Learning Associative Memories with Gradient Descent

Authors: Vivien Cabannes, Berfin Simsek, Alberto Bietti

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through theory and experiments, we provide several insights. ... We complement our analysis with experiments, investigating small multi-layer Transformer models with our associative memory viewpoint and identifying similar behaviors to those pinpointed in the simpler models.
Researcher Affiliation Collaboration 1Meta AI 2Flatiron. Correspondence to: <vivc@meta.com>.
Pseudocode No The paper does not contain any pseudocode or clearly labeled algorithm blocks.
Open Source Code No The paper does not provide any explicit statements about releasing source code or links to a code repository.
Open Datasets No We consider full-batch gradient descent on a dataset of 16 384 sequences of length 256 generated from the model described above with N = 64 tokens. ... The tokens following all non-trigger tokens are randomly sampled from a sequence-independent Markov model (namely, a character-level bigram model estimated from Shakespeare text data). The dataset is generated by the authors, and no specific link, DOI, or citation to a publicly available instance of this generated dataset is provided.
Dataset Splits No We consider full-batch gradient descent on a dataset of 16 384 sequences of length 256 generated from the model described above with N = 64 tokens. No explicit training, validation, or test dataset splits are provided.
Hardware Specification No The paper does not provide any specific details about the hardware (e.g., GPU models, CPU types, memory specifications) used to run the experiments.
Software Dependencies No The paper mentions 'pytorch convention' in Appendix A but does not list any specific software or library names with version numbers required for reproducibility.
Experiment Setup Yes We consider full-batch gradient descent on a dataset of 16 384 sequences of length 256 generated from the model described above with N = 64 tokens. ... Training losses are shown for different step-sizes η, and margins are shown for 5 different tokens. ... In Figure 6, we consider a setup with N = M = 5, f (x) = x, and p(x) 1/x, in different dimensions (with random embeddings).