HebbianLearning

Pytorch implementation of Hebbian learning algorithms to train deep convolutional neural networks. A neural network model is trained on various datasets both using Hebbian algorithms and SGD in order to compare the results. Hybrid models with some layers trained by Hebbian learning and other layers trained by SGD are studied. A semi-supervised approach is also considered, where unsupervised Hebbian learning is used to pre-train internal layers of a DNN, followed by SGD fine-tuning of a final linear classifier. The approach performs better than simple end-to-end backprop training in low sample efficiency scenarios.

We also introduce a preliminary version of the neurolab package, a simple, extensible, Pytorch-based framework for deep learning experiments, providing functionalities for handling experimental configurations, reproducibility, checkpointing/resuming experiment state, hyperparameter search, and other utilities.

This work is a continuation of https://github.com/GabrieleLagani/HebbianLearningThesis and https://github.com/GabrieleLagani/HebbianPCA, containing the latest updates and new Hebbian learning algorithms: Hebbian WTA variants (k-WTA, soft-WTA, ...), Hebbian PCA, Hebbian ICA.

IMPORTANT: the new implementation based on FastHebb is available, making Hebbian experiments more than 20x faster!!

In order to launch an experiment session, type:
PYTHONPATH=<project root> python <project root>/runexp.py --config <dotted.path.to.config.object> --mode <mode> --device <device> --clearhist --restart
Where <dotted.path.to.config.object> is the path, in dotted notation, to a dictionary (which can be defined anywhere in the code) containing the configuration parameters that you want to use for your experiment; <mode> is either train or test; <device> is, for example, cpu or cuda:0; the --clearhist flag can be used if you don't need to keep old checkpoint files saved on disk; and the --restart flag can be used to restart an experiment from scratch, instead of resuming from a checkpoint.

You can also use: PYTHONPATH=<project root> python <project root>/runstack.py --stack <dotted.path.to.stack.object> --mode <mode> --device <device> --clearhist --restart
to run full stack of experiments (i.e. a list, which can be defined anywhere in the code).

Example: python runstack.py --stack stacks.vision.all[cifar10] --mode train --device cuda:0 --clearhist

Author: Gabriele Lagani - gabriele.lagani@gmail.com