Dis-inhibitory neuronal circuits can control the sign of synaptic plasticity

Authors: Julian Rossbroich, Friedemann Zenke

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we demonstrate that our learning rule performs comparable to BP in deep neural networks trained on computer vision benchmarks. [...] Table 1: Test accuracy in % for networks trained with BP or dis-inhibitory feedback control. Reported values are mean stdev (n = 10).
Researcher Affiliation Academia 1 Friedrich Miescher Institute for Biomedical Research, Basel, Switzerland 2 Faculty of Science, University of Basel, Basel, Switzerland
Pseudocode No The paper provides mathematical equations and descriptions of the model and learning rules (e.g., Eq. 9, 10, 11), but it does not include a distinct pseudocode block or algorithm figure.
Open Source Code Yes The code to reproduce our results is publicly available at https://github.com/fmi-basel/disinhibitory-control.
Open Datasets Yes To that end, we implemented a MLP with excitatory-inhibitory microcircuit units. Networks were comprised of either one or three hidden layers and used a parameterized soft rectifier activation function for all units. [...] We trained the networks on MNIST [54] and Fashion-MNIST [55].
Dataset Splits Yes Table S3: Validation accuracy in % for networks trained with BP or dis-inhibitory feedback control. Reported values are mean +/stdev (n = 10). (See Appendix C.3 for full table).
Hardware Specification Yes All numerical simulations were implemented using Python 3.8.10 and were executed on NVIDIA Quadro RTX 5000 GPUs.
Software Dependencies Yes All numerical simulations were implemented using Python 3.8.10 and were executed on NVIDIA Quadro RTX 5000 GPUs. The software stack includes Jax [9], Flax [10], and Diffrax [11].
Experiment Setup Yes Table S1: Hyperparameters used for training networks on the student-teacher task (Fig. 4). [...] Table S2: Hyperparameters used for training networks on computer vision benchmarks. [...] Network weights were initialized using Xavier initialization and trained for 50 epochs with a batchsize of 100. [...] For all experiments, we used the ADAM optimizer with with learning rate 0.001.