Description:
- Image recognition and segmentation
-
Parallel computing
-
Machine vision
Abstract
USC researchers have developed an oscillatory retinal neuron (ORN) incorporating sensing and oscillation behavior necessary for in-sensor neuronal computing. Comprised of a silicon/graphene photodetector and an active inductor, it can perform many operations in the frequency domain and parallelize the feature map extraction tasks instrumental to convolutional neuronal networks. An ORN based in-sensor neuronal computing architecture can perform machine vision tasks with an energy cost of <20 fJ/OP, potentially going as low as 15 aJ/OP, almost ~10,000x lower than state-of-the-art NVIDIA A100 GPU.
Benefit
-
Can perform in-sensor neuronal computing
-
Can parallelize tasks in the frequency domain
-
Employs ultra-low power computing
Market Application
Brain-inspired in-sensor computing architectures offer a low power alternative to state-of-the-art machine learning systems based on traditional digital computing. Synaptic neuromorphic sensors have shown great promise in performing machine learning related tasks with significantly lower energy costs, however, lack of task parallelization has limited their applications. Neuronal sensors with oscillatory response, similar to biological neurons, can address this limitation by massively parallelizing machine vision tasks which will consequently lower the energy cost of computing by several orders of magnitude.
Other
Stage of Development