A Unified, Scalable Framework for Neural Population Decoding
Authors: Mehdi Azabou, Vinam Arora, Venkataramana Ganesh, Ximeng Mao, Santosh Nachimuthu, Michael Mendelson, Blake Richards, Matthew Perich, Guillaume Lajoie, Eva Dyer
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In a number of different tasks, we demonstrate that our pretrained model can be rapidly adapted to new, unseen sessions with unspecified neuron correspondence, enabling few-shot performance with minimal labels. This work presents a powerful new approach for building deep learning tools to analyze neural data and stakes out a clear path to training at scale. We evaluate the performance of our proposed approach on data from over 158 sessions from open electrophysiology datasets from seven non-human primates (NHPs), spanning over 27,373 units and 100 hours of recordings. We demonstrate that through pretraining on large amounts of data, we can transfer with very few samples (few-shot learning) and thus improve overall brain decoding performance. |
| Researcher Affiliation | Academia | 1 Georgia Tech, 2 Mila, 3 Université de Montréal, 4 Mc Gill University |
| Pseudocode | No | The paper does not contain any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Contact: {mazabou,evadyer}@gatech.edu. Project page and code: https://poyo-brain.github.io |
| Open Datasets | Yes | We curated a multi-lab dataset with electrophysiological recordings from motor cortical regions, where neural population activity has been extensively studied [10], and deep learning tools and benchmarks have recently been established [27]. In total, we aggregated 178 sessions worth of data, spanning 29,453 units from the primary motor (M1), premotor (PMd), and primary somatosensory (S1) regions in the cortex of 9 nonhuman primates (see Table 1). We make these datasets publicly available on the DANDI open repository [29]. |
| Dataset Splits | Yes | For every session, we holdout 20% of the trials for testing, and 10% for validation. |
| Hardware Specification | Yes | The large models are trained on a machine with an AMD EPYC 7452 32-Core Processor and 8 Nvidia A40 GPUs (48Gb memory), POYO-mp was trained for 2 days, and POYO-1 was trained for 5 days (both for a total of 400 epochs). Single-session models are trained with a single Nvidia Ge Force RTX 3090 GPU, and take less than an hour to train. Finetuning models is also done with a single GPU. |
| Software Dependencies | No | The paper mentions using the LAMB optimizer [28] but does not specify version numbers for any software dependencies or libraries (e.g., PyTorch, TensorFlow, Python). |
| Experiment Setup | Yes | We train the model with N = 512 latent tokens and a dimension D = 128. We use the LAMB optimizer [28], and employ a cosine decay of the learning rate at the end of training. For every session, we holdout 20% of the trials for testing, and 10% for validation. We use 1-GPU and 8-GPU setups for single-session and multi-session models, respectively. All details can be found in Appendix A. The model is trained using the LAMB optimizer [28] with weight decay. The learning rate is held constant, then decayed towards the end of training (last 25% of epochs), using a cosine decay schedule. Single-session models are trained on a single GPU with a batch size of 128 while large models are trained with 8 GPUs with a total batch size of 1400. Table 4: Hyperparameter Value Embedding Dimension 128 Head Dimension 64 Number of Latents 128 Depth 6 Number of Heads 8 FFN Dropout 0.3 Linear Dropout 0.3 Attention Dropout 0.3 Weight Decay 1e-4 Learning Rate 4e-3 Batch Size 128 |