NeuroAI Model Links Working Memory and Efficient Learning
Summary: Modern AI systems can read, converse, and analyze large datasets, yet they still face important limitations. Researchers in neuroAI have developed a new model inspired by the brain’s efficiency that enables more local, real-time synaptic adjustments to improve learning and reduce data movement.
The proposed approach lets individual artificial neurons receive feedback and update themselves on the fly, rather than waiting for entire network layers to be adjusted at once. By reducing the distance data must travel and allowing updates to happen continuously, this design improves energy efficiency and suggests a closer alignment between working memory and synaptic learning processes.
Key Facts:
- Inspired by the Brain: The model draws on biological principles to minimize data movement and mirror how brains update connections continuously.
- Real-Time Adjustment: Individual AI “neurons” receive feedback and adjust locally, enabling faster and more efficient learning.
- Potential Impact: This work points toward a new generation of AI that learns in a more brain-like, energy-efficient manner, strengthening the exchange between neuroscience and AI.
Source: CSHL
AI today appears remarkably capable—it reads, it talks, and it synthesizes vast amounts of data—but significant challenges remain.
“As impressive as ChatGPT and current AI systems are, they remain limited when interacting with the physical world and often require billions of examples to perform well on tasks such as math or writing,” explains Cold Spring Harbor Laboratory NeuroAI Scholar Kyle Daruwalla.
Looking for alternatives to the current computational model, Daruwalla and collaborators sought inspiration from the most energy-efficient computation system we know: the human brain. They focused on the primary inefficiency in modern computing—excessive data movement across network layers—and designed a learning rule that reduces that cost.
Traditional deep networks rely on back-propagation, which distributes error signals across many layers and requires large batches of data and considerable energy. The new method uses a three-factor Hebbian-style update informed by an information bottleneck perspective. Instead of propagating precise error signals backward through the entire network, each layer is trained in a way that is implicitly coordinated through feedforward connections and an auxiliary memory mechanism.
Crucially, the design allows synapses to be adjusted individually and continuously. “In our brains, connections change and adapt all the time,” Daruwalla notes. “It’s not that you pause everything, make adjustments, then resume. Learning and updating are ongoing.” This continuous, local updating reduces the need for global synchronization and shortens the distance data must travel, which can significantly lower energy consumption during training and deployment.
Beyond efficiency gains, the model provides a concrete link between working memory and synaptic updates. Working memory—our ability to hold and manipulate information over short periods while performing tasks—has long been theorized to support learning, but few computational rules have explicitly tied memory resources to synaptic change. The proposed rule creates that connection by using an auxiliary memory network that aggregates information across samples and informs local synaptic updates.
The auxiliary memory can be trained independently of the main network and does not require processing large batches simultaneously, making the approach compatible with neuromorphic hardware and spiking neural networks (SNNs), which simulate biologically realistic neuron behavior. Initial experiments show performance comparable to standard baselines on image classification tasks while revealing how memory capacity affects learning outcomes.
This perspective frames each network layer as balancing memory-informed compression against task performance. It naturally incorporates important aspects of neural computation—locality, memory, and efficiency—into a unified learning rule. If further validated and implemented on specialized hardware, the approach could help develop AI systems that learn more like humans, consuming far less energy and becoming more accessible for practical deployments.
About this artificial intelligence research news
Author: Sara Giarnieri
Source: CSHL
Contact: Sara Giarnieri – CSHL
Image: Image credited to Neuroscience News
Original Research: Open access. “Information bottleneck-based Hebbian learning rule naturally ties working memory and synaptic updates” by Kyle Daruwalla et al., Frontiers in Computational Neuroscience.
Abstract
Information bottleneck-based Hebbian learning rule naturally ties working memory and synaptic updates
Deep feedforward neural networks achieve strong performance across many domains but incur high energy costs during training and inference. Spiking neural networks (SNNs), which model biologically realistic neurons, promise greater efficiency on neuromorphic hardware when trained effectively. A major obstacle is that standard back-propagation is biologically implausible and difficult to implement directly on neuromorphic platforms.
Recent approaches sidestep back-propagation by training layers more independently using information bottleneck principles. These methods yield three-factor Hebbian updates where a global signal modulates local synaptic changes. However, the global signal typically requires combining multiple samples simultaneously, whereas biological systems process one sample at a time.
The authors propose a new three-factor update in which an auxiliary memory network captures cross-sample information for the global signal. This auxiliary network can be trained a priori, independently of the dataset used by the primary network. Experiments demonstrate comparable classification performance to conventional baselines and reveal an explicit relationship between working memory capacity and learning performance—a link not present in back-propagation-like schemes.
These results suggest a different view of learning in which each layer balances memory-informed compression with task objectives, bringing together memory, efficiency, and locality as central elements of neural computation.