Where Interpretability Works and Where It Breaks

Sensorimotor Modeling from Population to Individual Scale

Gil Raitses

Syracuse University

Studies

Kernel Structure
ORIGINAL STUDY

An Analytic Hazard Kernel for Optogenetically-Driven Larval Reorientation

Population-level modeling of sensorimotor dynamics using a gamma-difference kernel with two timescales.

HTML PDF Supplement
Clustering Illusion
FOLLOW-UP STUDY

Individual-Level Phenotyping Validation

Why phenotyping fails at the individual level and what experimental changes would fix it.

HTML PDF Suppl HTML Suppl PDF
Presentation Title
PRESENTATION

Mirna Lab General Meeting

47-slide presentation with supplemental materials.

View Slides PDF

Presentation Slides

Slide 1 Slide 2 Slide 3 Slide 4 Slide 5 Slide 6 Slide 7 Slide 8 Slide 9 Slide 10 Slide 11 Slide 12 Slide 13 Slide 14 Slide 15 Slide 16 Slide 17 Slide 18 Slide 19 Slide 20 Slide 21 Slide 22 Slide 23 Slide 24 Slide 25 Slide 26 Slide 27 Slide 28 Slide 29 Slide 30 Slide 31 Slide 32 Slide 33 Slide 34 Slide 35 Slide 36 Slide 37 Slide 38 Slide 39 Slide 40 Slide 41 Slide 42 Slide 43 Slide 44 Slide 45 Slide 46 Slide 47
1 / 47
Download PDF

Figure Gallery

Kernel Structure
Kernel Structure Gamma-difference kernel with fast excitatory and slow suppressive components
Habituation Dynamics
Habituation Dynamics Turn fraction increases across LED pulses in all four conditions
Behavioral States
Behavioral State Analysis Fractional state usage showing shift from running to turning
LOEO Validation
LOEO Validation Leave-one-experiment-out cross-validation results
Factorial Design
Factorial Analysis Effects of intensity and temporal pattern on kernel parameters
Population EDA
Population Summary Dataset overview with track completeness and effect sizes
Identifiability
Identifiability Problem Fisher Information analysis showing burst design advantage
Stimulation Protocols
Stimulation Protocols Comparison of continuous vs burst LED designs
Simulation Design
Simulated vs Empirical Validation of generative model against empirical data
Recommendations
Protocol Recommendations Data gap analysis and protocol modification options
Power Analysis
Power Analysis Required events per track for reliable phenotyping
Design Comparison
Design Comparison Event yield and estimation error across protocols
Clustering Illusion
The Clustering Illusion PCA shows apparent clusters that fail round-trip validation
Data Sparsity
Data Sparsity Event counts per track are insufficient for 6-parameter estimation
Hierarchical Shrinkage
Hierarchical Shrinkage Bayesian regularization pulls individuals toward population mean
Fast Responders
Genuine Outliers Only 8.6% of tracks show variation exceeding measurement noise
Validation Journey
Validation Journey From apparent phenotypes to continuous variation
Model Comparison
Model Comparison Gamma-difference vs raised cosine kernel forms

Key Findings

Population Success

Gamma-difference kernel accurately models population-level reorientation dynamics with fast excitation at 0.3s and slow suppression at 4s.

Individual Failure

Individual phenotyping fails due to sparse data averaging 25 events per track versus 100 required for reliable estimation.

Design Insight

Burst stimulation extracts 10 times more Fisher Information per event than continuous protocols.

Path Forward

Protocol modification with burst trains and extended recording would enable individual phenotyping.