Abstractors

This is the code repository associated with the paper

"Abstractors and relational cross-attention: An inductive bias for explicit relational reasoning in Transformers" --- Awni Altabaa, Taylor Webb, Jonathan D. Cohen, John Lafferty.

This paper appears in ICLR 2024. The arXiv version is here: https://arxiv.org/abs/2304.00195. The project webpage contains a high-level summary of the paper and can be found here: https://awni00.github.io/abstractor.

The following is an outline of the repo:

• abstracters.py implements the main variant of the abstracter module with positional symbols. symbol_retrieving_abstractor.py implements an abstractor with symbol-retrieval via symbolic attention. abstractor.py is a 'simplified' implementation that avoids using tensorflow's MHA layer.
• autoregressive_abstractor.py implements sequence-to-sequence abstractor-based architectures. seq2seq_abstracter_models.py is an older, less general, implementation of sequence-to-sequence models.
• multi_head_attention.py is a fork of tensorflow's implementation of MultiHeadAttention which we have adjusted to support different kinds of activation functions applied to the attention scores. transformer_modules.py includes implementations of different Transformer modules (e.g.: Encoders, Decoders, etc.). Finally, attention.py implements different attention mechanisms for Transformers and Abstractors (including relational cross-attention).
• The experiments directory contains the code for all experiments in the paper. See the readme's therein for details on the experiments and instructions for replicating them.