Testing Predictive Coding | Dimitris Pinotsis, PhD
McGovern Institute for Brain Research

Testing Predictive Coding | Dimitris Pinotsis, PhD

Add to CalendarAmerica/New_YorkTesting Predictive Coding | Dimitris Pinotsis, PhD11/17/2021 4:00 pm11/17/2021 5:00 pmBuilding 46,46-3189
November 17, 2021
4:00 pm - 5:00 pm
Location
Building 46,46-3189
Contact
Nunziata@mit.edu
Predictive Coding (PC) is a theory of brain function based on ideas from Bayesian inference and machine learning. It suggests that the brain tries to understand the world by generating predictions about the information that should be present in the world. The brain searches for these predictions. In this context, brain activity represents the brain’s predictions and the discrepancies between them and the information the brain receives, called prediction errors. Then, understanding the world amounts to minimizing prediction errors.
 
Despite its successes, whether PC is implemented by the brain is an open question. I will discuss tests that assess evidence in support of PC in brain data. I will use  data from animal and human studies and different brain imaging modalities and tasks to show that the brain seemed to represent predictions of sensory signals.  Also, that  these predictions were different in schizophrenics compared to controls.
 
Then, I will discuss new mathematical tools that allow one to perform new tests of PC and similar theories:

1) Tests bridging scale: hypotheses about the micro scale by analysing non-invasive human M/EEG data obtained at the macro scale. As an illustration, I will show how to test one of the tenets of PC; that deep cortical layer activity represents predictions and superficial activity represents prediction errors. This uses mathematical tools from statistical decision theory.

 

2) Tests about neuronal connections that carry PC signals. As an illustration, I will show that a well-known behavioral effect in psychophysics, known as the oblique effect, can be explained in terms of sparser microscopic connectivity and metabolic efficiency. I will also discuss a potential explanation of representational drift. This uses a combination of mathematical tools from machine learning and dynamical systems.

Time permitting, I will finish with some recent work that applies PC in consumer neuroscience.

Speaker:
Dimitris Pinotsis, PhD
Associate Professor, Centre for Mathematical Neuroscience and Psychology    Department of Psychology , University of London — City     

 

ZOOM: https://mit.zoom.us/j/95382665126

IN PERSON: 46-3189

Upcoming Events