Neuromorphic Plasticity Chip NeuroPlast: When Chips Learn in Real Time Like the Brain

Article

In November 2028, Beijing's Institute of Brain-Like Intelligence released NeuroPlast — the world's first neuromorphic chip to achieve synaptic plasticity during inference. Unlike conventional neuromorphic chips that merely simulate brain structure at the design stage, NeuroPlast continuously reshapes its connections during operation, much like a biological neural network.

"Current neuromorphic chips are frozen brains — once manufactured, their structure is fixed," explained project lead Professor Liu Ming. "NeuroPlast is a living brain — it learns as it works."

NeuroPlast's core innovation is its reconfigurable analog synapse array. The chip contains 256 analog synapse units, each with connection strengths adjustable in real time through external input. When the chip processes new input patterns, relevant synaptic connections are strengthened while irrelevant ones are weakened — mirroring the Hebbian learning rule in biological brains.

NeuroPlast is manufactured using a 14nm process with 12 million plastic synapse nodes. In continuous learning tests, the chip achieved a 40% speed improvement and 60% energy reduction after processing the same task type 100 times. On the ImageNet continual learning benchmark, NeuroPlast achieved 89% accuracy in learning new categories without forgetting old ones — 25 percentage points higher than conventional GPU solutions.

Power consumption peaks at just 1.2 watts, one-eighth that of a comparable GPU. This makes it particularly suited for deployment in scenarios requiring continuous adaptation — such as autonomous vehicles performing transfer learning across cities, or industrial robots transferring skills between production lines.

SenseTime has announced plans to integrate NeuroPlast into its autonomous driving platform, targeting mass production for vehicles in Q3 2029. Horizon Robotics and Black Sesame Technologies are also evaluating the chip for edge AI applications.

However, NeuroPlast's continuous learning capability introduces new challenges. Because the chip constantly changes its internal state during operation, traditional chip verification methods struggle to ensure predictable behavior. "A chip that is continuously learning — you can't easily specify its behavioral logic at any given moment," commented a chip verification engineer. "This is problematic for safety-critical applications."

The research team is developing formal verification-based behavioral analysis tools for NeuroPlast but acknowledges that deployment in safety-critical scenarios will require significantly more validation time.