Latent Variable Sequence Identification for Cognitive Models with Neural Network Estimators
This paper introduces a neural network-based approach that extends neural Bayes estimation to infer time-varying latent variable sequences from both tractable and intractable cognitive models, thereby enabling broader model-based neural analyses and theory testing.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Human behavior is rarely a simple, straight line from stimulus to response. When a person or an animal makes a choice, that decision is often the visible tip of a much larger, invisible process happening inside the mind. Scientists call these hidden processes "latent variables." They are the internal states—like a sudden shift in attention, a changing belief about which action will bring a reward, or a momentary rule the brain is following—that drive what we see on the outside. Understanding these hidden states is crucial for researchers who want to know how the brain supports thinking and learning. For decades, the standard way to uncover these internal states has been to build a mathematical model of the behavior, find the best numbers that fit the model to the data, and then run the model again to see what it says the hidden states were. However, this traditional method hits a hard wall when the models become too complex. Many interesting theories about how the mind works involve calculations that are so difficult to solve that the standard mathematical tools simply cannot find the answers. This has left a vast landscape of cognitive theories unexplored because scientists lacked a way to peek inside the black box of complex behavior.
A team of researchers at the University of California, Berkeley, has developed a new tool to bypass this wall, allowing scientists to see these hidden states even in the most complicated models. They call their method LaseNet. Instead of trying to solve the difficult math equations directly, the researchers trained an artificial intelligence system to learn the relationship between what an animal does and what it is likely thinking. They did this by creating thousands of simulated experiments where they knew the exact answers. In these simulations, they generated data that looked like real behavior—choices, rewards, and mistakes—and paired it with the known hidden states that caused them. They fed this massive library of "what happened" and "what was thinking" into a type of neural network, a computer system designed to recognize patterns in sequences of events. The network learned to look at a sequence of actions and outcomes and predict the hidden mental state that produced them, effectively learning to translate behavior into thought without ever needing to solve the complex underlying math.
The researchers tested this new approach on a variety of cognitive models, ranging from simple learning scenarios to complex, hierarchical decision-making tasks. In cases where the old methods worked, LaseNet performed just as well, accurately tracking the hidden variables. But its true power showed up in the models where the old methods failed or struggled. For models where the math was too difficult to solve directly, the traditional approaches either could not work at all or required extremely slow, custom-built statistical tools that were prone to error. LaseNet, by contrast, handled these difficult cases with ease. In simulations, it identified hidden states with a level of precision that surpassed the best existing techniques, recovering the true internal states with far less error. It successfully tracked both continuous variables, like the changing value of a reward, and discrete variables, like a sudden switch in strategy or attention.
To prove this wasn't just a method that worked on made-up data, the team applied LaseNet to real data collected from mice. In one experiment, mice were navigating a task where rewards changed unpredictably, requiring them to constantly update their internal rules. The researchers used LaseNet to infer the mice's internal attention states and their expectations of reward. The results matched the patterns found by traditional methods where those methods could be applied, but LaseNet provided a clear, continuous picture of the mice's internal states throughout the entire task. In another study involving a visual decision-making task, the method successfully identified when the mice were fully engaged versus when they were biased toward one side, revealing the dynamic shifts in their mental focus. The approach worked for individual animals, meaning it could be used to study the unique cognitive style of a single subject rather than just averaging out the behavior of a whole group.
The significance of this work lies in its ability to open doors that were previously locked. By removing the need for complex, often impossible mathematical calculations, LaseNet allows researchers to test a much wider range of theories about how the mind works. Scientists can now propose models that were previously too difficult to analyze and still get answers about the hidden processes driving behavior. The method is flexible enough to handle different types of data and different kinds of mental states, and it works efficiently on standard computers without requiring massive computing power. While the tool does not yet provide a way to measure the uncertainty of its own guesses, and it relies on the quality of the models used to train it, it represents a major step forward. It transforms the study of cognitive models from a field limited by mathematical tractability into one limited only by the creativity of the theories researchers wish to explore, offering a clearer window into the dynamic, hidden machinery of the mind.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.