We introduce a stereo correspondence system implemented fully on event-based digital hardware, using a fully graph-based non von-Neumann computation model, where no frames, arrays, or any other such data-structures are used. This is the first time that an end-to-end stereo pipeline from image acquisition and rectification, multi-scale spatiotemporal stereo correspondence, winner-take-all, to disparity regularization is implemented fully on event-based hardware. Using a cluster of TrueNorth neurosynaptic processors, we demonstrate their ability to process bilateral event-based inputs streamed live by Dynamic Vision Sensors (DVS), at up to 2,000 disparity maps per second, producing high fidelity disparities which are in turn used to reconstruct, at low power, the depth of events produced from rapidly changing scenes. Experiments on real-world sequences demonstrate the ability of the system to take full advantage of the asynchronous and sparse nature of DVS sensors for low power depth reconstruction, in environments where conventional frame-based cameras connected to synchronous processors would be inefficient for rapidly moving objects. System evaluation on event-based sequences demonstrates a ~200 × improvement in terms of power per pixel per disparity map compared to the closest state-of-the-art, and maximum latencies of up to 11ms from spike injection to disparity map ejection.

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.

Deep neural networks (DNN) have been shown to be very effective at solving challenging problems in several areas of computing, including vision, speech, and natural language processing. However, traditional platforms for implementing these DNNs are often very power hungry, which has lead to significant efforts in the development of configurable platforms capable of implementing these DNNs efficiently. One of these platforms, the IBM TrueNorth processor, has demonstrated very low operating power in performing visual computing and neural network classification tasks in real-time. The neuron computation, synaptic memory, and communication fabrics are all configurable, so that a wide range of network types and topologies can be mapped to TrueNorth. This reconfigurability translates into the capability to support a wide range of low-power functions in addition to feed-forward DNN classifiers, including for example, the audio processing functions presented here.In this work, we propose an end-to-end audio processing pipeline that is implemented entirely on a TrueNorth processor and designed to specifically leverage the highly-parallel, low-precision computing primitives TrueNorth offers. As part of this pipeline, we develop an audio feature extractor (LATTE) designed for implementation on TrueNorth, and explore the tradeoffs among several design variants in terms of accuracy, power, and performance. We customize the energy-efficient deep neuromorphic networks structures that our design utilizes as the classifier and show how classifier parameters can trade between power and accuracy. In addition to enabling a wide range of diverse functions, the reconfigurability of TrueNorth enables re-training and re-programming the system to satisfy varying energy, speed, area, and accuracy requirements. The resulting system's end-to-end power consumption can be as low as$14.43\text{mW}$, which would give up to 100 hours of continuous usage with button cell batteries (CR3023$1.5\; \text{Whr}$) or 450 hours with cellphone batteries (iPhone 6s$6.55\; \text{Whr}$).

We demonstrate how to map Local Binary Patterns (LBP), a class of leading feature extractors, onto a neuromorphic processor such as TrueNorth, a silicon expression of a non-von Neumann, low-power, spiking-based, brain-inspired processor. The application is presented in the form of a texture feature extractor that can process 8-bit grayscale video at 30fps. While consuming less than 140mW of power, this neuromorphic implementation provides a rotation and contrast insensitive characterization of texture, with similar accuracy as a standard von Neumann implementation of the same algorithm. The successful mapping of an important vision routine on a neuromorphic architecture is indicative of an alternative paradigm for addressing the von Neumann bottleneck, which is currently placing severe constraints on the processing speed, power consumption, reliability, scalability, programmability and mobility of vision algorithms. This also introduces a new methodology for the design of vision algorithms for power efficient, asynchronous, mobility-targeted applications.

With recent advances in silicon technology, previously intractable Deep Neural Network (DNN) solutions to complex visual, auditory, and other sensory perception problems are now practical for real-time, energy constrained systems. One such advancement is IBM's TrueNorth neurosynaptic processor, containing 1 million neurons and 256 million synapses, consuming 65mW of power, and capable of operating in real-time for a variety of applications. In this work, we explore how auditory features can be extracted on the TrueNorth processor using low numerical precision while maintaining algorithmic fidelity for DNN based spoken digit recognition on isolated words from the TIDIGITS dataset. Further, we show that our Low-power Audio Transform with TrueNorth Ecosystem (LATTE) is capable of achieving a 24× reduction in energy for feature extraction over a baseline FPGA implementation using standard MFCC audio features, while only incurring a 3 - 6% accuracy penalty.

