Abstract

Computational models are essential tools for understanding brain function, but inferring their parameters from experimental data remains challenging. Traditional likelihood-based inference is often intractable for realistic models of neural circuits and behavior. In this talk, I present how simulation-based inference (SBI) provides a solution by learning the relationship between parameters and data from simulations.

Talk Outline

  1. Introduction to SBI in Neuroscience

    • Why neuroscience models are challenging for inference
    • The role of simulators in computational neuroscience

  2. Mixed Neural Likelihood Estimation (MNLE)

    • Handling mixed discrete-continuous data from behavioral experiments
    • Application to drift-diffusion models of decision-making
    • Comparison with traditional methods

  3. Connectomics Applications

    • Inferring synaptic connectivity rules from anatomical data
    • Scaling to realistic cortical circuits
    • Validation on electron microscopy reconstructions

  4. Software Tools

    • The sbi package: making SBI accessible
    • Integration with neuroscience simulators
    • Best practices for practitioners

Key Results

  • MNLE achieves comparable accuracy to specialized methods with 6 orders of magnitude fewer simulations
  • Successfully identified connectivity rules in dense cortical reconstructions
  • Open-source tools have been adopted by multiple neuroscience labs

Venue

Bernstein Center for Computational Neuroscience
Humboldt University Berlin
Berlin, Germany

Recording

[Contact for recording availability]