We are developing neuFlow: a dataflow computer for bio-inspired visual processing aimed at speeding up the computations of holistic vision systems. NeuFlow is developed jointly at NYU’s Computational Learning Lab and Purdue’s e-lab. Our approach is to provide a general-purpose system that can be programmed like a standard PC. Our system is based on a streaming micro-controller and a runtime reconfigurable 2D grid of compute elements. The reconfiguration capabilities are somewhat similar to those of programmable logic devices, with the major difference that reconfiguration can occur at runtime, allowing very diverse computations to be performed on the grid.

Our Dataflow Computer is especially well suited for convolutional neural networks. These networks have been successfully used in many recognition and classification tasks including generic recogntion/online learning, action recogntion, object recognition, face detection, pedestrian detection and robot navigation, and combined with our Dataflow Computer, they can perform in real-time on megapixel-size video streams.
Applications are in robotic vision, micro-robots, UAVs, imaging sensor, wireless phones, and other embedded vision systems that require low cost and high-speed implementations of synthetic vision system capable of recognizing and categorizing objects and scenes in real-time.

Here is a recent talk on our system:

This website gathers all sorts of information about NeuFlow, and more generally about holistic hardware systems and dataflow computing for vision.

Note: this project has been sponsored by Xilinx University program, and several federal grants to Yann LeCun and Eugenio Culurciello. We also thank PicoComputing for designing with us the latest version of neuFlow hardware on V6 550LT.