Abstract not provided

Drawing on neuroscience, we have developed a parallel, event-driven kernel for neurosynaptic computation, that is efficient with respect to computation, memory, and communication. Building on the previously demonstrated highly optimized software expression of the kernel, here, we demonstrate True North, a co-designed silicon expression of the kernel. True North achieves five orders of magnitude reduction in energy to-solution and two orders of magnitude speedup in time-to solution, when running computer vision applications and complex recurrent neural network simulations. Breaking path with the von Neumann architecture, True North is a 4,096 core, 1 million neuron, and 256 million synapse brain-inspired neurosynaptic processor, that consumes 65mW of power running at real-time and delivers performance of 46 Giga-Synaptic OPS/Watt. We demonstrate seamless tiling of True North chips into arrays, forming a foundation for cortex-like scalability. True North's unprecedented time-to-solution, energy-to-solution, size, scalability, and performance combined with the underlying flexibility of the kernel enable a broad range of cognitive applications.

Marching along the DARPA SyNAPSE roadmap, IBM unveils a trilogy of innovations towards the TrueNorth cognitive computing system inspired by the brain's function and efficiency. The sequential programming paradigm of the von Neumann architecture is wholly unsuited for TrueNorth. Therefore, as our main contribution, we develop a new programming paradigm that permits construction of complex cognitive algorithms and applications while being efficient for TrueNorth and effective for programmer productivity. The programming paradigm consists of (a) an abstraction for a TrueNorth program, named Corelet, for representing a network of neurosynaptic cores that encapsulates all details except external inputs and outputs; (b) an object-oriented Corelet Language for creating, composing, and decomposing corelets; (c) a Corelet Library that acts as an ever-growing repository of reusable corelets from which programmers compose new corelets; and (d) an end-to-end Corelet Laboratory that is a programming environment which integrates with the TrueNorth architectural simulator, Compass, to support all aspects of the programming cycle from design, through development, debugging, and up to deployment. The new paradigm seamlessly scales from a handful of synapses and neurons to networks of neurosynaptic cores of progressively increasing size and complexity. The utility of the new programming paradigm is underscored by the fact that we have designed and implemented more than 100 algorithms as corelets for TrueNorth in a very short time span.

Marching along the DARPA SyNAPSE roadmap, IBM unveils a trilogy of innovations towards the TrueNorth cognitive computing system inspired by the brain's function and efficiency. The non-von Neumann nature of the TrueNorth architecture necessitates a novel approach to efficient system design. To this end, we have developed a set of abstractions, algorithms, and applications that are natively efficient for TrueNorth. First, we developed repeatedly-used abstractions that span neural codes (such as binary, rate, population, and time-to-spike), long-range connectivity, and short-range connectivity. Second, we implemented ten algorithms that include convolution networks, spectral content estimators, liquid state machines, restricted Boltzmann machines, hidden Markov models, looming detection, temporal pattern matching, and various classifiers. Third, we demonstrate seven applications that include speaker recognition, music composer recognition, digit recognition, sequence prediction, collision avoidance, optical flow, and eye detection. Our results showcase the parallelism, versatility, rich connectivity, spatio-temporality, and multi-modality of the TrueNorth architecture as well as compositionality of the corelet programming paradigm and the flexibility of the underlying neuron model.

Inspired by the function, power, and volume of the organic brain, we are developing TrueNorth, a novel modular, non-von Neumann, ultra-low power, compact architecture. TrueNorth consists of a scalable network of neurosynaptic cores, with each core containing neurons, dendrites, synapses, and axons. To set sail for TrueNorth, we developed Compass, a multi-threaded, massively parallel functional simulator and a parallel compiler that maps a network of long-distance pathways in the macaque monkey brain to TrueNorth. We demonstrate near-perfect weak scaling on a 16 rack IBM® Blue Gene®/Q (262144 CPUs, 256 TB memory), achieving an unprecedented scale of 256 million neurosynaptic cores containing 65 billion neurons and 16 trillion synapses running only 388x slower than real time with an average spiking rate of 8.1 Hz. By using emerging PGAS communication primitives, we also demonstrate 2x better real-time performance over MPI primitives on a 4 rack Blue Gene/P (16384 CPUs, 16 TB memory).

