Search

In mobile and edge devices, always-on keyword spotting (KWS) is an essential function to detect wake-up words. Recent works achieved extremely low power dissipation down to $\sim500$ nW [1]. However, most of them adopt noise-dependent training, i.e. training for a specific signal-to-noise ratio (SNR) and noise type [1], and therefore their accuracies degrade for different SNR levels and noise types that are not targeted in the training (Fig. 9.9.1, top left). To improve robustness, so-called noise-independent training can be considered, which is to use the training data that includes all the possible SNR levels and noise types [2]. But, this approach is challenging for an ultra-low-power device since it demands a large neural network to learn all the possible features. A neural network of a fixed size has its own memory capacity limit and reaches a plateau in accuracy if it has to learn more than its limit (Fig. 9.9.1, top right). On the other hand, it is known that biological acoustic systems employ a simpler process, called divisive energy normalization (DN), to maintain accuracy even in varying noise conditions [3]. In this work, therefore, by adopting such a DN, we prototype a normalized acoustic feature extractor chip (NAFE) in 65nm. The NAFE can take an acoustic signal from a microphone and produce spike-rate coded features. We pair NAFE with a spiking neural network (SNN) classifier chip [4], creating the end-to-end KWS system. The proposed system achieves 89-to-94% accuracy across -5 to 20dB SNRs and four different noise types on HeySnips [5], while the baseline without DN achieves a much lower accuracy of 71-87%. NAFE consumes up to 109nW and the KWS system 570nW.

Despite recent advances in the understanding of olfactory signal processing [1], [2], [3], [4], [5], robust odorant sensing in complex environments with time-varying odorant identities and concentrations remains an open problem. Particularly, the operational principles of biological and biomimetic olfactory sensors define a new class of sampling problems in which the odorant identity and intensity are multiplicatively coupled into a volatile signal format. We solve the sampling problem by developing the Odorant Encoding Machine (OEM), a biomimetic system based on the latest insights in the architectural organization of the fruit fly early olfactory system. The OEM provides event-driven sensing, reconstruction and robust representation of odorant identity as a combinatorial code of multidimensional spike trains. Like its biological counterpart, OEM 1) decouples odorant identity and concentration encoding via a predictive coding circuit, 2) enables real-time responses to changing odorant input through an on-off circuit, and 3) provides robust representation of odorant identity with a real-time hashing circuit. Furthermore, the OEM is directly applicable for future in silico implementations.

Computing technology has been a backbone of our society. Its importance is hard to overemphasize. Today, we again confirm its extreme importance with recent advances in deep neural networks. Those emerging workloads impose an unprecedented amount of arithmetic complexity and data access beyond our existing computing systems can barely handle. Particularly, mobile and embedded computing systems will face a major challenge in achieving energy-efficient computing for truly enabling intelligent systems. In this talk, we will discuss the emerging analog and mixed-signal circuit techniques to improve energy efficiency. We will discuss two recent cases using such techniques, one on the speech recognition processor in hybrid analog and digital circuits and the other on the embedded SRAM circuits that support analog-mixed-signal in-memory (in-bitcell) computing for convolutional and deep neural networks.

