Hierarchical interpretations for neural network predictions

Authors: Chandan Singh, W. James Murdoch, Bin Yu

ICLR 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the utility of ACD on both long short term memory networks (LSTMs) (Hochreiter & Schmidhuber, 1997) trained on the Stanford Sentiment Treebank (SST) (Socher et al., 2013) and CNNs trained on MNIST (Le Cun, 1998) and Image Net (Russakovsky et al., 2015). Through human experiments, we show that ACD produces intuitive visualizations that enable users to better reason about and trust DNNs. We also demonstrate that ACD s hierarchy is robust to adversarial perturbations (Szegedy et al., 2013) in CNNs.
Researcher Affiliation Academia Chandan Singh Department of EECS UC Berkeley c singh@berkeley.edu W. James Murdoch Department of Statistics UC Berkeley jmurdoch@berkeley.edu Bin Yu Department of Statistics, EECS UC Berkeley binyu@berkeley.edu
Pseudocode Yes Algorithm 1 Agglomeration algorithm. ACD(Example x, model, hyperparameter k, function CD(x, blob; model))
Open Source Code Yes Code and scripts for running ACD and experiments available at https://github.com/csinva/ acd
Open Datasets Yes We demonstrate the utility of ACD on both long short term memory networks (LSTMs) (Hochreiter & Schmidhuber, 1997) trained on the Stanford Sentiment Treebank (SST) (Socher et al., 2013) and CNNs trained on MNIST (Le Cun, 1998) and Image Net (Russakovsky et al., 2015).
Dataset Splits No The paper mentions 'validation set' in Table 1 but does not provide explicit training/validation/test dataset splits (e.g., percentages or sample counts).
Hardware Specification No The paper states 'All models are implemented using Py Torch' but does not provide any specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments.
Software Dependencies No The paper states 'All models are implemented using Py Torch' but does not provide specific version numbers for PyTorch or any other software dependencies.
Experiment Setup No The paper mentions training models using 'standard best practices' and 'standard Py Torch example', but it does not provide specific hyperparameters like learning rate, batch size, number of epochs, or other detailed training settings.