In the eye gaze tracking problem, the goal is to determine where on a monitor screen a computer user is looking, ie., the gaze point. Existing systems generally have one of two limitations: either the head must remain fixed in front of a stationary camera, or, to allow for head motion, the user must wear an obstructive device. We introduce a 3D eye tracking system where the head motion is allowed without the need for markers or worn devices. We use a pair of stereo systems: a wide angle stereo system detects the face and steers an active narrow FOV stereo system to track the eye at high resolution. For high resolution tracking, the eye is modeled in 3D, including the corneal ball, pupil and fovea. We discuss the calibration of the stereo systems, the eye model, eye detection and tracking, and we close with an evaluation of the accuracy of the estimated gaze point on the monitor.

In this paper, we describe experiments conducted to explain observed differences in the bright pupil response of human eyes. Many people observe the bright pupil response as the red-eye effect when taking flash photography. However, there is significant variation in the magnitude of the bright pupil response across the population. Since many commercial gaze-tracking systems use the infrared bright pupil response for eye detection, a clear understanding of the magnitude and cause of the bright pupil variation gives critical insight into the robustness of gaze tracking systems. This paper documents studies we have conducted to measure the bright pupil differences using infrared light and hypothesis factors that lead to these differences.

We describe a real-time computer vision system and algorithms that extracts customer behavior information by detecting and tracking multiple people as they wait and watch advertisements on a billboard or a new product promotion at a stand. Our system segments each frame into foreground regions which contains multiple people. Foreground regions are further segmented into individuals using a temporal segmentation of foreground and motion cues and global shape constraints on 2D Silhouettes. A 2D dynamic appearance templates is used to track people. The system can provide online customer information, such as, number of people currently watching the billboard, their gender, and offline customer data, such as, how long each people looked at the billboard. Experimental results demonstrate robustness and real-time performance of the algorithm.

We describe a monocular real-time computer vision system that identifies shopping groups by detecting and tracking multiple people as they wait in a checkout line or service counter. Our system segments each frame into foreground regions which contains multiple people. Foreground regions are further segmented into individuals using a temporal segmentation of foreground and motion cues. Once a person is detected, an appearance model based on color and edge density in conjunction with a mean-shift tracker is used to recover the person's trajectory. People are grouped together as a shopping group by analyzing interbody distances. The system also monitors the cashier's activities to determine when shopping transactions start and end. Experimental results demonstrate the robustness and real-time performance of the algorithm.

Reliable detection and tracking of eyes is an important requirement for attentive user interfaces. In this paper, we present a methodology for detecting eyes robustly in indoor environments in real-time. We exploit the physiological properties and appearance of eyes as well as head/eye motion dynamics. Infrared lighting is used to capture the physiological properties of eyes, Kalman trackers are used to model eye/head dynamics, and a probabilistic based appearance model is used to represent eye appearance. By combining three separate modalities, with specific enhancements within each modality, our approach allows eyes to be treated as robust features that can be used for other higher-level processing.

This paper presents a multiple face detector based on a robust pupil detection technique. The pupil detector uses active illumination that exploits the retro-reflectivity property of eyes to facilitate detection. The detection range of this method is appropriate for interactive desktop and kiosk applications. Once the location of the pupil candidates are computed, the candidates are filtered and grouped into pairs that correspond to faces using heuristic rules. To demonstrate the robustness of the face detection technique, a dual-mode face tracker was developed, which is initialized with the most salient detected face. Recursive estimators are used to guarantee the stability of the process and combine the measurements from the multi-face detector and a feature correlation tracker. The estimated position of the face is used to control a pan-tilt servo mechanism in real-time, that moves the camera to keep the tracked face always centered in the image.

