On Blackbox Backpropagation and Jacobian Sensing

Authors: Krzysztof M. Choromanski, Vikas Sindhwani

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

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate efficient backpropagation through noisy blackbox layers in a deep neural net, improved data-efficiency in the task of linearizing the dynamics of a rigid body system, and the generic ability to handle a rich class of input-output dependency structures in Jacobian estimation problems. (...) 4 Experiments (...) 4.1. Sparse Jacobian Recovery: (...) 4.2. Training Convolutional Neural Networks with Blackbox Nodes: (...) 4.3. Jacobian of manipulator dynamics:
Researcher Affiliation Industry Krzysztof Choromanski Google Brain New York, NY 10011 kchoro@google.com Vikas Sindhwani Google Brain New York, NY 10011 sindhwani@google.com
Pseudocode No The paper describes algorithmic steps (e.g., ADMM iterations) in text and equations, but does not include a formally labeled 'Pseudocode' or 'Algorithm' block.
Open Source Code No The paper does not include an unambiguous statement that the authors are releasing their source code for the work described, nor does it provide a direct link to a code repository for their methodology.
Open Datasets Yes We introduce a blackbox layer between the convolutional layers and the fully connected layers of a standard MNIST convnet.
Dataset Splits No The paper mentions 'training and validation error' and 'validation error' but does not provide specific percentages or sample counts for dataset splits (e.g., 80/10/10 split) or reference predefined splits with citations for reproducibility.
Hardware Specification No The paper does not provide any specific hardware details such as GPU or CPU models, processor types, or memory specifications used for running its experiments.
Software Dependencies No The paper mentions software like 'Tensor Flow [5]', 'Torch [2]', and 'MIT’s Drake planning and control toolbox [25]', but does not provide specific version numbers for any of these components.
Experiment Setup Yes The blackbox node is a standard Re LU layer that takes as input 32-dimensional vectors, 32 32-sized weight matrix and a bias vector of length 32, and outputs a 32 dimensional representation. The minibatch size is 16. We inject truncated Gaussian noise in the output of the layer and override its default gradient operator in Tensor Flow with our LP-based rainbow procedure. We use Gaussian perturbation directions and sample measurements by forward evaluation calls to the Tensor Flow Op inside our custom blackbox gradient operator.