Code to run the simulations of the paper:

Giorgia Dellaferrera, Gabriel Kreiman

Presented at ICML 2022: https://proceedings.mlr.press/v162/dellaferrera22a.html

We run the experiments with the following:

Numpy framework (fully connected models): Python 3.9.5, Numpy 1.19.5, Keras 2.5.0

Pytorch framework (convolutional models): Python 3.7.10, Numpy 1.19.2, Pytorch 1.6.0

The notebook Tutorial_PEPITA_FullyConnectedNets_CIFAR-10.ipynb provides a simple tutorial on how to implement and run the PEPITA training scheme for fully connected models. The entire framework is pytorch-based. The settings and results are the same as reported in the paper.

The training for 100 epochs takes approximately 1.5 hours on CPU.

The experiments are run through main.py, which uses functions in functions.py and utils.py. The entire framework is numpy-based and relies on the keras library to load the datasets.

For example, to run PEPITA with the standard settings on the MNIST dataset:

python main.py --exp_name Experiment1 \
    --learn_type ERIN --n_runs 1 --train_epochs 100 \
    --sample_passes 2 --n_samples all --eta 0.1 --dropout 0.9 \
    --eta_decay --mnist --validation --batch_size 64 \
    --update_type mom --w_init he_uniform \
    --build auto --struct uniform --start_size 1024 --n_hlayers 1 --act_hidden relu --act_out softmax

Note that the training scheme for PEPITA is denoted as ERIN (ERror-INput). If you train with PEPITA (ERIN), make sure to use the setting --sample_passes 2, to have for each input two forward passes.

Substitute --learn_type ERIN with --learn_type BP to train the network with backpropagation. Remember to set --sample_passes 1.

The experiments are run through main_pytorch.py, which uses functions in models.py. The entire framework is pytorch-based.

For example, to run PEPITA with the standard settings on the MNIST dataset:

python main_pytorch.py --exp_name Experiment2 \
    --learn_type ERIN --n_runs 1 --train_epochs 100 \
    --eta 0.01 --dropout 0.9 --Bstd 0.05 \
    --eta_decay --dataset mn --batch_size 50 \
    --update_type mom --w_init he_uniform \
    --model Net1conv1fcXL

The argument Bstd defines the standard deviation of the projection matrix. Here we use B instead of F (paper) to denote the projection matrix to avoid confusion with torch.nn.functional.

The notebook plot_compute_slowness.ipynb contains the function to extract the convergence rate as "slowness" parameter.

@InProceedings{pmlr-v162-dellaferrera22a,
  title = 	 {Error-driven Input Modulation: Solving the Credit Assignment Problem without a Backward Pass},
  author =       {Dellaferrera, Giorgia and Kreiman, Gabriel},
  booktitle = 	 {Proceedings of the 39th International Conference on Machine Learning},
  pages = 	 {4937--4955},
  year = 	 {2022},
  editor = 	 {Chaudhuri, Kamalika and Jegelka, Stefanie and Song, Le and Szepesvari, Csaba and Niu, Gang and Sabato, Sivan},
  volume = 	 {162},
  series = 	 {Proceedings of Machine Learning Research},
  month = 	 {17--23 Jul},
  publisher =    {PMLR},
  pdf = 	 {https://proceedings.mlr.press/v162/dellaferrera22a/dellaferrera22a.pdf},
  url = 	 {https://proceedings.mlr.press/v162/dellaferrera22a.html},
}