Predify: Augmenting deep neural networks with brain-inspired predictive coding dynamics

Authors: Bhavin Choksi, Milad Mozafari, Callum Biggs O'May, B. ADOR, Andrea Alamia, Rufin VanRullen

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

Reproducibility Variable Result LLM Response
Research Type Experimental We show that implementing this strategy into two popular networks, VGG16 and Efficient Net B0, improves their robustness against various corruptions and adversarial attacks.
Researcher Affiliation Academia Bhavin Choksi Cer Co CNRS, UMR 5549 & Université de Toulouse bhavin.choksi@cnrs.fr
Pseudocode Yes Pseudocode 1 Predictive Coding Iterations
Open Source Code Yes Predify is an ongoing project available on Git Hub3 under GNU General Public License v3.0. The footnote points to https://github.com/miladmozafari/predify.
Open Datasets Yes We select VGG16 and Efficient Net B0, two different pre-trained feedforward networks on Image Net, and augment them with the proposed predictive coding dynamics. and We use Image Net-C, a benchmarking dataset for noise robustness provided by [8]
Dataset Splits Yes We select VGG16 and Efficient Net B0, two different pre-trained feedforward networks on Image Net, and augment them with the proposed predictive coding dynamics. and To this end, we inject additive Gaussian noise to the Image Net validation set, and monitor the models performance across timesteps.
Hardware Specification No The paper states 'See Appendix' for hardware specifications, but the Appendix content is not provided in the given text.
Software Dependencies No The paper mentions 'Py Torch [26]' but does not provide specific version numbers for software dependencies in the main text.
Experiment Setup Yes For both the networks, after training the feedback deconvolution layers, we freeze all of the weights, and set the values of hyperparameters to βn = 0.8, λn = 0.1, and αn = 0.01 for all the encoders/decoders in Equation (2).