Rochester Engineers Develop Analog Networks for Image Recognition

Engineers at the University of Rochester, in collaboration with researchers from Rice University and UCLA, are developing biologically inspired analog hardware based on predictive coding networks to improve the energy efficiency of artificial intelligence systems used in autonomous drones and self-driving cars. Funded by up to $7.2 million from the Defense Advanced Research Projects Agency (DARPA) over 54 months, the research aims to move beyond conventional digital neural networks and their computationally intensive back propagation method, instead replicating the brain’s hierarchical process of prediction and correction. The team intends to build and test digital image recognition systems using existing complementary metal oxide semiconductor (CMOS) technology, initially focusing on static image classification before progressing to more complex perception tasks.

Mimicking Biological Neural Networks

Engineers at the University of Rochester are developing new analog hardware utilising predictive coding networks, an approach inspired by theories concerning how the brain processes visual information and maintains a model of its environment. This development represents a departure from conventional neural networks, which are typically implemented on digital hardware initially designed for general-purpose computing tasks, and prioritises energy efficiency for machine learning applications. The research is motivated by findings from neuroscientists that the back propagation mechanism, commonly used in developing neural networks, is biologically implausible, suggesting the brain employs a different perceptual process.

Predictive coding, as theorised and previously researched – notably with early foundational work by the late Dana Ballard of the University of Rochester – involves a hierarchical system of prediction and correction, mirroring the process of paraphrasing information and refining understanding based on feedback. The project, a collaborative effort including researchers from Rice University and UCLA, has received up to $7.2 million in funding from the Defense Advanced Research Projects Agency (DARPA) over a 54-month period to facilitate the development of these biologically inspired networks for digital image recognition. The initial focus of this research is on classifying static images, with the long-term objective of translating this technology to more complex perception tasks required by autonomous drones and self-driving vehicles, contingent upon achieving comparable performance to existing digital systems.

The proposed analog system will be fabricated using established technologies, specifically complementary metal oxide semiconductor (CMOS), rather than relying on experimental devices, ensuring feasibility and scalability. This approach aims to address the high-power consumption associated with current digital computers used in artificial intelligence systems, whilst leveraging the efficiency of analog circuits for machine learning tasks, as detailed in this latest research article. The team’s success in developing a functional prototype could demonstrate the viability of biologically inspired architectures for advanced perception systems in autonomous applications.

Funding and Future Applications

The funding for this project totals up to $7.2 million from the Defense Advanced Research Projects Agency (DARPA), allocated over a 54-month period, to support the development of biologically inspired predictive coding networks for digital image recognition. These networks will be built on analog circuits, representing a departure from conventional state-of-the-art neural networks developed on digital hardware for computer vision. The initial prototype will focus on classifying static images, with the expectation that, should the analog system achieve performance comparable to existing digital approaches, it can be adapted for more complex perception tasks.

The research article details that the system will not utilise experimental devices in its construction, but will instead be manufactured using existing technologies, specifically complementary metal oxide semiconductor (CMOS). This selection of materials and manufacturing processes is intended to ensure the feasibility and scalability of the proposed analog system, facilitating its potential application in autonomous drones and self-driving cars. The collaborative team leading this development includes researchers from the University of Rochester, Rice University, and UCLA, building upon a history of computer vision research at Rochester, including early work on predictive coding networks by the late Dana Ballard.