Skip to content

AutoLFADS and RADICaL on Google Cloud Platform

This tutorial is a programming-beginner friendly, step-by-step walkthrough on applying AutoLFADS and RADICaL, deep learning tools that can be trained to uncover dynamics from single trial neural population data, using the computational resources of Google Cloud Platform.

Quick introduction

AutoLFADS is the combination of Latent Factor Analysis via Dynamical Systems (LFADS), a deep learning method to infer latent dynamics from single-trial neural population data, with Population Based Training (PBT), an automatic hyperparameter tuning framework. AutoLFADS was originally designed for modeling spiking activity [1,2], and recently extended for modeling EMG activity [3]. RADICaL is a recent adaptation of AutoLFADS for modeling 2-photon calcium imaging data [4]. This tutorial walks through instructions and examples for all three applications.

Specifically, this tutorial focuses on running AutoLFADS and RADICaL on Google Cloud Platform (GCP), which allows the use of these methods using computational resources available for rent on the cloud. Thus, as long as the user has access to GCP and the ability to pay for the usage of GPUs, then this tutorial can be used to apply AutoLFADS and RADICaL to neural population data without any need for local hardware.

Requirements for this tutorial

This tutorial is specifically designed for researchers and scientists with neural population data who are interested in using powerful deep-learning technology to uncover dynamics underlying neural populations. Fundamentally, this tutorial is designed so that significant programming knowledge or experience with deep-learning technology are NOT required.

In order to use this tutorial, it is suggested that:

  • Have neural population data in the format neurons x trial-length x number of trials, with sequence length <100 timesteps.
  • Have access to Matlab and basic familiarity with it
  • Have an email account associated with a university/institution and can pay for the usage of GPUs on the Google Cloud platform. For an estimated rate, a 2.5 hour AutoLFADS or RADICaL run with 8 GPUs (example run in tutorial) will cost ~$7. For more detailed pricing on GPUs, go to https://cloud.google.com/compute/gpus-pricing.

Requesting GPU quota

Google Cloud Platform enforces a GPU quota to prevent unforseen spikes in usage. Requesting additional quota must be done 24-48 hours in advance of a run. If interested in running AutoLFADS or RADICaL through GCP, it is recommended to first request additional GPU quota. Instructions are detailed in the First Time Set-up section.

References

[1] Pandarinath C. et al. Inferring single-trial neural population dynamics using sequential auto-encoders. Nature Methods (2018). Paper link

[2] Keshtkaran MR*, Sedler AR*. et al. A large-scale neural network training framework for generalized estimation of single-trial population dynamics. Accepted in principle at Nature Methods (2022). bioRxiv link

[3] Wimalasena LN. et al. Estimating muscle activation from EMG using deep learning-based dynamical systems models. In revision. bioRxiv

[4] Zhu F. et al. A deep learning framework for inference of single-trial neural population activity from calcium imaging with sub-frame temporal resolution. In revision. bioRxiv link