Tobi Delbrück

This paper describes a 128 times 128 pixel CMOS vision sensor. Each pixel independently and in continuous time quantizes local relative intensity changes to generate spike events. These events appear at the output of the sensor as an asynchronous stream of digital pixel addresses. These address-events signify scene reflectance change and have sub-millisecond timing precision. The output data rate depends on the dynamic content of the scene and is typically orders of magnitude lower than those of conventional frame-based imagers. By combining an active continuous-time front-end logarithmic photoreceptor with a self-timed switched-capacitor differencing circuit, the sensor achieves an array mismatch of 2.1% in relative intensity event threshold and a pixel bandwidth of 3 kHz under 1 klux scene illumination. Dynamic range is > 120 dB and chip power consumption is 23 mW. Event latency shows weak light dependency with a minimum of 15 mus at > 1 klux pixel illumination. The sensor is built in a 0.35 mum 4M2P process. It has 40times40 mum <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> pixels with 9.4% fill factor. By providing high pixel bandwidth, wide dynamic range, and precisely timed sparse digital output, this silicon retina provides an attractive combination of characteristics for low-latency dynamic vision under uncontrolled illumination with low post-processing requirements.

Hardware implementations of spiking neurons can be extremely useful for a large variety of applications, ranging from high-speed modeling of large-scale neural systems to real-time behaving systems, to bidirectional brain-machine interfaces. The specific circuit solutions used to implement silicon neurons depend on the application requirements. In this paper we describe the most common building blocks and techniques used to implement these circuits, and present an overview of a wide range of neuromorphic silicon neurons, which implement different computational models, ranging from biophysically realistic and conductance-based Hodgkin-Huxley models to bi-dimensional generalized adaptive integrate and fire models. We compare the different design methodologies used for each silicon neuron design described, and demonstrate their features with experimental results, measured from a wide range of fabricated VLSI chips.

Event-based dynamic vision sensors (DVSs) asynchronously report log intensity changes. Their high dynamic range, sub-ms latency and sparse output make them useful in applications such as robotics and real-time tracking. However they discard absolute intensity information which is useful for object recognition and classification. This paper presents a dynamic and active pixel vision sensor (DAVIS) which addresses this deficiency by outputting asynchronous DVS events and synchronous global shutter frames concurrently. The active pixel sensor (APS) circuits and the DVS circuits within a pixel share a single photodiode. Measurements from a 240×180 sensor array of 18.5 μm 2 pixels fabricated in a 0.18 μm 6M1P CMOS image sensor (CIS) technology show a dynamic range of 130 dB with 11% contrast detection threshold, minimum 3 μs latency, and 3.5% contrast matching for the DVS pathway; and a 51 dB dynamic range with 0.5% FPN for the APS readout.

Deep spiking neural networks (SNNs) hold the potential for improving the latency and energy efficiency of deep neural networks through data-driven event-based computation. However, training such networks is difficult due to the non-differentiable nature of spike events. In this paper, we introduce a novel technique, which treats the membrane potentials of spiking neurons as differentiable signals, where discontinuities at spike times are considered as noise. This enables an error backpropagation mechanism for deep SNNs that follows the same principles as in conventional deep networks, but works directly on spike signals and membrane potentials. Compared with previous methods relying on indirect training and conversion, our technique has the potential to capture the statistics of spikes more precisely. We evaluate the proposed framework on artificially generated events from the original MNIST handwritten digit benchmark, and also on the N-MNIST benchmark recorded with an event-based dynamic vision sensor, in which the proposed method reduces the error rate by a factor of more than three compared to the best previous SNN, and also achieves a higher accuracy than a conventional convolutional neural network (CNN) trained and tested on the same data. We demonstrate in the context of the MNIST task that thanks to their event-driven operation, deep SNNs (both fully connected and convolutional) trained with our method achieve accuracy equivalent with conventional neural networks. In the N-MNIST example, equivalent accuracy is achieved with about five times fewer computational operations.

We present the first gesture recognition system implemented end-to-end on event-based hardware, using a TrueNorth neurosynaptic processor to recognize hand gestures in real-time at low power from events streamed live by a Dynamic Vision Sensor (DVS). The biologically inspired DVS transmits data only when a pixel detects a change, unlike traditional frame-based cameras which sample every pixel at a fixed frame rate. This sparse, asynchronous data representation lets event-based cameras operate at much lower power than frame-based cameras. However, much of the energy efficiency is lost if, as in previous work, the event stream is interpreted by conventional synchronous processors. Here, for the first time, we process a live DVS event stream using TrueNorth, a natively event-based processor with 1 million spiking neurons. Configured here as a convolutional neural network (CNN), the TrueNorth chip identifies the onset of a gesture with a latency of 105 ms while consuming less than 200 mW. The CNN achieves 96.5% out-of-sample accuracy on a newly collected DVS dataset (DvsGesture) comprising 11 hand gesture categories from 29 subjects under 3 illumination conditions.