Traditionally, computational cognitive models have been hand-designed to implement specific hypothesized cognitive mechanisms. For example, Q-learning models have been fit to reinforcement learning behavior, and Bayesian models have been applied to inference. These models are interpretable by design, but often provide an imperfect fit to experimental data. An alternative approach employs artificial neural networks (ANNs), which typically gain predictive power at the expense of being more difficult to interpret. This lecture will highlight the strengths and limitations of both the classic and the ANN approach, and showcase recent research in both traditions. We will also describe new methods that combine the predictive power of ANNs with the interpretability of classic models.
During the tutorial, participants will be working in Python notebooks. The notebooks will have all necessary boilerplate code prepared ahead of time, and participants will have access to a custom kernel with all necessary libraries installed. Code for models (Q-learning, Ideal Observer, DDM, ANN), visualization tools, and other functions (e.g., model fit, loss, training loop) will be included.
The tutorial will begin with models familiar to many members of the CCN community (e.g. Q-learning, Ideal Observers, DDMs). Participants will make plots showing the predictions of underfit and overfit models on heldout data. We will introduce the two notions of quality-of-fit: quantitative (does the model assign high likelihood to held-out data?) and qualitative (does model-generated behavior reproduce the key features of the real dataset). Participants will make plots that illustrate important quantities such as the parameter fits, and the loss landscape and its properties. We will then segue to a very different approach for modeling behavioral data, using ANNs. We will compare the qualitative and quantitative fit achieved by classic models to that achieved by simple ANNs, and participants will readily appreciate how strikingly ANNs outperform classic models. Participants will implement a few methods to interpret ANNs that have been trained to fit data. One type of method involves constructing lower-dimensional projections of the network dynamics. Another approach is applying cognitive science methods to trained neural networks, for example probing their behavior in key experimental settings. If time permits, participants may explore the method of fitting hybrid models that have both neural and hand-crafted components.
Colab can be found here.
And GitHub here.
The materials can all be accessed via the colab. However we recommend downloading the materials from GitHub in case there are wifi limitations.