Voice user interfaces (UIs) are highly compelling for wearable and mobile devices. They have the advantage of using compact and ultra-low-power (ULP) input devices (e.g. passive microphones). Together with ULP signal acquisition and processing, voice UIs can give energy-harvesting acoustic sensor nodes and battery-operating devices the sought-after capability of natural interaction with humans. Voice activity detection (VAD), separating speech from background noise, is a key building block in such voice UIs, e.g. it can enable power gating of higher-level speech tasks such as speaker identification and speech recognition [1]. As an always-on block, the power consumption of VAD must be minimized and meanwhile maintain high classification accuracy. Motivated by high power efficiency of analog signal processing, a VAD system using analog feature extraction (AFE) and mixed-signal decision tree (DT) classifier was demonstrated in [2]. While it achieved a record of 6μW, the system requires machine-learning based calibration of the DT thresholds on a chip-to-chip basis due to ill-controlled AFE variation. Moreover, the 7-node DT may deliver inferior classification accuracy especially under low input SNR and difficult noise scenario, compared to more advanced classifiers like deep neural networks (DNNs) [1,3]. Although heavy computational load in conventional floating-point DNNs prevents their adoption in embedded systems, the binarized neural networks (BNNs) with binary weights and activations proposed in [4] may pave the way to ULP implementations. In this paper, we present a 1μW VAD system utilizing AFE and a digital BNN classifier with an event-encoding A/D interface. The whole AFE is 9.4x more power-efficient than the prior art [5] and 7.9x than the state-of-the-art digital filter bank [6], and the BNN consumes only 0.63μW. To avoid costly chip-wise training, a variation-aware python model of the AFE was created and the generated features were used for offline BNN training. Measurements show 84.4%/85.4% mean speech/non-speech hit rate with 1.88%/4.65% 1-σ standard deviation among 10 dies using the same weights for 10dB SNR speech with restaurant noise.

Video Time Decoding Machines faithfully reconstruct bandlimited stimuli encoded with Video Time Encoding Machines. The key step in recovery calls for the pseudo-inversion of a typically poorly conditioned large scale matrix. We investigate the realization of time decoders employing only neural components. We show that Video Time Decoding Machines can be realized with recurrent neural networks, describe their architecture and evaluate their performance. We provide the first demonstration of recovery of natural and synthetic video scenes encoded in the spike domain with decoders realized with only neural components. The performance in recovery using the latter decoder is not distinguishable from the one based on the pseudo-inversion matrix method.

We consider the problem of reconstructing finite energy stimuli from a finite number of contiguous spikes. The reconstructed signal satisfies a consistency condition: when passed through the same neuron, it triggers the same spike train as the original stimulus. The recovered stimulus has to also minimize a quadratic smoothness criterion. We show that under these conditions, the problem of recovery has a unique solution and provide an explicit reconstruction algorithm for stimuli encoded with a population of integrate-and-fire neurons. We demonstrate that the quality of reconstruction improves as the size of the population increases. Finally, we demonstrate the efficiency of our recovery method for an encoding circuit based on threshold spiking that arises in neuromorphic engineering.

Time encoding is a real-time asynchronous mechanism of mapping analog amplitude information into multidimensional time sequences. We investigate the exact representation of analog video streams with a Time Encoding Machine realized with a population of spiking neurons. We also provide an algorithm that perfectly recovers streaming video from the spike trains of the neural population. Finally, we analyze the quality of recovery of a space-time separable video stream encoded with a population of integrate-and-fire neurons and demonstrate that the quality of recovery increases as a function of the population size.

Time encoding is a real-time asynchronous mechanism of mapping amplitude information into a time sequence. We investigate fast algorithms for the recovery of time encoded bandlimited signals and construct an algorithm that has provably low computational complexity. We also devise a fast algorithm that is parameter-insensitive.

Time encoding is a novel real-time asynchronous mechanism for encoding amplitude information into a time sequence. The analog bandlimited input is fed into a simple nonlinear neuron-like circuit that generates a strictly increasing time sequence based on which the signal can be reconstructed. The heart of the reconstruction is solving a system of ill-conditioned linear equations. The paper shows that the equations can be manipulated so that the reconstruction becomes feasible using a cellular neural network (CNN) with a banded system matrix. In particular, the system is first transformed into a smaller well-conditioned system, and, then, the Lanczos process is used to lay it out into a set of even smaller systems characterized by a set of tridiagonal matrices. Each of these systems can directly be solved by CNNs, whereas the preprocessing (transformation and Lanczos algorithm) and simple postprocessing phases can be partly or fully implemented by using the digital capabilities of the CNN universal machine (CNN-UM). Each step of the proposed formulation is confirmed by numerical (digital) simulations.

