The Shape of Learning: Intrinsic Dimensions in Transformer-Based Models
Препринт нашей новой работы! Оказалось, что языковые модели «упаковывают» свои репрезентации в очень маленькое пространство с внутренней размерностью не больше 60. И при этом анизотропия на средних слоях трансформеров-декодеров стремится к единице! Получается, эмбеддинги из середины модели расположены вдоль одной линии.
Еще одно интересное наблюдение — обучение LLM делится на две фазы: расширение и последующее сжатие активаций (см. картинку). А перед взрывами лосса их размерность немного подрастает.
The Shape of Learning: Intrinsic Dimensions in Transformer-Based Models
Препринт нашей новой работы! Оказалось, что языковые модели «упаковывают» свои репрезентации в очень маленькое пространство с внутренней размерностью не больше 60. И при этом анизотропия на средних слоях трансформеров-декодеров стремится к единице! Получается, эмбеддинги из середины модели расположены вдоль одной линии.
Еще одно интересное наблюдение — обучение LLM делится на две фазы: расширение и последующее сжатие активаций (см. картинку). А перед взрывами лосса их размерность немного подрастает.
Right now the digital security needs of Russians and Ukrainians are very different, and they lead to very different caveats about how to mitigate the risks associated with using Telegram. For Ukrainians in Ukraine, whose physical safety is at risk because they are in a war zone, digital security is probably not their highest priority. They may value access to news and communication with their loved ones over making sure that all of their communications are encrypted in such a manner that they are indecipherable to Telegram, its employees, or governments with court orders. For example, WhatsApp restricted the number of times a user could forward something, and developed automated systems that detect and flag objectionable content. "He has kind of an old-school cyber-libertarian world view where technology is there to set you free," Maréchal said. Since January 2022, the SC has received a total of 47 complaints and enquiries on illegal investment schemes promoted through Telegram. These fraudulent schemes offer non-existent investment opportunities, promising very attractive and risk-free returns within a short span of time. They commonly offer unrealistic returns of as high as 1,000% within 24 hours or even within a few hours. But Telegram says people want to keep their chat history when they get a new phone, and they like having a data backup that will sync their chats across multiple devices. And that is why they let people choose whether they want their messages to be encrypted or not. When not turned on, though, chats are stored on Telegram's services, which are scattered throughout the world. But it has "disclosed 0 bytes of user data to third parties, including governments," Telegram states on its website.
from in
