Cornell Tech Demonstrates Optical Receiver That Can Update AI Memory With Light
Researchers at Cornell Tech have demonstrated a new optical receiver design that can update a robot's AI directly via light, presented last month at the IEEE/JSAP Symposium on VLSI Technology & Circuits.
The system works by beaming an array of light—resembling a QR code pattern—onto the receiver. Unlike conventional approaches where light sensors merely capture data that is then processed, this design uses the resulting photocurrents to directly alter the receiver's memory. The optical code can carry AI model parameters rather than just pointing to external data.
According to Jae-sun Seo, an associate professor of electrical and computer engineering at Cornell Tech, the motivation stems from the growing memory demands of AI systems. Modern AI chips often lack space to store all model parameters, forcing designs to rely on external dynamic random-access memory (DRAM). The electrical connections required to shuttle data between DRAM and processors consume significant energy.
The researchers envision the technology could lower energy consumption in data centers, self-driving vehicles, and edge applications such as AI-powered robots, where efficient memory updates are critical for responsive operation.
The work addresses a fundamental challenge in scaling AI systems: as models grow larger, the energy cost of moving data between storage and processing units becomes a limiting factor. By transmitting model parameters optically directly into memory, the approach could streamline this bottleneck.