GLM, regression models, and latent variables

Difficulties experienced in understanding machine learning techniques often stem from lack of clarity concerning more basic statistical models and fundamental considerations, including the various regression models that can all be subsumed under the General Linear Model.

These courses will provide a refresher on the basics of the General Linear Model and various fitting approaches that fall under its umbrella, collectively showing how 'traditional' inferential statistics form the basis for machine learning.

Lessons of this Course

Machine Learning I (Intro Lecture)

Duration:

33:58

Speaker:

Cristina Savin

This lecture is part of the Neuromatch Academy (NMA), a massive, interactive online summer school held in 2020 that provided participants with experiences spanning from hands-on modeling experience to meta-science interpretation skills across just about everything that could reasonably be included in the label "computational neuroscience".

This lecture provides an overview of generalized linear models (GLM) and contains links to 2 tutorials, lecture/tutorial slides, suggested reading list, and 3 recorded question and answer sessions.

Tutorial 1: Generalized Linear Models (GLMs)

Duration:

8:09

Speaker:

Anqi Wu

This tutorial covers Generalized Linear Models (GLMs), which are a fundamental framework for supervised learning. In this tutorial, the objective is to model a retinal ganglion cell spike train by fitting a temporal receptive field: first with a Linear-Gaussian GLM (also known as ordinary least-squares regression model) and then with a Poisson GLM (aka "Linear-Nonlinear-Poisson" model). This tutorial also covers a special case of GLMs, logistic regression, and learn how to ensure good model performance. This tutorial is designed to run with retinal ganglion cell spike train data from Uzzell & Chichilnisky 2004.

Tutorial 2: Classifiers and regularizers

Duration:

6:42

Speaker:

Anqi Wu

This tutorial covers the implementation of logistic regression, a special case of GLMs used to model binary outcomes. Oftentimes the variable you would like to predict takes only one of two possible values. Left or right? Awake or asleep? Car or bus? In this tutorial, we will decode a mouse's left/right decisions from spike train data.

Objectives of this tutorial:

Learn about logistic regression, how it is derived within the GLM theory, and how it is implemented in scikit-learn
Apply logistic regression to decode choices from neural responses
Learn about regularization, including the different approaches and the influence of hyperparameters

Machine Learning II (Outro Lecture)

Duration:

29:30

Speaker:

I. Memming Park

This lecture further develops the concepts introduced in Machine Learning I.

Statistical models for neural data: from GLMs to latent variables (Part 1)

Duration:

1:45:48

Speaker:

Jonathan Pillow

Part 1 of 2 of a tutorial on statistical models for neural data