Online Learning Of Neural Computations From Sparse Temporal Feedback

Authors: Lukas Braun, Tim Vogels

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

Reproducibility Variable Result LLM Response
Research Type Experimental The results are separated into a theoretical and an empirical part. In the theoretical part, we derive the partial derivatives (sec. 3.1) that are required for the parameter update rule (eq. 1) and motivate their event-based versions (sec. 3.2 and derivation in appendix E). In the empirical part, we first show that the EDS rule can recover teacher s parameters in LIF and LRF neurons (sec. 3.3). Subsequently, we investigate the influence and functionality of the EDS factor and compare it to the Super Spike surrogate gradient [23] in a lesion study (sec. 3.4). Finally, we test the robustness of the algorithm to temporal noise in the teacher signal (sec. 3.5).
Researcher Affiliation Academia 1. Department of Experimental Psychology, University of Oxford, Oxford, United Kingdom 2. Bernstein Center for Computational Neuroscience Berlin, Berlin, Germany 4. Institute of Science and Technology Austria, Klosterneuburg, Austria
Pseudocode No The paper does not include a clearly labeled 'Pseudocode' or 'Algorithm' block, nor does it present structured steps in a code-like format.
Open Source Code Yes C++ code to replicate all simulations and plots is publicly available1 under GPLv3 license and uses the MPL 2.0 licensed Eigen software library v3.3.7 [40]. 1https://github.com/lukas-braun/learning-neural-computations
Open Datasets No The paper uses a 'teacher-student paradigm' where a teacher neuron generates target spike times, and input spike trains are homogeneously Poisson distributed. This describes a data generation process rather than the use of a pre-existing publicly available dataset.
Dataset Splits No The paper describes a continuous online learning process using a teacher-student paradigm, and thus does not refer to traditional dataset splits like training, validation, and test sets with specific percentages or counts.
Hardware Specification Yes For example, 30 independent simulations of 12,000 minutes of simulated time at 1ms temporal resolution (as in Fig. 3), take less than 30 minutes of real time on a single AMD Ryzen 5950x.
Software Dependencies Yes C++ code to replicate all simulations and plots is publicly available1 under GPLv3 license and uses the MPL 2.0 licensed Eigen software library v3.3.7 [40].
Experiment Setup Yes To speed up learning, we use the Adam optimiser [33] with default parameters (β1 = 0.9, β2 = 0.999, ϵ = 1e 8) and per parameter learning rates, which are scaled according to the range from which target values are sampled (ηw = 35e 6, ητs = 7e 4, ητm = 28e 4 , ηVr = 7e 5 and η w = 8e 5, ηb = 15e 6, ηω = 33e 7, η Vr = 8e 5 η Ir = 8e 5).