Decoupled Neural Interfaces using Synthetic Gradients

Authors: Max Jaderberg, Wojciech Marian Czarnecki, Simon Osindero, Oriol Vinyals, Alex Graves, David Silver, Koray Kavukcuoglu

ICML 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section we perform empirical expositions of the use of DNIs and synthetic gradients, first by applying them to RNNs in Sect. 3.1 showing that synthetic gradients extend the temporal correlations an RNN can learn. Secondly, in Sect. 3.2 we show how a hierarchical, two-timescale system of networks can be jointly trained using synthetic gradients to propagate error signals between networks. Finally, we demonstrate the ability of DNIs to allow asynchronous updating of layers a feed-forward network in Sect. 3.3.
Researcher Affiliation Industry 1Deep Mind, London, UK. Correspondence to: Max Jaderberg <jaderberg@google.com>.
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide concrete access to source code for the methodology described.
Open Datasets Yes Copy and Repeat Copy We first look at two synthetic tasks Copy and Repeat Copy tasks from (Graves et al., 2014). [...] Language Modelling We also applied our DNI-enabled RNNs to the task of character-level language modelling, using the Penn Treebank dataset (Marcus et al., 1993).
Dataset Splits Yes We measure error in bits per character (BPC) as in (Graves, 2013), perform early stopping based on validation set error, and for simplicity do not perform any learning rate decay.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory amounts) used for running its experiments.
Software Dependencies No The paper does not provide specific ancillary software details with version numbers.
Experiment Setup Yes The implementation details of the RNN models are given in Sect. D.2 in the Supplementary Material. [...] For full experimental details please refer to Sect. D.2 in the Supplementary Material. [...] Full experimental details can be found in Sect. D.3 in the Supplementary Material. [...] More details are given in the Supplementary Material.