Projecting intermediate representations onto the vocabulary is an increasingly popular interpretation tool for transformer-based LLMs, also known as the logit lens. We propose a quantitative extension to this approach and define spectral filters on intermediate representations based on partitioning the singular vectors of the vocabulary embedding and unembedding matrices into bands. We find that the signals exchanged in the tail end of the spectrum are responsible for attention sinking (Xiao et al. 2023), of which we provide an explanation. We find that the loss of pretrained models can be kept low despite suppressing sizable parts of the embedding spectrum in a layer-dependent way, as long as attention sinking is preserved. Finally, we discover that the representation of tokens that draw attention from many tokens have large projections on the tail end of the spectrum.

San Francisco (CNN Business)It's hard for the average person to tell Dani, Lenore, and Bella apart: They all sport fashionably fuzzy brown coats and enjoy a lot of the same activities, like playing in icy-cold water and, occasionally, ripping apart a freshly caught fish. Melanie Clapham is not the average person. As a bear biologist, she has spent over a decade studying these grizzly bears, who live in Knight Inlet in British Columbia, Canada, and developed a sense for who is who by paying attention to little things that make them different. "I use individual characteristics -- say, one bear has a nick in its ear or a scar on the nose," she said. But Clapham knows most people don't have her eye for detail, and the bears' appearances change dramatically over the course of a year -- such as when they get winter coats and fatten up before denning -- which makes it even harder to distinguish between, say, Toffee and Blonde Teddy.

Researchers say the new technology, termed BearID, created a'non-invasive' technique to study the animals. Despite a decade of behavioural research on grizzly bears in B.C.'s Knight Inlet, Melanie Clapham still has trouble telling some individual bears apart. Brown bears, which include grizzly bears, can change dramatically in their appearance during their younger years and, unlike other wildlife that has spots or stripes, they lack distinguishing markings on their bodies. Ms. Clapham, a conservation biologist and postdoctoral research fellow at the University of Victoria, dreamed of technology that could help her individually identify these furry mammals. While she was looking for a tech team to make that idea possible, south of the border, Ed Miller and Mary Nguyen, two Silicon Valley engineers who are also outdoor and wildlife enthusiasts, had started a project to develop machine-learning models that could be adapted to grizzly bears.

Grizzly bears have domed shoulders, tall foreheads, and pale-tipped fur that gives them their grizzled appearance. If you're comparing two bears, one might be lighter or darker in color, or fatter for hibernation. But for the most part, there's no universal, unique marker a person can use to tell two bears apart. This issue is a challenge for scientists like University of Victoria wildlife conservationist Melanie Clapham, whose research on grizzly bear behavior requires her to monitor individual bears over years, Adam van der Zwan reports for CBC. But now, Clapham and her research team have developed a solution: facial recognition for bears. Bears grow and shrink a lot depending on the season, and their appearance changes frequently during their 20- to 25-year-long lifespans.