Credit Assignment in Neural Networks through Deep Feedback Control

Update: There is a new version of the code base available here, which incorporates both the code for DFC from the paper "Credit Assignment in Neural Networks through Deep Feedback Control", and the code for Strong-DFC from the paper "Minimizing Control for Credit Assignment with Strong Feedback".

This repository is the official implementation of the NeurIPS 2021 submission "Credit Assignment in Neural Networks through Deep Feedback Control".

Install Python packages

All the needed Python libraries can be installed with conda by running:

$ conda env create -f DFC_environment.yml

Running the methods

One can run the various methods on any feedforward fully connected neural network architecture by calling the main.py script and specifying the needed command-line arguments. Run python3 main.py --help for a documentation on all command-line arguments.

To avoid typing all hyperparameters into the command window, one can also define a config file and use the run_config.py script with the specified config file.

$ python3 run_config.py --config_module=configs.configs_mnist.mnist_dfc_ssa

We provide the hyperparameter configurations used for all results in Table 1 of the NeurIPS 2021 submission in the directory configs.

Generating the computer vision results

We used 5 random weight initializations for computing the results in Table 1. To automatically run all 5 random seeds and save the results (together with mean and std) in a .csv file, run the seed_robustness.py script provided with the correct config file.

$ python3 run_seed_robustness.py --config_module=configs.configs_mnist.mnist_dfc_ssa --name=mnist_dfc_ssa.csv

Creating figure 1D

For creating figure 1D, which provides an example of the DFC dynamics, run:

$ python3 run_toy_exp_fig1D.py

This figure will be saved in logs/toy_experiments/fig1D.

Creating figure 2

For creating figure 2, run:

$ python3 run_toy_exp_fig2.py

This figure will be saved in logs/toy_experiments/fig2.

Creating figure 3

For creating figure 3, we need to run all methods separately in pipelines and then run a plotting script. To run the pipelines, run:

$ python3 run_pipeline.py --config_module=configs.toy_experiments.configs_fig3.config_dfc --experiment_name=fig3_dfc --out_dir=logs/toy_experiments/fig3

for all configs in configs/toy_experiments/configs_fig3 (remember to change the config_module and experiment_name accordingly).

Note that the pipeline of DFC also generates Figure S1 in logs/toy_experiments/fig3/figS1. Next, to create Figure 3, which will be saved in logs/toy_experiments/fig3, run:

$ python3 run_toy_exp_fig3.py

Creating figures of Section C.2 and C.3

For creating the feedback learning figures of Section C.2 and C.3, run the following pipeline script:

$ python3 run_pipeline.py --config_module=configs.toy_experiments.configs_figC2C3.fig_S2_linear --experiment_name=fig_S2_linear --out_dir=logs/toy_experiments/figC2C3

for all configs in configs/toy_experiments/configs_figC2C3 (remember to change the config_module and experiment_name accordingly).

Creating the alignment figures of Section E.5

For creating the alignment figures of Sections S5-S12 in the appendices, run the following script with the desired config file (e.g., here is shown for MNIST with trained feedback weights):

$ python3 run_pipeline.py --config_module=configs.pipeline_configs.mnist --experiment_name=mnist

Method names

For historical reasons, the naming of the methods in the code base does not always correspond to the naming in the NeurIPS submission. The table below provides a translation sheet.

NeurIPS Submission	Code Base
DFC	cont
DFC-SS	ss
DFC-SSA	ndi
MN	GNT
DFC	tpdi

To use DFC, provide --network_type=DFC and --grad_deltav_cont. To use DFC-SS, provide --network_type=DFC without --grad_deltav_cont. To use DFC-SSA, provide --network_type=DFC and --ndi.

References

This code base is built further upon the code base of Meulemans et al. 2020, 'A Theoretical Framework for Target Propagation', with the Apache 2.0 license: https://www.apache.org/licenses/LICENSE-2.0