Summary of Next-generation recurrent network models for cognitive neuroscience

00:00:00 - 00:50:00

This video discusses next-generation recurrent network models for cognitive neuroscience. The models are designed to address the limitations of current models, which are based on a single task. The recurrent neural networks used in this paper allow for the modeling of complex behavior, and the paper shows that these networks can be used to match data to models. This work is still in its early stages, but it suggests that recurrent neural networks could be a more naturalistic way to study cognition.

• 00:00:00 Robert Young, a new assistant professor at the Mit School of Brain and Cognitive Science, will discuss his work using recurrent neural networks in cognitive neuroscience. He believes that these models have a number of advantages over traditional computational models, but they also have disadvantages.
• 00:05:00 The recurrent neural networks (RNNs) used in this paper are helpful for cognitive neuroscience because they allow us to model complex behavior, such as the activity observed in prefrontal cortex. The paper also shows that these networks can be used to match data to models, which is helpful for cognitive neuroscience research.
• 00:10:00 This video discusses a new approach to cognitive neuroscience that uses recurrent neural networks. The approach is aimed at moving beyond the limitations of current models, which are based on a single task. The research was done by a group of scientists who collaborated with Maddie Javakar, David Sasilo, and Xiao Jinguang. They used a recurrent neural network to train on tasks that are commonly used in cognitive neuroscience research, including memory, decision making, and multi-sensory integration. While this work is still in its early stages, it suggests that recurrent neural networks could be a more naturalistic way to study cognition.
• 00:15:00 The video discusses next-generation recurrent network models for cognitive neuroscience. The models are trained using simple stochastic gradient descent, and task variance is used to quantify engagement in different tasks. Results show that clusters of neurons correspond to causal modules, and different activation functions produce different results.
• 00:20:00 The video discusses next-generation recurrent network models for cognitive neuroscience. The models are designed to circumvent the "catastrophic forgetting" problem, in which previously learned information is forgotten when the network is used for a different task. The models are also able to quantify the impact of continued learning on the neural representation.
• 00:25:00 Next-generation recurrent network models for cognitive neuroscience focus on the task variance for one task minus the test variance for another task, which gives a number between -1 and 1 for each neuron. This data can be plotted to create a distribution, which can be compared to results from prefrontal cortex data recorded from monkeys doing very similar tasks. It appears that at least in these two tasks, the data is more consistent with the network trained with continual learning.
• 00:30:00 The video presents a hypothesis that working memory is more active when it needs to be manipulated, and demonstrates how recurrent networks can be used to test this hypothesis.
• 00:35:00 The authors discuss how the next-generation recurrent network models for cognitive neuroscience can be used to understand how a rat's behavior changes based on its experience on a decision making task. They present results from different rats, showing that some rats tend to repeat the correct choice more often, while others tend to alternate between the correct and incorrect choices. When training a network on this task, the rats do not exhibit this behavior.
• 00:40:00 This video discusses how to build more sophisticated recurrent networks for cognitive neuroscience, using examples of tasks that are more difficult for animals to perform. The video concludes by discussing some of the scientific questions that remain unanswered for networks of this type.
• 00:45:00 This video discusses how to build models that can capture the diversity of brain areas, and how to balance machine learning with science. One idea is to have area specific long-range connectivity, and another is to use readout functions that are similar to the actual areas in the brain.
• 00:50:00 The video discusses next-generation recurrent network models for cognitive neuroscience, discussing the importance of quantitative metrics and intellectual insights. It also mentions the work of Francis Son, Xi Jing, Nick Moss, and David Soselow.