Transformers have become the dominant model class in the last few years for large data, but their quadratic complexity in terms of sequence length has plagued them until now. Fastformer claims to be the fastest and most performant linear attention variant, able to consume long contexts at once. This is achieved by a combination of additive attention and elementwise products. OUTLINE: 0:00 - Intro & Outline 2:15 - Fastformer description 5:20 - Baseline: Classic Attention 10:00 - Fastformer architecture 12:50 - Additive Attention 18:05 - Query-Key element-wise multiplication 21:35 - Redundant modules in Fastformer 25:00 - Problems with the architecture 27:30 - Is this even attention? Abstract: Transformer is a powerful model for text understanding.

Transformers are dominating Deep Learning, but their quadratic memory and compute requirements make them expensive to train and hard to use. Many papers have attempted to linearize the core module: the attention mechanism, using kernels - for example, the Performer. However, such methods are either not satisfactory or have other downsides, such as a reliance on random features. This paper establishes an intrinsic connection between linearized (kernel) attention and the much older Fast Weight Memory Systems, in part popularized by Jürgen Schmidhuber in the 90s. It shows the fundamental limitations of these algorithms and suggests new update rules and new kernels in order to fix these problems. The resulting model compares favorably to Performers on key synthetic experiments and real-world tasks.

Hopfield Networks are one of the classic models of biological memory networks. This paper generalizes modern Hopfield Networks to continuous states and shows that the corresponding update rule is equal to the attention mechanism used in modern Transformers. It further analyzes a pre-trained BERT model through the lens of Hopfield Networks and uses a Hopfield Attention Layer to perform Immune Repertoire Classification. My Video on Attention: https://youtu.be/iDulhoQ2pro My Video on BERT: https://youtu.be/-9evrZnBorM

Surprisingly, they outperform Bayesian Networks, which are - in theory - doing the same thing. This paper investigates how Deep Ensembles are especially suited to capturing the non-convex loss landscape of neural networks.