SIMPL: Scalable and hassle-free optimisation of neural representations from behaviour

Authors: Tom George, Pierre Glaser, Kimberly Stachenfeld, Caswell Barry, Claudia Clopath

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We first validate SIMPL on a dataset of synthetically generated 2D grid cells. Next, we apply SIMPL to rodent electrophysiological hippocampal data (Tanni et al., 2022) and show it modifies the latent space in an incremental but significant way: optimised tuning curves are better at explaining held-out neural data and contain sharper, more numerous place fields allowing for a reinterpretation of previous experimental results. Finally, we apply SIMPL to somatosensory dataset for a monkey performing a centre-out reaching task (Chowdhury et al., 2020). ... To assess performance we track to the log-likelihood of training and test spikes (see Appendix C.5 for how we partition the dataset).
Researcher Affiliation Collaboration Tom M George Sainsbury Wellcome Centre, UCL EMAIL; Pierre Glaser Gatsby Computational Neuroscience Unit, UCL; Kimberly Stachenfeld Google Deep Mind & Columbia University; Caswell Barry Dept. of Cell and Developmental Biology, UCL; Claudia Clopath Sainsbury Wellcome Centre, UCL & Imperial College London EMAIL
Pseudocode Yes Algorithm 1 SIMPL: An algorithm for optimizing tuning curves and latents from behaviour
Open Source Code Yes We provide an open-source JAX-optimised (Bradbury et al., 2018) implementation of our code2. 2Code and a demo can be found at: https://github.com/Tom George1234/SIMPL
Open Datasets Yes Next, we apply SIMPL to rodent electrophysiological hippocampal data (Tanni et al., 2022)... Finally, we apply SIMPL to somatosensory dataset for a monkey performing a centre-out reaching task (Chowdhury et al., 2020).
Dataset Splits Yes To assess performance we track to the log-likelihood of training and test spikes (see Appendix C.5 for how we partition the dataset). ... In our experiments, we used a test fraction of 10% and held out speckled data segments of length 1 second to evaluate the performance of the model.
Hardware Specification No One-hour recordings of 200 neurons (106 spikes) takes 1 minute to run on a CPU laptop. ... Except for GPDM, which required a GPU, all techniques were run and timed on a CPU.
Software Dependencies No We provide an open-source JAX-optimised (Bradbury et al., 2018) implementation of our code.
Experiment Setup Yes Table 1: Hyperparameters settings Dataset v σ dx dt E Artificial Grid Cell Dataset (Fig. 2) 0.4 ms 1 0.02 m 0.02 m 0.1 s 10 Real Hippocampal Dataset (Fig. 3) 1.0 ms 1 0.1 m 0.04 m 0.2 s 10 Motor task dataset (2D) 6 (Fig. 4c&d) 1.0 0.1 0.02 0.05 s 10 Motor task dataset (4D) (Fig. 4e) 1.5 0.09 0.1 0.05 s 10