A time encoding machine, consisting of a feedback loop containing an adder, an integrator and a Schmitt trigger, encodes amplitude information into a time sequence. We demonstrate how to construct a time decoding machine that perfectly recovers the amplitude information from the time sequence and is trigger parameter insensitive. We derive bounds on the error in signal recovery introduced by the quantization of the time sequence. We compare these with the recovery error introduced by the quantization of the amplitude of the bandlimited signal when irregular sampling is employed. Under Nyquist-type rate conditions, quantization of a bandlimited signal in the time and amplitude domains are shown to be largely equivalent methods of information representation.

A time encoding machine is a real-time asynchronous mechanism for encoding amplitude information into a time sequence. We investigate the operating characteristics of a machine consisting of a feedback loop containing an adder, a linear filter and a Schmitt trigger. We show how to recover, loss-free, the amplitude information of a bandlimited signal from the time sequence.

This paper presents a decentralized auction-based approach to pricing of edge-allocated bandwidth in a differentiated services model for the Internet. The players in this architecture are users, one raw-capacity seller per network and one broker per service per network. With the progressive second price auction mechanism as the basic building block, we conduct a game theoretic analysis, deriving optimal strategies for buyers and brokers, and show the existence of network-wide equilibria. We investigate the system dynamics by simulating a scenario with three interconnected networks, and two types of services built on the proposed standard expedited forwarding (EF) and assured forwarding (AF) per-hop-behaviors.

A metropolitan area network is described that guarantees quality of service on the hardware level. The network supports three classes of traffic. Switching is based on the principle of asynchronous time sharing, i.e., the three traffic classes enter any switch fabric sequentially. The system architecture consists of a set of generic switch fabrics that are implemented as buses, rings, or link schedulers. A switching node typically consists of a ring that interconnects a set of access buses; these nodes are interconnected via T3/DS3 links. A flexible hardware design that supports the traffic control architecture in order to guarantee the quality of service as negotiated at call setup is also presented. >

The authors focus on a gateway design for the interconnection of ILAN (integrated local area network) and ISDN (integrated services digital network). A key question is how to maintain the quality of service for each class of traffic as guaranteed by both networks. The emphasis of the study is on the main issues involved in the interconnection of MAGNET (a network testbed at Columbia University) to PBX through a standard ISDN user-network interface. One issue, in particular, is the protocol conversion to resolve the incompatibilities, e.g. transforming voice between packet-switched form and circuit-switched form. A multilayer protocol conversion model is proposed which provides equivalent service at different layers for each class of traffic. Another issue is the development of a high-performance gateway architecture. A multimicroprocessor system architecture which exploits most parallel features in the execution of protocol conversion functions is also presented. >

A two dimensional, linear state-space model of the cochlea of the human ear, the receiver side of the speech communications paradigm, is investigated. The ear is viewed as a serial to parallel converter of information: the input to the cochlear model is a time dependent single channel signal, whereas the output of the cochlear model is a spatially discretized representation of the displacement of the basilar membrane from its rest position. A characterization of the system response of the cochlear filters is given. An approach to detecting known signals at the input using only the information available at the output based upon methods akin to 'beamforming' for a linear array is presented. The detector threshold is set using an adaptive technique that uses spatially local information. Simulation results showing the capabilities of the detection system for processing speech signals are also discussed. >

The authors investigate the structural representation of real-time network control and management of asynchronous transfer modes (ATM)-based networks. The network architecture consists of four planes that model the information transport, connection management and control, resource monitoring and management, and resource management and control functions. A fifth plane represents the architecture for network management based on the open systems interconnection (OSI) management paradigm and provides for operations, administration and maintenance tasks. The storage of information is based on a set of inheritance and containment trees of network objects. >

The CCITT Common Channel Signalling System Hash 7 is a packet-switched network used to carry call signalling and control messages independently of the underlying telecommunications network. This paper derives the mean and worst-case packet delay experienced in a 6 node network under conditions involving multiple node failures using BCMP analysis. The result is generalized to be applicable to both symmetric and asymmetric traffic loads. Numerical results are based on actual performance measurements on the Singapore National network. >

