/pdf/neural-networks-and-deep-learning-3ccs3ain8p.pdf

Neural Networks and Deep Learning

The new era of cognitive computing brings forth the grand challenge of developing systems capable of processing massive amounts of noisy multisensory data. This type of intelligent computing poses a set of constraints, including real-time operation, low-power consumption and scalability, which require a radical departure from conventional system design. Brain-inspired architectures offer tremendous promise in this area. To this end, we developed TrueNorth, a 65 mW real-time neurosynaptic processor that implements a non-von Neumann, low-power, highly-parallel, scalable, and defect-tolerant architecture. With 4096 neurosynaptic cores, the TrueNorth chip contains 1 million digital neurons and 256 million synapses tightly interconnected by an event-driven routing infrastructure. The fully digital 5.4 billion transistor implementation leverages existing CMOS scaling trends, while ensuring one-to-one correspondence between hardware and software. With such aggressive design metrics and the TrueNorth architecture breaking path with prevailing architectures, it is clear that conventional computer-aided design (CAD) tools could not be used for the design. As a result, we developed a novel design methodology that includes mixed asynchronous–synchronous circuits and a complete tool flow for building an event-driven, low-power neurosynaptic chip. The TrueNorth chip is fully configurable in terms of connectivity and neural parameters to allow custom configurations for a wide range of cognitive and sensory perception applications. To reduce the system’s communication energy, we have adapted existing application-agnostic very large-scale integration CAD placement tools for mapping logical neural networks to the physical neurosynaptic core locations on the TrueNorth chips. With that, we have successfully demonstrated the use of TrueNorth-based systems in multiple applications, including visual object recognition, with higher performance and orders of magnitude lower power consumption than the same algorithms run on von Neumann architectures. The TrueNorth chip and its tool flow serve as building blocks for future cognitive systems, and give designers an opportunity to develop novel brain-inspired architectures and systems based on the knowledge obtained from this paper.

TrueNorth: Design and Tool Flow of a 65 mW 1 Million Neuron Programmable Neurosynaptic Chip

Guided by brain-like ‘spiking’ computational frameworks, neuromorphic computing—brain-inspired computing for machine intelligence—promises to realize artificial intelligence while reducing the energy requirements of computing platforms. This interdisciplinary field began with the implementation of silicon circuits for biological neural routines, but has evolved to encompass the hardware implementation of algorithms with spike-based encoding and event-driven representations. Here we provide an overview of the developments in neuromorphic computing for both algorithms and hardware and highlight the fundamentals of learning and hardware frameworks. We discuss the main challenges and the future prospects of neuromorphic computing, with emphasis on algorithm–hardware codesign. The authors review the advantages and future prospects of neuromorphic computing, a multidisciplinary engineering concept for energy-efficient artificial intelligence with brain-inspired functionality.

Towards spike-based machine intelligence with neuromorphic computing.

Deep learning in spiking neural networks

Deep neural networks such as Convolutional Networks (ConvNets) and Deep Belief Networks (DBNs) represent the state-of-the-art for many machine learning and computer vision classification problems To overcome the large computational cost of deep networks, spiking deep networks have recently been proposed, given the specialized hardware now available for spiking neural networks (SNNs) However, this has come at the cost of performance losses due to the conversion from analog neural networks (ANNs) without a notion of time, to sparsely firing, event-driven SNNs Here we analyze the effects of converting deep ANNs into SNNs with respect to the choice of parameters for spiking neurons such as firing rates and thresholds We present a set of optimization techniques to minimize performance loss in the conversion process for ConvNets and fully connected deep networks These techniques yield networks that outperform all previous SNNs on the MNIST database to date, and many networks here are close to maximum performance after only 20 ms of simulated time The techniques include using rectified linear units (ReLUs) with zero bias during training, and using a new weight normalization method to help regulate firing rates Our method for converting an ANN into an SNN enables low-latency classification with high accuracies already after the first output spike, and compared with previous SNN approaches it yields improved performance without increased training time The presented analysis and optimization techniques boost the value of spiking deep networks as an attractive framework for neuromorphic computing platforms aimed at fast and efficient pattern recognition

Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing

Deep-learning neural networks such as convolutional neural network (CNN) have shown great potential as a solution for difficult vision problems, such as object recognition. Spiking neural networks (SNN)-based architectures have shown great potential as a solution for realizing ultra-low power consumption using spike-based neuromorphic hardware. This work describes a novel approach for converting a deep CNN into a SNN that enables mapping CNN to spike-based hardware architectures. Our approach first tailors the CNN architecture to fit the requirements of SNN, then trains the tailored CNN in the same way as one would with CNN, and finally applies the learned network weights to an SNN architecture derived from the tailored CNN. We evaluate the resulting SNN on publicly available Defense Advanced Research Projects Agency (DARPA) Neovision2 Tower and CIFAR-10 datasets and show similar object recognition accuracy as the original CNN. Our SNN implementation is amenable to direct mapping to spike-based neuromorphic hardware, such as the ones being developed under the DARPA SyNAPSE program. Our hardware mapping analysis suggests that SNN implementation on such spike-based hardware is two orders of magnitude more energy-efficient than the original CNN implementation on off-the-shelf FPGA-based hardware.

Spiking Deep Convolutional Neural Networks for Energy-Efficient Object Recognition

A safety monitoring system for a workspace area. The workspace area related to a region having automated moveable equipment. A plurality of vision-based imaging devices capturing time-synchronized image data of the workspace area. Each vision-based imaging device repeatedly capturing a time synchronized image of the workspace area from a respective viewpoint that is substantially different from the other respective vision-based imaging devices. A visual processing unit for analyzing the time-synchronized image data. The visual processing unit processes the captured image data for identifying a human from a non-human object within the workspace area. The visual processing unit further determining potential interactions between a human and the automated moveable equipment. The visual processing unit further generating control signals for enabling dynamic reconfiguration of the automated moveable equipment based on the potential interactions between the human and the automated moveable equipment in the workspace area.

Vision system for monitoring humans in dynamic environments

Described is a system for rapid object detection combining structural information with bio-inspired attentional mechanisms. The system oversegments an input image into a set of superpixels, where each superpixel comprises a plurality of pixels. For each superpixel, a bounding box defining a region of the input image representing a detection hypothesis is determined. An average residual saliency (ARS) is calculated for the plurality of pixels belonging to each superpixel. Each detection hypothesis that is out of a range of a predetermined threshold value for object size is eliminated. Next, each remaining detection hypothesis having an ARS below a predetermined threshold value is eliminated. Then, color contrast is calculated for the region defined by the bounding box for each remaining detection hypothesis. Each detection hypothesis having a color contrast below a predetermined threshold is eliminated. Finally, the remaining detection hypotheses are output to a classifier for object recognition.

Rapid object detection by combining structural information from image segmentation with bio-inspired attentional mechanisms

Described, is a system for object detection via multi-scale attentional mechanisms. The system receives a multi-band image as input. Anti-aliasing and downsampling processes are performed to reduce the size of the multi-band image. Targeted contrast enhancement is performed on the multi-band image to enhance a target color of interest. A response map for each target color of interest is generated, and each response map is independently processed to generate a saliency map. The saliency map is converted into a set of detections representing potential objects of interest, wherein each detection is associated with parameters, such as position parameters, size parameters, an orientation parameter, and a score parameter. A post-processing step is applied to filter out false alarm detections in the set of detections, resulting in a final set of detections. Finally, the final set of detections and their associated parameters representing objects of interest is output.

Robust static and moving object detection system via attentional mechanisms

Described is a system for representing, storing, and reconstructing an input signal. The system constructs an index of unique polychronous groups (PCGs) from a spiking neuron network. Thereafter, a basis set of spike codes is generated from the unique PCGs. An input signal can then be received, with the input signal being spike encoded using the basis set of spike codes from the unique PCGs. The input signal can then be reconstructed by looking up in a reconstruction table, for each unique PCG in the basis set in temporal order according to firing times, anchor neurons. Using a neuron assignment table, an output location can be looked up for each anchor neuron to place a value based on the firing times of each unique PCG. Finally, the output locations of the anchor neurons can be compiled to reconstruct the input signal.

Yang Chen

Papers

Spiking Deep Convolutional Neural Networks for Energy-Efficient Object Recognition

Vision system for monitoring humans in dynamic environments

Rapid object detection by combining structural information from image segmentation with bio-inspired attentional mechanisms

Robust static and moving object detection system via attentional mechanisms

System for representing, storing, and reconstructing an input signal