With its ability to read, converse, process vast amounts of data, and provide business recommendations, modern artificial intelligence is more human-like than ever before. However, AI still has some significant limitations.
Kyle Daruwala, a NeuroAI researcher at Cold Spring Harbor Laboratory (CSHL), points out that while current AI technologies like ChatGPT are impressive, they still have limitations when it comes to interacting with the physical world: Even tasks like solving math problems or writing essays require billions of training examples to perform well.
Daruwala has been exploring unconventional ways to develop AI that can overcome these computational challenges, and he may have identified a promising approach.
Data movement is key. Currently, most of the energy used in modern computing comes from transmitting data. Artificial neural networks made up of billions of connections often require data to travel long distances. So Daruwala wanted to take inspiration from the human brain, which is known for its immense computational power and energy efficiency.
Inspired by the way the human brain absorbs new information, Daruwala devised a new way for AI algorithms to efficiently move and process data. This new design allows individual AI “neurons” to adapt and receive feedback in real time, rather than waiting for an entire circuit to update at the same time. As a result, data doesn't have to travel as far and can be processed instantly.
“In our brains, connections are constantly changing and adjusting.” Daruwala says: “It's not like you just pause everything, adjust, and then become yourself again.”
A new machine learning model supports a previously untested hypothesis linking working memory to academic achievement and learning. Working memory is the mental system that enables you to focus on a task while retrieving stored information and past experiences.
“In neuroscience, there are theories about how working memory circuits drive learning, but nothing as concrete as our rules that actually link the two. And that was one of the good things we stumbled upon here. The theory led to the rule that in order to regulate each synapse individually, this working memory needs to exist alongside it.” Daruwala says:
Daruwalla's design could usher in a new era of AI that learns like humans do. This advancement would not only improve AI's efficiency and accessibility, but it would also represent a landmark moment for neuro-AI. Long before ChatGPT spoke its first digital word, neuroscience has provided valuable information for AI, and AI may soon reciprocate.
Journal References:
- Kyle Daruwala and Mikko Lipasti. Information bottleneck-based Hebbian learning rules naturally link working memory and synaptic updating. Frontiers in Computational Neuroscience, 2024. DOI: 10.3389/fncom.2024.1240348