For services offered by B-ISDNs many corporate telecommunication departments will write contracts with their users to provide classes of network services and to guarantee certain levels of performance in terms of, for instance, end-to-end response time and availability. In general a level of performance can be quantified in terms of the quality (QOS) of some service offered by the network. QOS management permits the customer to monitor and control directly the utilization, performance and related parameters. In the context of this paper network management is taken to mean the more restricted QOS management at the subscriber-level. >

This paper analyzes a CCSS#7 (Common Channel Signalling System #7) network supporting database services. One example of such a database service is the 800 number telephone service in which the database translates 800 toll free numbers into the actual address numbers. One question in designing such a database system is: should it be centralized or rather be divided into several distributed smaller databases? This paper attempts to answer the question from the system performance point of view. The authors derive the total system delay for some different cases regarding the number and locations of the databases, and then compare them under various system conditions.

We introduce a methodology for establishing a connection between two arbitrary nodes of a multiclass broadband network that secures sufficient resources at every intermediate node along the selected route, and thereby guarantees quality of service (QOS). We describe how a virtual path (VP) mechanism may be implemented within our framework, and compare the performance achieved by such a scheme with that of a virtual circuit (VC) approach. The notion of contract region is introduced as a mean of achieving coordination among distributed Virtual Path controllers to ensure that the state of each network link remains within its schedulable region and the cell level QOS is guaranteed. >

An architecture capable of supporting real-time traffic monitoring and control, management, and information transport for broadband networks is outlined. It embeds the management, the traffic control, and the information transport architectures. The traffic control architecture (TCA) is characterized both as a resource manager and an extended machine. The traffic control model employed consists of a statistical and a dynamic database and a set of distributed algorithms that interact by writing and reading into these databases. The latter are part of a knowledge database capable of supporting monitoring, control, and fault management tasks. The associated tasks operate on two time scales: a fast time scale with reaction times to events on the order of milliseconds and a slow time scale where actions in response to events take place within seconds. The TCA and the information transport architecture run on the fast time scale whereas the management architecture runs on the slow time scale. >

The traffic control architecture of an integrated network is modeled as a real-time system that asynchronously interacts with its environment, i.e., the network through an interface. The interface is specified by a set of state variables to which sensors are attached for monitoring. The entity-relationship model is used to represent the software and hardware objects in the network. The interface is modeled using the entity-relationship model and a computation model. How to translate the specification of the environment and the interface into an object-oriented system is described. >

The authors report their attempt at building a highly concurrently connection management platform in the OMNI (Open Multimedia Networking Initiative). Connection set up and release are one of the most frequent signalling activities in the network. In order to handle the growth of connection volume and service types, it is important to design a signalling platform that explores concurrency, allows different control algorithms to be implemented, and to make it easy to redistribute control intelligence. Based on these principles, they developed a connection management platform and quantified their benefits. They have implemented three different connection set up algorithms to study different design alternatives of the platform. In addition, they compared the performance and scalability of these three algorithms.

To eliminate the use of conventional encryption and to improve protocol efficiency, one-way hash function based authenticated key distribution protocols have appeared in the literature following two distinct approaches: the first approach employs one-way hash function and bit-wise XOR operation; while the second approach makes use of one-way hash function and polynomial interpolation. In this paper, we present a new technique for constructing one-way hash function based protocols for authenticated key distribution. The technique makes use of systematic linear block codes in error-control coding theory, results in more efficient protocols compared with their counterparts in the literature, and in effect, unifies the above two approaches in designing such protocols. As a by product, a secure broadcast protocol is also given which is capable of distributing a data message of any size in one broadcast protocol message.

