STATS320: Statistical Methods for Neural Data Analysis

STATS320: Machine Learning Methods for Neural Data Analysis

Cross-listed as NBIO220 and CS339N Instructor: Scott Linderman Winter Quarter, 2021 Stanford University

Course Description

With modern high-density electrodes and optical imaging techniques, neuroscientists routinely measure the activity of hundreds, if not thousands, of cells simultaneously. Coupled with high-resolution behavioral measurements, these datasets offer unprecedented opportunities to learn how neural circuits function. This course will study statistical machine learning methods for analysing such datasets, including: spike sorting, calcium deconvolution, and voltage smoothing techniques for extracting relevant signals from raw data; markerless tracking methods for estimating animal pose in behavioral videos; state space models for analysis of high-dimensional neural and behavioral time-series; point process models of neural spike trains; and deep learning methods for neural encoding and decoding. We will develop the theory behind these models and algorithms and then apply them to real datasets in the in-class coding labs and final project.

Labs

The course is organized around eight coding labs. Each lab develops a minimal implementation of a state-of-the-art method from scratch in a self-contained Google Colab notebook. These are widely used techniques for analyzing neural and behavioral data, and through the labs you’ll get a deep understanding of how these methods work under the hood:

1. A simple spike sorting algorithm [Solutions (upon request)]
2. Kilosort: Spike sorting by Deconvolution [Solutions (upon request)]
3. CNMF: Calcium deconvolution via constrained NMF [Solutions (upon request)]
4. DeepLabCut: Markerless pose tracking with CNNs [Solutions (upon request)]
5. DeepRetina: Deep encoding models of retinal spike trains [Solutions (upon request)]
6. Kalman Smoothers: Decoding movement from neural data [Solutions (upon request)]
7. MoSeq: Autoregressive HMMs for animal movements [Solutions (upon request)]
8. SLDS: Switching LDS model of neural data [Solutions (upon request)]

Schedule

The lectures develop the theory behind the methods developed in lab. I’ve organized the course into three units: signal extraction, encoding and decoding, and unsupervised modeling of neural and behavioral data. At the end, you’ll work on a final project in which you will use, explore, or extend the techniques studied in class.

Unit I: Extracting biological signals from raw data

Unit II: Encoding and decoding models for neural data

Unit III: Unsupervised models of neural and behavioral data

Final Projects