New Photonic Chip Enables Unlimited Lifelong Learning for AI
-
Proposes L2ONN, a neuromorphic photonic architecture for lifelong learning that can incrementally learn tens of tasks in one model while avoiding catastrophic forgetting.
-
Leverages intrinsic sparsity and parallelism of optics with adaptive photonic neuron connections and multi-spectrum representations for different tasks.
-
Achieves 14x larger capacity and over 10x better efficiency than electronic neural networks while maintaining accuracy.
-
Demonstrates performance on challenging vision, audio, and medical classification datasets, learning up to 14 tasks with 96.3% neuron connections.
-
Fabricates on-chip version that experimentally verifies lifelong learning capability by incrementally implementing multiple tasks.