The Lipschitz Constant of Self-Attention

Authors: Hyunjik Kim, George Papamakarios, Andriy Mnih

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To demonstrate the practical relevance of our theoretical work, we formulate invertible self-attention and use it in a Transformer-based architecture for a characterlevel language modelling task. We compare its test log-likelihood and stability to dot-product self-attention.
Researcher Affiliation Industry 1Deep Mind, UK. Correspondence to: Hyunjik Kim <hyunjikk@google.com>.
Pseudocode No The paper describes algorithms and mathematical formulations in text and equations but does not provide structured pseudocode or algorithm blocks.
Open Source Code No The paper does not explicitly state that source code for the described methodology is publicly available, nor does it provide a link to a code repository.
Open Datasets Yes at character-level language modelling on the Penn Treebank dataset (Marcus et al., 1993).
Dataset Splits Yes tuning the hyperparameters on a validation set.
Hardware Specification No This was the deepest model we could fit on a single GPU, and we expect to be able to train even deeper models with these two. (Section 5.4) - This mentions a GPU but no specific model or specifications are provided.
Software Dependencies No The paper does not provide specific version numbers for any software libraries, frameworks (e.g., PyTorch, TensorFlow), or programming languages used.
Experiment Setup Yes In practice, this leads to instabilities in training for DP-MHA, hence requiring careful tuning of the learning rate schedule for training deeper Transformer models: linear warmup and square root decay, as detailed in Appendix H.