summary: Using data from the game “Ebb and Flow,” researchers are training a machine learning algorithm to mimic the human ability to switch attention between tasks. This finding sheds new light on cognitive control and may extend our current understanding of disorders characterized by cognitive control deficits, such as bipolar disorder and schizophrenia.
sauce: UT Austin
Scientists use data from supercomputers and games ebb and flow We train a deep learning model that mimics the human behavior of “task switching,” shifting attention from one task to another.
This basic research is important for helping scientists understand cognitive control. Cognitive control involves the basic mental processes that allow us to focus on the task at hand, but also allow us to flexibly disengage from the task when needed. These abilities are taxed by the game Ebb and Flow studied by researchers.
This research may also help us understand diseases in which patients exhibit deficits in cognitive control, such as bipolar disorder and schizophrenia.
During the game, the player uses the arrow keys on the keyboard to indicate which direction the green leaves are pointing and which direction the red leaves are moving. Alternating green and red leaves should train your mastery of this task-switching game. Mental flexibility, as the player must repeatedly shift focus from one task to another.
Postdoctoral researcher Paul Jaffe, collaborating with Professor Russell Poldrak of the Department of Psychology at Stanford University, said:
Jaffe and Poldrack, co-authors of a study that developed a new, more realistic task switching model, nature human behavior January 2023.
Existing models of cognitive processing assemble simple components in a ‘top-down’, rigorous fashion.
“They make a lot of assumptions about how the mind performs tasks, or have other limitations, such as not being able to actually fit data from participants,” Jaffe said. says.
Jaffe and colleagues developed a framework for modeling human behavior in cognitive tasks called task-DyVA. It uses a dynamic neural network that takes task stimuli as inputs and produces task responses as outputs, much like a human would engage in a task.
“The Task DyVA framework not only allowed us to fit the vast amount of Ebb and Flow data available, but also allowed us to model the individual differences of participants,” said Jaffe. For each person’s data he can fit one model and see how the models differ. You can then examine the neural network, the “brain” of the model, to understand how it performs its tasks. ”
The team employed a machine learning algorithm called a variational autoencoder. This is a method developed to handle inference and learning on difficult probabilistic models.
The research team was assigned to the Maverick2 supercomputer at the Texas Advanced Computing Center (TACC). The system is dedicated to machine learning workloads accomplished by a framework utilizing graphics processing units (GPUs) that can utilize his 24-node NVidia GTX 1080 Ti GPUs. , with 4 GPUs in the node, and his 3 nodes with 2 NVidia P100 GPUs each.
“TACC was essential to accomplishing this task due to the availability of GPUs, which are inherently optimized for computing many matrix multiplications very quickly. This is a frequently used operation in deep learning models like the one used in this study,” said Jaffe.
Russell Poldrack said, “GPUs can greatly speed up fitting and testing of machine learning models. Allocating to Maverick2 has allowed us to move forward with this work much more quickly than we would have been without this resource.” .”
The researchers used Maverick2 supercomputing resources and an existing anonymized data set from 140 Ebb and Flow participants aged 20-89 to develop a modeling framework and eventually , asked questions about how the brain performs tasks by analyzing models.
“We looked under the hood of these models to understand how they perform their tasks. The two tasks of the switching task are represented by different regions of the model’s latent space, which is an abstract representation of the variables involved in this particular task.The two different regions of the model’s “brain” are represented by I could see it doing each task,” Jaffe says.
This finding may explain why there is a ‘switch cost’, a slow response when switching between tasks, because it takes time for activity to move from one area of the brain to another. Moreover, this model can explain why dividing these tasks is advantageous for the brain.
This supports the idea of a 2022 study by scientists Musslick and Cohen.
“We found that separating the task into these two different brain regions actually made the model more robust, making it less likely that noise would interfere with the task in each of these brain regions. This allows the brain to perform each task very well without being confused by signals from other tasks,” Jaffe added.
In the future, the scientific team will adapt the model to other tasks and train the model to perform multiple tasks so that people can achieve generalization from their limited experience and perform a huge number of complex tasks. I’m looking into understanding and developing a new model that can explain how to do it. we encounter in our daily life.
For example, fMRI brain scans, a technique employed by the Poldrack Lab, can be fitted to the model to obtain neural and behavioral data. “Then we can start to understand how the brain generates these complex behaviors, which is one of the long-term goals of the task-DyVA framework,” he said. says Jaffe.
Poldrack Lab is currently using Pathways Allocation on TACC’s Frontera supercomputer to process a large number of openly shared fMRI datasets.
Mr Jaffe said: Supercomputing resources like TACC are essential to perform this important work. ”
About this Artificial Intelligence and Cognitive Research News
author: Jorge Salazar
sauce: UT Austin
contact: Jorge Salazar – UT Austin
image: image is public domain
Original research: closed access.
“Modeling Human Behavior in Cognitive Tasks Using Latent Dynamic Systems” by Paul Jaffe et al. nature human behavior
overview
Modeling Human Behavior in Cognitive Tasks Using Latent Dynamics Systems
Response time data collected from cognitive tasks are the cornerstone of psychological and neuroscience research, but existing models for these data either make strong assumptions about the data generation process or are limited to modeling a single trial. is limited to
We introduce Task-DyVA, a deep learning framework that trains expressive dynamic systems to reproduce a set of observed response times in data from individual human subjects. A model fitted to a large task-switching dataset captured subject-specific behavioral differences with high temporal accuracy, including task-switching costs.
Through perturbation experiments and analysis of the latent dynamics of the model, we find that the cost of switching can be reasonably explained in terms of the trade-off between stability and flexibility. Our framework can therefore be used to discover interpretable cognitive theories that explain how the brain dynamically takes action.
