Deep equilibrium networks are sensitive to initialization statistics

Authors: Atish Agarwala, Samuel S Schoenholz

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We begin by training a fully-connected DEQ on MNIST (Le Cun et al., 1989). ... We next examine the effects of the matrix ensembles on a DEQ using a vanilla transformer layer from (Al-Rfou et al., 2019) as the base of the DEQ layer, trained on Wikitext-103 (Merity et al., 2016).
Researcher Affiliation Industry 1Google Research, Brain Team.
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code No The paper refers to basing its experiments on 'a Haiku implementation of a DEQ transformer (Khan, 2020)', but does not explicitly state that the code specific to the methodology described in this paper is open-source or provide a link to it.
Open Datasets Yes We begin by training a fully-connected DEQ on MNIST (Le Cun et al., 1989). ... trained on Wikitext-103 (Merity et al., 2016).
Dataset Splits No The paper mentions training on datasets and evaluating test error/loss, but does not explicitly provide details about train/validation/test dataset splits or cross-validation setup.
Hardware Specification Yes We trained on TPUv3.
Software Dependencies No The paper mentions using 'Haiku implementation' and 'sentencepiece tokenizer' but does not provide specific version numbers for these or any other software dependencies.
Experiment Setup Yes Learning rate tuning of an ADAM optimizer with momentum of 0.9 suggested an optimal learning rate of 10−2 for all the conditions studied. We trained on Wikitext-103 (Merity et al., 2016) with a batch size of 512 and a context length of 128. We ran for 20 steps of the Broyden solver. For the experiments with multiple seeds, we used a learning rate of 10−3 with a linear warmup for 2×10^3 steps, followed by a cosine learning rate decay for 5×10^4 steps.