might save the day for all. In fact I started to wonder if we might see her more in the future. But alas...it was not meant to be. The quarren gang boss is quite fun and has some superb lines such as follow the fool and something to the effect of all of Enloe's ideas wind up killing his men.
The biggest surprise is the Maker. What is he now? What has he ever been. Loved how his consciousness (visage was your wording) moved from one robot body to another and in fact multiples at once. I must admit the villain was not necessarily grabbing my interest as say our fiendish Baron. That is a character development to start off a Republic Peace officer from Naboo to now an evil minion of the Emperor. But the Maker's story is one that needs to be understood. Looking forward to knowing about the dwellers. (edited; forgot to mention them the first go around.)
Did I miss something or did the droid that accompanied Enloe die in the battle? Where is he? It is possible my computer which was jumping some while I read hid the droid's demise from me. I will have to reread. (EDIT: Asked and answered. The droid died quickly)
So the questions continue to build: 1) How long will Enloe be here? 2) What will the outside world be like once he does escape? 3) Will Freelo have changed much? 4) What possible Solo or Rebels characters might Enloe meet when he does leave? I can't even ask what about Alyssa and Calden; I am thinking Calden is going to be a real imperial goose stepper.
Last edited by UKHistory on Sat Jun 09, 2018 10:04 am, edited 2 times in total.
Just read it a second time, wow. Really noticed the dynamic action poses, how did you put the guy in midair being thrust away from an explosive impact?? And the viper droid crashing into the wall as it's hit?! Amazing. |
1. Field of the Invention
This invention relates generally to testing methods and apparatus for semiconductor devices. More particularly, the invention pertains to a method and apparatus for measuring localized temperatures present on semiconductor devices and the like for research and development purposes.
2. State of the Art
Modern integrated circuit (IC) devices are commonly formed by joining the electrically active bond pads of a semiconductor die to the conductive lead fingers of a leadframe with metal wires. The wire bonding process may comprise:
a. thermocompression bonding, which uses pressure and elevated temperature, typically 300-400xc2x0 C. to bond the wire ends to the bond pads and leadframe;
b. thermosonic bonding, in which ultrasonic energy is combined with compression at temperatures of about 150xc2x0 C.; or
c. ultrasonic bonding, in which ultrasonic energy is typically applied at ambient temperatures. This method is generally limited to some specific metals such as aluminum or aluminum alloy wires on aluminum or gold pads.
As is well known, the functionality of manufactured electronic devices depends upon successful bonding of the wires to the bond pads of the die and to the lead fingers.
In each of thermocompression bonding and thermosonic bonding, reliability of the bonding process depends upon the temperatures of the elements being joined.
It is important for a semiconductor device manufacturer to have the capability for evaluating the quality of conductor bonds, such as wire bonds, leadframe to bump bonds, etc. Evaluation of the bonding process includes, e.g., destructive ball shear tests and wire bond pull tests as well as contaminant tests such as by spectrographic analysis.
In addition, thermal analysis of the die and leadframe may be done during the conductor bonding operations to yield an indication as to wire bonding quality. Thus, for example, U.S. Pat. No. No. 5,500,502 of Horita et al. describes a process for bonding a leadframe to a bump using laser irradiation. The state of contact between the leadframe and the bump is then tested using the intensity of the emitted infrared radiation as a measure of the leadframe temperature. Knowing the time lapse between the laser radiation and the measured temperature, the temperature as a function of time may be calculated, particularly a threshold temperature correlated to bond effectiveness and the resulting quality of the wire bond.
The Horita et al. method does not address the testing of wire bonds. Furthermore, the method depends upon the emission and reflection of infrared radiation, which varies with the surface characteristics of the material whose temperature is being measured. As is well known, both semiconductor dies and leadframes are made of a variety of materials, each of which may have a differing emission/reflection temperature function when laser-irradiated. In addition, a wide variety of materials is used for doping semiconductor dice and for coating dice. For example, U.S. Pat. No. No. 5,256,566 of Bailey teaches the coating of dice with polysilicon. Thus, the infrared temperature meter must be calibrated for each material, making temperature measurements labor intensive.
Furthermore, the presence of contaminants on the die or leadframe surfaces will affect the accuracy of the Horita et al. method.
A method and apparatus for accurately measuring the temperature of very small areas of surfaces, independent of the surface composition, are desirable for research and development purposes in the semiconductor die area.
The present invention is directed to a method and apparatus for accurately measuring the temperature of precisely defined areas of surfaces of materials having a wide variety of compositions, such as a semiconductor die and/or leadframe.
An apparatus and method for producing a computer-generated thermal map of the surface of a semiconductor die and/or attached leadframe, wafer, or other object are described herein. The apparatus may be used to measure, compile, collate, plot, and display temperatures of a die and its associated leadframe fingers for evaluating a manufacturing process. The apparatus may be configured to back-calculate measured real-time temperatures to a predetermined initial time for preparing thermal maps, e.g., initial or maximum temperatures as a function of location and time.
The apparatus includes (a) a fiber-optic temperature sensor mounted on the bondhead of a wire bonding machine, and connected to (b) a thermometer apparatus which calculates a temperature based on the sensor output via a (c) signal isolation trigger box having a circuit which is connected to the ultrasonic generator output of the wire bonding machine, whereby a temperature measurement is initiated, and to (d) a computer having software for controlling the wire bonding machine and trigger box and for storing and collating temperature measurements (and other measurements) from the thermometer controller and wire bonding machine.
The invention may be applied to temperature measurements on a die, wafer, semiconductor device at any stage of construction, or surfaces of other objects of interest. The temperature measurements may be xe2x80x9crasteredxe2x80x9d over the surface by the stage controller, using any desired increment of movement, because the temperature sensor tip may have a size approximating the size of the area of which the temperature is to be measured. |
[Effect of 5-bromodeoxyuridine in vivo and in vitro on the development of the avian skeleton].
The effects produced in vivo and in organ culture on the differentiation of the somitic mesenchyme by the thymidine analog, 5-bromodeoxyuridine, are dependant on the experimental conditions. The analog can give rise to irreversible or reversible blockade of cell differentiation or to inhibitory effects of the biosynthesis of several macromolecules normally secreted by the somitic cells. In vivo, the analog causes skeletal malformations affecting particularly the lumbo-sacral area and the hind limbs. |