Workshop on Computational Cognitive Modeling of Behavioral and Neural Data (August 2017)

Modelling Challenge Winners 2017!

Thanks to everyone who participated in the challenge! We had twelve final models competing for the best spot! Below please see the winners.

Modeling Learning & Choice – Expected Value & Perceptual Uncertainty RL: Avinash Vaidya

Cross-Validation & Modelling Practices – Expected Value & Perceptual Uncertainty RL: Apoorva Bhandari

Modeling Memory – DDM: Milena Rmus


Dates: 8/14 – 8/28

Location: Metcalf 104

Overview: This two-week workshop aims to provide researchers across a variety of fields and backgrounds with the necessary toolkit to use computational models for hypothesis testing and quantitative fitting of behavioral data and brain-behavior relationships. The workshop will be useful both for computation “novices” with limited or no background, as well as those with a more advanced computational background who have either not mastered the science of model selection and parameter estimation, or would like to learn more details on specific classes of models.

Week 1 will provide an introduction to computational modeling, focusing on tips and tricks, as well as pitfalls and perils, and how to judiciously apply models to inform experimental studies of brain and behavior (including choices and response time distributions). To provide concrete guidance we will focus on a few classes of models – RL (reinforcement learning), Bayesian inference, drift diffusion models of perceptual and reward-based decision making, and signal detection models of memory – but the approaches are general and can be applied to other domains. We will also introduce how neural data can be informed by, and can reciprocally inform, the fits to behavioral data. This week will include three sessions of theoretical and hands-on tutorials on how to code these models, quantitatively fit them to the data, perform goodness-of-fit checks, model selection and parameter estimation, and more.

In Week 2 all participants will have the chance to apply these skills as part of a modeling challenge. Participants will be given access to a novel human behavioral dataset at the interface of visual perception, decision making, reinforcement learning and episodic memory (all in one!). You will be able to test your ideas about different aspects of these processes (and their interaction) by constructing models and fitting them to the data. The best and/or most creative model(s) will be awarded prizes, including an iPad as first prize and/or stipend for conference travel. Winners will be decided based on several criteria, including overall model fits to held-out data (more points will be awarded for better performance across multiple domains), but also on other aspects such as innovation and creativity.

Requirements: basic knowledge of Matlab (or sufficient knowledge of another software to code own models without example code), and the ability to commit 3-4 hours a day to the workshop. No previous experience with modeling or probability theory is required, but basic probability would be helpful. For the modeling challenge, participants can use whatever program they want to write their models (e.g., some of the provided toolboxes we will consider are in Python), as long as they provide a matlab-compatible set of parameters and likelihood values for model comparison at the end.

Participation is limited to 20. Please use this form to sign up.

Schedule: This workshop will take place over two weeks (and a final day for announcing challenge winners). Week 1 will consist of two daily sessions of ~1½-2 hours each (possibly broken up by lunch?), and an optional third session with more in-depth discussion and advanced modeling techniques. Week 2 has a flexible schedule: everyone can work on their models whenever they want, but there will be a space provided every morning for anyone who wants to work in a common space with the other participants. There will be a daily one-hour (ish) coffee break to meet and discuss progress and challenges.

Competition Dataset: After the tutorial week, everyone will be given access to the same behavioral dataset. This is a real dataset recently collected on Amazon’s Mechanical Turk, on a learning and memory task that asks how people’s perceptual and reward learning interact to drive decisions, and how various perceptual and reward factors may influence subsequent memory. An example model will be provided, but everyone is free to come up with their own ideas; the ones that best capture behavior, the most original ones etc. will be awarded at the end of the competition. Participants are free to work in teams of two.

Detailed Schedule

Week 1:

  • Day 1: Introduction to computational modeling: why is modeling useful, who can use it and how? Basic overview of different types of models and the main challenges for picking and comparing them.
  • Day 2: Reinforcement Learning (RL) models
    • Session 1: introduction to RL, basic overview, model-based vs model-free learning etc.
    • Session 2: Q-Learning/SARSA/actor-critic: fitting RL models to example data
    • Session 3: RL model comparison; parameter recoverability and goodness-of-fit tests; using RL models to inform neural data analysis
  • Day 3: Bayesian Models
    • Session 1: Quick primer on probability theory and Bayes
    • Session 2: Fitting a simple Bayesian model to (same) example data
    • Session 3: MLE methods, posterior distributions, using posterior predictive checks to test model accuracy
  • Day 4: Drift-diffusion models (DDM)
    • Session 1: Introduction to drift-diffusion models and what they bring to the table compared to the RL and Bayes models
    • Session 2: Fitting a basic DDM to example choice and response time data
    • Session 3: HDDM: using hierarchical Bayesian estimation in DDM to capture both individual and population parameters
  • Day 5: Model comparison and Recap
    • Session 1: overview and discussion of the different models and their advantages and disadvantages
    • Session 2: model comparison: methods, challenges; in-depth discussion on interpreting parameters; relative and absolute goodness-of-fit checks;

Week 2:

  • Day 1: Introduction to the modeling competition: going over the data set, the task, potential open questions, setting the schedule for the rest of the week etc.
  • Days 2 – 5: Individual/Collaborative work on own models. Participants will be provided with a space to work in, but not required to be there (can work on their models anywhere!). There will be a daily coffee break around 3 pm, during which we’ll all have a chance to talk about how our modeling is going and brainstorm ideas and solutions over (sponsored) coffee.
  • Day 8 (technically week 3): Everyone will do a quick (1-2 minutes) overview of their proposed models, and the winning model(s) will be announced!