Skip to main content
Search papers, people, events & news Esc to close
Online · Talks and panel discussions

Interpretability in Psychological Models

Four talks and four panel discussions on what it means for a psychological model to be interpretable, and whether the idea still holds in an era of machine-learned cognition.

When
15 October 2026 Thursday
Time

21:30 Hong Kong · 14:30 London · 09:30 New York · 06:30 Los Angeles · 00:30 Melbourne (+1)

Where
Online — Microsoft Teams
Register Now

Free and open to all Add to calendar

I

About

Computational models can serve many purposes in psychological and cognitive science. They can describe patterns in data, formalise theories, and generate predictions. Researchers often distinguish between normative, descriptive, and process models; between models designed to explain, predict, or reproduce behaviour. But these categories do not tell us what makes a model interpretable, or whether interpretability means the same thing across modelling traditions.

Is interpretability a single property that a model possesses or lacks? Does it lie in the meaning of a model's parameters, the transparency of its functional form, its relation to psychological constructs or external measures, or something else? And how does what counts as interpretable depend on who is using the model, for what purpose, and at what level of description? These questions arise across models ranging from evidence accumulation and Bayesian approaches to cognitive architectures and large neural networks. This event explores these different ways of thinking about interpretability and how they might help us evaluate and use models in psychological science.

Questions in play

01

Model class

Are process models inherently more interpretable than descriptive or machine-learned models?

02

Parsimony

Does parsimony make a model easier to understand, or can a simple model still conceal consequential assumptions?

03

Distinctions

How should interpretability be distinguished from explanation, prediction, identifiability, psychological plausibility, scope, and generalisability?

04

Understanding

When a model is opaque, can interpretation methods provide genuine understanding, or only a more accessible description of its behaviour?

The format

Four short talks will open different perspectives on model interpretability, each followed by a structured panel discussion. The talks provide a starting point, while the panels are encouraged to broaden the conversation and bring in their own questions and perspectives.

The aim

Rather than seeking a single definition of interpretability, the event aims to clarify where different modelling traditions agree and disagree, and to develop a more useful shared language for discussing interpretability in psychological science.

II

Programme

Four short talks, each followed by a 30-minute guided discussion with its panel.

Time · CEST Programme
Welcome
Fusing neural networks with process models to aid interpretability

Adam Sanborn · Talk 15 min · Discussion 30 min

Finding the core of human cognition

Marcel Binz · Talk 15 min · Discussion 30 min

Break
Interpretability as a bridge to understanding

Adrian Erasmus · Talk 15 min · Discussion 30 min

On models, prediction, and scientific roles thereof

Olivia Guest · Talk 15 min · Discussion 30 min

Audience Q&A
Closing
1 Session

Fusing neural networks with process models to aid interpretability

Cognitive models have been limited to operating on researcher-defined feature spaces, but advances in neural networks allow for models that can “see” complex stimuli. However, neural networks trained to map complex stimuli to human responses are difficult to interpret, despite advances in techniques for interpreting network activations. Here I discuss approaches for fusing process models with neural networks to enhance interpretability, including process models that operate on neural-network-defined similarity spaces and in particular on those that train the neural network on the ground truth and add plausible perceptual or decision processes to produce human-like responses.

In discussion

2 Session

Finding the core of human cognition

What are the core components of human cognition? To approach this question, we build large-scale cognitive models that capture human behavior across hundreds of experiments. The winning model is based on a small set of interpretable principles: a phoneme-level input representation, a Hebbian-like episodic memory that stores prediction errors instead of raw content, a context-dependent forgetting mechanism, a selective output gate that controls when memory is allowed to guide behavior, and hierarchical application of these principles. We find that these simple components are sufficient to outperform larger black-box models and to reproduce 73% of the effects found in human behavior.

In discussion

3 Session

Interpretability as a bridge to understanding

This talk offers an account and typology of interpretation as the bridge from one difficult- or impossible-to-understand explanation of a model to another, hopefully more understandable explanation of that model. Through describing the different types of interpretation methods applied in computational cognitive science, this account places understanding at the center of discussions of interpretability and delineates the kind(s) of understanding we should expect. I argue that interpretation is limited in that it cannot provide explanatory understanding, the kind of understanding commonly ideally sought for through applying such methods.

In discussion

4 Session

On models, prediction, and scientific roles thereof

To interpret models, we can benefit from answering questions such as: What are models and what role do they play in cognitive theorising? I will present a perspective that diverges from the current mainstreams of our fields. Models are not a container for observations through being fit to data and should not be held to the standard of providing us with quantitative predictions. I will also sketch out an evaluative account of modelling to inter alia protect against confusions with other model types, such as with inferential statistical models of the data, and to guard against inferring success prematurely. Ultimately, modelling can only play its unique role of mediating from theory to data and back, if and only if we buttress its special scientific status.

In discussion

III

Attending

Registration

Anyone interested in the topic is welcome to attend. Registration is open; the joining link is emailed to you as soon as you register.

Audience Q&A

The closing Q&A is open to the whole audience. Questions can be put to any of the speakers or panels from the afternoon.

Recording

The session will be recorded. For the moment the recording is for the coordinating team's own reference; we would ask the speakers and panel members first before sharing any of it more widely.

← Back to All Events