Video Player

Video by Kevin D Schmidt

Nikolaus Kriegeskorte - Comparing models by their predictions of representational geometries and topologies

Air Force Research Laboratory

May 10, 2024 | 01:23:26

Description: In this edition of QuEST, we will have an extended session with Niko Kriegeskorte on geometric analyses of brain representations

Abstract:

Understanding the brain-computational mechanisms underlying cognitive functions requires that we implement our theories in task-performing models and adjudicate among these models on the basis of their predictions of brain representations and behavioral responses. Previous studies have characterized brain representations by their representational geometry, which is defined by the representational dissimilarity matrix (RDM), a summary statistic that abstracts from the roles of individual neurons (or responses channels) and characterizes the discriminability of stimuli. The talk will cover (1) recent methodological advances implemented in Python in the open-source RSA3 toolbox that support unbiased estimation of representational distances and model-comparative statistical inference that generalizes simultaneously to the populations of subjects and stimuli from which the experimental subjects and stimuli have been sampled, and (2) topological representational similarity analysis (tRSA), an extension of representational similarity analysis (RSA) that uses a family of geo-topological summary statistics that generalizes the RDM to characterize the topology while de-emphasizing the geometry. Results show that topology-sensitive characterizations of population codes are robust to noise and interindividual variability and maintain excellent sensitivity to the unique representational signatures of different neural network layers and brain regions.

Key Moments and Questions in the video include:
Focus on methods development
RSA3 Toolbox: github.com/rsagroup/rsatoolbox
Representational Similarity Analysis version 3
Representational Similarity Analysis
Studying vision
Activity patterns as representations of the stimuli
Neural network model
Representational geometry, representational dissimilarity matrix
Euclidean distance
Representational dissimilarity matrix (RDM)
RDM estimator
Distance from noisy data are positively biased
Two true response patterns
Noisy response patterns
Removing Bias
Square Mahalanobis distance
Crossnobis distance estimator
RDM Comparator
Accounting for dependency among dissimilarity estimates by whitening
Dissimilarity estimation error covariance
Whitened Pearson RDM correlation
Whitened cosine RDM similarity
Topological RSA
Representational geodesics matrix (RGDM)
Turning an RDM into a weighted graph
Distance matrix
Geo-topological matrices
Adjacency matrix
Family of geo-topological distance transforms
Identifying subject’s brain regions
Identifying which layer of a neural network generated the data
RDM estimator
Biased distance estimators
Euclidean distance
Pearson correlation distance
Mahalanobis distance
Poisson-KL estimator
Unbiased:
Crossnobis estimator
Linear-discriminant t
Crossvalidated Poisson-KL estimator
Choosing a combination of RDM estimator and RDM comparator
Flexible RDM models
Data RDM
Selection model
Weighted component model
Manifold model
Fitting and testing in cross validation
Model evaluation
RSA3: new capabilities