Feedback Transformers: Addressing Some Limitations of Transformers with Feedback Memory (Explained)
519#ai #science #transformers
Autoregressive Transformers have taken over the world of Language Modeling (GPT-3). However, in order to train them, people use causal masking and sample parallelism, which means computation only happens in a feedforward manner. This results in higher layer information, which would be available, to not be used in the lower layers of subsequent tokens, and leads to a loss in the computational capabilities of the overall model. Feedback Transformers trade-off training speed for access to these representations and demonstrate remarkable improvements in complex reasoning and long-range dependency tasks.
OUTLINE:
0:00 - Intro & Overview
1:55 - Problems of Autoregressive Processing
3:30 - Information Flow in Recurrent Neural Networks
7:15 - Information Flow in Transformers
9:10 - Solving Complex Computations with Neural Networks
16:45 - Causal Masking in Transformers
19:00 - Missing Higher Layer Information Flow
26:10 - Feedback Transformer Architecture
30:00 - Connection to Attention-RNNs
36:00 - Formal Definition
37:05 - Experimental Results
43:10 - Conclusion & Comments
Paper: https://arxiv.org/abs/2002.09402
My video on Attention: https://youtu.be/iDulhoQ2pro
ERRATA: Sometimes I say "Switch Transformer" instead of "Feedback Transformer". Forgive me :)
Abstract:
Transformers have been successfully applied to sequential, auto-regressive tasks despite being feedforward networks. Unlike recurrent neural networks, Transformers use attention to capture temporal relations while processing input tokens in parallel. While this parallelization makes them computationally efficient, it restricts the model from fully exploiting the sequential nature of the input. The representation at a given layer can only access representations from lower layers, rather than the higher level representations already available. In this work, we propose the Feedback Transformer architecture that exposes all previous representations to all future representations, meaning the lowest representation of the current timestep is formed from the highest-level abstract representation of the past. We demonstrate on a variety of benchmarks in language modeling, machine translation, and reinforcement learning that the increased representation capacity can create small, shallow models with much stronger performance than comparable Transformers.
Authors: Angela Fan, Thibaut Lavril, Edouard Grave, Armand Joulin, Sainbayar Sukhbaatar
Links:
TabNine Code Completion (Referral): http://bit.ly/tabnine-yannick
YouTube: https://www.youtube.com/c/yannickilcher
Twitter: https://twitter.com/ykilcher
Discord: https://discord.gg/4H8xxDF
BitChute: https://www.bitchute.com/channel/yannic-kilcher
Minds: https://www.minds.com/ykilcher
Parler: https://parler.com/profile/YannicKilcher
LinkedIn: https://www.linkedin.com/in/yannic-kilcher-488534136/
BiliBili: https://space.bilibili.com/1824646584
If you want to support me, the best thing to do is to share out the content :)
If you want to support me financially (completely optional and voluntary, but a lot of people have asked for this):
SubscribeStar: https://www.subscribestar.com/yannickilcher
Patreon: https://www.patreon.com/yannickilcher
Bitcoin (BTC): bc1q49lsw3q325tr58ygf8sudx2dqfguclvngvy2cq
Ethereum (ETH): 0x7ad3513E3B8f66799f507Aa7874b1B0eBC7F85e2
Litecoin (LTC): LQW2TRyKYetVC8WjFkhpPhtpbDM4Vw7r9m
Monero (XMR): 4ACL8AGrEo5hAir8A9CeVrW8pEauWvnp1WnSDZxW7tziCDLhZAGsgzhRQABDnFy8yuM9fWJDviJPHKRjV4FWt19CJZN9D4n