Neural Attention Memory

We propose a novel perspective of the attention mechanism by reinventing it as a memory architecture for neural networks, namely Neural Attention Memory (NAM). NAM is a memory structure that is both readable and writable via differentiable linear algebra operations. We explore three use cases of NAM: memory-augmented neural network (MANN), few-shot learning, and efficient long-range attention. First, we design two NAM-based MANNs of Long Short-term Memory (LSAM) and NAM Turing Machine (NAM-TM) that show better computational powers in algorithmic zero-shot generalization tasks compared to other baselines such as differentiable neural computer (DNC). Next, we apply NAM to the N-way K-shot learning task and show that it is more effective at reducing false positives compared to the baseline cosine classifier. Finally, we implement an efficient Transformer with NAM and evaluate it with long-range arena tasks to show that NAM can be an efficient and effective alternative for scaled dot-product attention.

I-BERT: Inductive Generalization of Transformer to Arbitrary Context Lengths

Self-attention has emerged as a vital component of state-of-the-art sequence-to-sequence models for natural language processing in recent years, brought to the forefront by pre-trained bi-directional Transformer models. Its effectiveness is partly due to its non-sequential architecture, which promotes scalability and parallelism but limits the model to inputs of a bounded length. In particular, such architectures perform poorly on algorithmic tasks, where the model must learn a procedure which generalizes to input lengths unseen in training, a capability we refer to as inductive generalization. Identifying the computational limits of existing self-attention mechanisms, we propose I-BERT, a bi-directional Transformer that replaces positional encodings with a recurrent layer. The model inductively generalizes on a variety of algorithmic tasks where state-of-the-art Transformer models fail to do so. We also test our method on masked language modeling tasks where training and validation sets are partitioned to verify inductive generalization. Out of three algorithmic and two natural language inductive generalization tasks, I-BERT achieves state-of-the-art results on four tasks.