Recorded 04 November 2024. Sanjoy Dasgupta of the University of California, San Diego, presents "Neural machines: From fruit flies to embedded devices" at IPAM's Naturalistic Approaches to Artificial Intelligence Workshop. Abstract: We study a class of neural machines, biological or human-made, that operate as follows: at each time step, they receive input and, in constant time, respond with an action or output; as part of this process, they are able to read or write to a fixed-size memory. These machines should offer rigorous performance guarantees for various lifelong learning tasks, and be able adapt gracefully to a changing environment. In this talk I will cover two examples: 1. With Saket Navlakha, Chuck Stevens, and others, we have studied the architecture of the fly’s olfactory system and identified data structures and algorithms for similarity search, novelty detection, habituation, and continual learning. We find that these are elegant, have strong mathematical guarantees, and are easily ported to digital computers. 2. With Tajana Rosing, Anthony Thomas, and others, we have explored the construction of low-power embedded devices based on Kanerva’s notion of hyperdimensional computing. These are low-precision machines that allow only elementary arithmetic operations, but with a data representation similar in spirit to the fly’s they are able to provably solve a variety of basic retrieval and information processing tasks. Learn more online at: https://www.ipam.ucla.edu/programs/workshops/workshop-iii-naturalistic-approaches-to-artificial-intelligence/?tab=overview