The Neuroscience and AI Connection

Study Notes

The speaker is David Friedman, a professor in neurobiology and neuroscience at the University of Chicago.
His research focuses on electrophysiological approaches for recording neuronal population activity in awake non-human primates trained to perform complex behavioral tasks.
He also investigates neuronal computations of higher order perceptual and cognitive functions and designs biologically inspired AI approaches.
His research is supported by NIH, NSF, DOD, and private foundations.
He established his lab at the University of Chicago in 2008 and has trained numerous graduate students and post-doc researchers.
He has received several awards, including the Trollent research award from the National Academy of Sciences and the NSF career award.
The speaker is interested in understanding how the brain makes sense of visual scenes and how it guides decisions and actions.
He believes that AI can learn from the brain's mechanisms to enhance its capabilities and flexibility.
Neuroscience research has inspired AI breakthroughs, and AI is accelerating neuroscience research by becoming better models for how the brain works.
The speaker's background is in experimental neuroscience of vision and cognition, and he has been at the University of Chicago for 15 years.
The speaker's lab at University of Chicago focuses on the interface between experimental neuroscience and AI.
The brain interprets signals from the outside world through the visual system, starting with photons entering the eye and triggering electrical spikes to be transmitted to the brain for processing.
The brain processes basic sensory features in the occipital lobe before recognizing the meaning or identity of what is being seen in higher order areas.
The brain then determines the course of action based on factors like current behavioral context, goals, and expected consequences.
Monkeys can be trained to play video games and perform tasks while their brain signals are recorded using multi-electrode arrays.
The lab is studying how the brain learns visual categories and the neural computations underlying categorical decisions.
The parietal cortex is a good candidate for mediating cognitive functions like decision making due to its position between sensory and motor processing and its connections to cognitive networks.
The lab is focusing on a network of brain areas in the parietal cortex.
The lab is decoding large populations of neurons simultaneously recorded in monkeys to understand the format of information encoding in the brain.
The research is a massive decoding problem aimed at understanding how cognitive processes unfold in the brain.
The lateral inter parietal area is important for visually guided actions.
Area Mt is a higher order visual motion processing area that represents the direction of motion in an image.
Monkeys were trained to make decisions about visual motion using a categorization task.
The task required the monkeys to categorize the first stimulus, remember it during a delay period, and compare it to the second stimulus.
Neurons in area Mt encoded the physical features of the stimulus, while neurons in area lip encoded the category membership of the motion directions.
The researchers used artificial neural networks to train on similar tasks and understand how signals are converted from a sensory format to a cognitive format.
The artificial neural networks were trained using back propagation through time and were able to perform the task with high accuracy.
The researchers analyzed the patterns of activity in the artificial networks and found category-selective neural encoding similar to that seen in the monkey brain.
The artificial neural networks provide insight into how signals are processed and converted in the brain.
The researchers hope to use this information to generate a wiring diagram that shows how signals are converted from a sensory format to a cognitive format in the brain.
Neurons in the primate brain show different patterns of activity during a cognitive task.
Artificial neural networks trained on the same task as the monkeys also show similar activity patterns.
The similarity between the activity patterns suggests that the artificial neural networks are finding solutions to the task that have something in common with mechanisms in the primate brain.
Artificial neural networks suffer from catastrophic forgetting, where learning a new task erases knowledge of previous tasks.
A novel approach called context-dependent gating has been developed to alleviate forgetting in artificial neural networks.
Context-dependent gating is inspired by how the primate brain can switch between tasks and learn multiple tasks without forgetting previous ones.
The approach uses top-down connections to gate the activity of artificial neural networks to select which neural ensembles will be recruited to encode each new task or memory.
Context-dependent gating goes a long way toward alleviating forgetting in artificial neural networks and adds almost no computational overhead to the process.
The approach has been tested on a large set of tasks and achieved near-perfect accuracy.
The approach can potentially enhance the representational capacity of a single neural network to store more information and memories.

Description

Test your knowledge on the intersection of neuroscience and artificial intelligence with this quiz based on a lecture by David Friedman, a neurobiology and neuroscience professor at the University of Chicago. Explore how the brain processes visual information, guides decisions and actions, and how AI can learn from the brain's mechanisms. Learn about the research methods used in the lab, including recording neuronal population activity in awake non-human primates and decoding large populations of neurons simultaneously. Test your understanding of artificial neural networks, their strengths, weaknesses,