DeepSITH: Efficient Learning via Decomposition of What and When Across Time Scales

Authors: Brandon Jacques, Zoran Tiganj, Marc Howard, Per B Sederberg

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

Reproducibility Variable Result LLM Response
Research Type Experimental We compare Deep SITH to LSTMs and other recent RNNs on several time series prediction and decoding tasks. Deep SITH achieves results comparable to state-of-the-art performance on these problems and continues to perform well even as the delays are increased.
Researcher Affiliation Academia Brandon G. Jacques Department of Psychology University Of Virginia bgj5hk@virginia.eduZoran Tiganj Department of Computer Science Indiana University ztiganj@iu.eduMarc W. Howard Department of Psychological and Brain Sciences Boston University marc777@bu.eduPer B. Sederberg Department of Psychology University of Virginia pbs5u@virginia.edu
Pseudocode No No structured pseudocode or algorithm blocks were found in the paper.
Open Source Code Yes We provide all the code for Deep SITH and the subesquent analysis in our github here.
Open Datasets Yes In the MNIST task [22], handwritten numerical digits can be identified by neural networks with almost 100% accuracy utilizing a convolutional neural network (CNN).
Dataset Splits Yes The Deep SITH network is trained with a batch size of 64, with a cross entropy loss function, with a training/test split of 80%-20%.
Hardware Specification No We did not ran anything that could not be ran on a single GPU
Software Dependencies No The paper mentions "Py Torch machine learning framework [19]" but does not provide specific version numbers for PyTorch or any other software dependencies.
Experiment Setup Yes Table 1 summarizes the hyperparameters used in the experiments. ... The Deep SITH network is trained with a batch size of 64, with a cross entropy loss function, with a training/test split of 80%-20%. In between each layer we applied batch normalization, and applied a step learning rate annealing after every third of the training epochs (2e-3, 2e-4, 2e-5).