Presents a new covariant basis, dubbed the quasi-orthogonal Q-spline basis, for the space of n-degree periodic uniform splines with k knots. This basis is obtained analogously to the B-spline basis by scaling and periodically translating a single spline function of bounded support. The construction hinges on an important theorem involving the asymptotic behavior (in the dimension) of the inverse of banded Toeplitz matrices. The authors show that the Gram matrix for this basis is nearly diagonal, hence, the name "quasi-orthogonal". The new basis is applied to the problem of approximating closed digital curves in 2D images by least-squares fitting. Since the new spline basis is almost orthogonal, the least-squares solution can be approximated by decimating a convolution between a resolution-dependent kernel and the given data. The approximating curve is expressed as a linear combination of the new spline functions and new "control points". Another convolution maps these control points to the classical B-spline control points. A generalization of the result has relevance to the solution of regularized fitting problems.

In image retrieval based on color, the weighted distance between color histograms of two images, represented as a quadratic form, may be defined as a match measure. However, this distance measure is computationally expensive and it operates on high dimensional features (O(N)). We propose the use of low-dimensional, simple to compute distance measures between the color distributions, and show that these are lower bounds on the histogram distance measure. Results on color histogram matching in large image databases show that prefiltering with the simpler distance measures leads to significantly less time complexity because the quadratic histogram distance is now computed on a smaller set of images. The low-dimensional distance measure can also be used for indexing into the database.>

The paper presents a new approach to least-squares spline fitting of curves. A new approximately orthogonal basis, the Q-spline basis, for n-degree uniform spline space is developed. Using the Q-spline basis, it is shown that least squares spline fitting can be approximated via a single fixed sized inner product for each control point. Another convolution maps these Q-spline control points to the classical B-spline control points. Tight error bounds on the approximation induced errors are derived. Finally a procedure for discrete least squares spline fitting via convolution is presented along with several examples. A generalization of the result has relevance to the solution of regularized fitting problems.>

On-line collections of images are growing larger and more common, and tools are needed to efficiently manage, organize, and navigate through them. The authors have developed a prototype system called QBIC which allows complex multi-object and multi-feature queries of large image databases. The queries are based on image content-the colors, textures, shapes, and positions of images and the objects/regions they contain. The system computes numeric features to represent the image properties and uses similarity measures based on these features for image retrieval. The focus of the paper is its user interface which allows a user to graphically pose and refine queries based on multiple visual properties of images and their objects.>

We describe how the QBIC (Query By Image Content) system handles "multi-*" queries-queries on large image collections involving multifeatures of each image as a whole and of multiple objects within each image. The queries are based on properties of image content-such as colors, textures, shapes, and edges. The system computes a set of features to describe the above properties, uses distance-like measures on the features to provide similarity based retrieval, and has a graphical interface that enable users pose queries visually. In this paper, we present QBIC indexing algorithms that allow these "multi-*" queries to run efficiently.

In this panel session, the following topics are discussed: artificial intelligence in business; artificial intelligence in multimedia; neural networks as a tool for artificial intelligence: software engineering for knowledge-based systems: and artificial intelligence as a solution for software engineering.>

The object-oriented language for image and vision environments (OLIVE), which is intended to make it easier to develop efficient, portable applications is presented. OLIVEs principal object types, called images and loci (abstractions of point sets and geometric entities), and their corresponding operations, including the use of locuses as generalized indexes for images, are defined. Several examples of OLIVE for typical image processing and machine vision tasks are presented. Issues concerning the implementation of OLIVE, including a hardware architecture that simplifies the implementation while enhancing its performance, are discussed.>

The storage requirements for component labeling and feature extraction operations are unknown a priori. Whenever large images are processed, many labels, and thus a large amount of storage, may be required, making hardware implementation difficult. The proposed labeling procedure eliminates memory overflow by enabling the reuse of memory locations in which features of nonactive labels had been stored. The storage requirement for the worst case conditions is analyzed and is shown to be realizable. The basic procedure can be implemented in two modes, an interrupted mode or a parallel mode. A hardware design is presented.

Several commercially available image processing systems do not have comparators in their arithmetic processors. This fact makes the computation of local minima/maxima of digital images or pixel-by-pixel minima/maxima between two or more images seem impossible. This letter describes an efficient technique for computing maxima and minima without the use of a hardware comparator. The procedure requires only table look-up and simple arithmetic operations, and therefore can be efficiently implemented in commercially available pipeline image processors.