Description:
- Computing
- Neural networks
- Artificial Intelligence
Abstract
USC researchers propose a novel edge intelligence computing paradigm that employs in-pixel, massively parallel analog computing. The system utilizes monolithically or heterogeneously integrated memory and image sensors based on multi-bit, multi-channel memory embedded pixels. This paradigm would increase power, performance, and area benefit compared to traditional computing structures, allowing for mapping neural networks on sensors for quick decision making for a wide range of complex machine learning tasks.
Benefit
- Improved power, performance, and area benefit
- Enables massively parallel analog computing with sensor
- Reduces data density
- Data transfer without compromising for accuracy
- In-pixel processing
- Multi-object tracking
Market Application
Traditionally, pixel, memory, and computing elements are separate entities in a vision sensor. This separation can compromise size, weight, and power and lead to bandwidth, data processing, and switching speed bottlenecks. In this arrangement, the data produced by a sensor must be transmitted to a remote computing chip for analysis and decision making, leading to limited throughput, excessive energy consumption for data transfer, and data security concerns. As modern AI computations become increasingly server-centric, methods to enable intelligent distributed computation on edge devices like sensor chips to cloud-based servers are needed.
Stage of Development
- Simulation tested
- Available for exclusive and non-exclusive licensing