In noncooperative networks users make control decisions that optimize their individual performance measure. Focusing on routing, two methodologies for architecting noncooperative networks are devised, that improve the overall network performance. These methodologies are motivated by problem settings arising in the provisioning and the run time phases of the network. For either phase, Nash equilibria characterize the operating point of the network. The goal in the provisioning phase is to allocate link capacities that lead to systemwide efficient Nash equilibria. The solution of such design problems is, in general, counterintuitive, since adding link capacity might lead to degradation of user performance. For systems of parallel links, it is shown that such paradoxes cannot occur and that the optimal solution coincides with the solution in the single-user case. Extensions to general network topologies are derived. During the run time phase, a manager controls the routing of part of the network flow. The manager is aware of the noncooperative behavior of the users and makes its routing decisions based on this information while aiming at improving the overall system performance. We obtain necessary and sufficient conditions for enforcing an equilibrium that coincides with the global network optimum, and indicate that these conditions are met in many cases of interest. >

A key consideration in building differentiated network services is the feasibility of maintaining stable and consistent service level agreements across multiple networks where allocations are made only on the edges. To investigate this, we consider a game theoretic model of capacity provisioning in a differentiated services Internet. The players are one raw-capacity seller per network, one broker per service per network, and users, to play the roles of wholesellers, retailers and end users respectively in a two-tier wholeseller/retailer market. Based on this model, we are able to construct an explicit necessary and sufficient condition for the stability of the game, which determines the sustainability of a given set of SLA configurations among peering ISP. The analytical results are validated with simulations of user and broker dynamics, using distributed progressive second price auctions as the spot market mechanism in a scenario with three interconnected networks, and two services.

The exponential growth of computer networking demands massive upgrades of capacity in existing networks. Traditional capacity design methodologies, developed with the single-class networking paradigm in mind, overlook the noncooperative structure of modern networks. Consequently, such design approaches entail the danger of degraded performance when resources are added to a network, a phenomenon known as the Braess paradox. The present paper proposes methods for efficiently adding resources to a noncooperative network of general topology. It is shown that the paradox is avoided when resources are added across the network, rather than on a local scale, and when upgrades are focused on direct connections between sources and destinations. The relevance of these findings to modern networks is demonstrated.

The Binding Architecture is an open architecture for building multimedia networks that must guarantee quality of service (QOS). We explore the notion of a service in the Binding Architecture and propose a conceptual model for building scalable multimedia distribution services based on it. We begin first by examining the relation between resources, their abstractions and the services that can be built from them and use this to derive a general model for binding. Based on this model, we identify a general set of capabilities required for building any multimedia distribution service. We also describe how these capabilities can be incorporated into our view of the service creation process. Finally, we augment our discussion with the description of an example service which we have developed using this paradigm.

The design and implementation of the transport layer of a native ATM (asynchronous tranfer mode) protocol stack and its embedding into an overall architecture that provides end-to-end quality of service (QOS) is presented. Within this architecture, the typical transport functionalities are enlarged with QOS monitoring and adaptation mechanisms. A QOS-based API (application programming interface) is proposed that shields application programmers from the complexity of QOS management and control. Both unicast and multicast connections supporting interactive multimedia applications are considered.

A proactive distributed cooperative control algorithm for resource scheduling of broadband networks based on asynchronous time sharing is presented. The problem setting for distributed, cooperative proactive scheduling is described. An overview of the distributed cooperative algorithm is presented. The general control principles underlying the effectiveness of the cooperative algorithm are discussed. The formal details of the distributed cooperative algorithm and its underlying single noise scheduling algorithm are considered. Simulation results comparing the distributed cooperative algorithm with other networkwide scheduling algorithms are reported. >

An object-oriented model for network independent monitoring and performance evaluation in real-time is presented. the model describes the behavior of network objects by a set of state variables and a set of invariants. The model is used to evaluate quality-of-service characteristics of different classes of traffic in an integrated network testbed. An objective-driven measurement strategy to establish the dynamic database of the network is discussed. A statistical database derived from the dynamic database is also obtained. >

Real-time traffic measurements on MAGNET II, an integrated network testbed based on asynchronous time sharing (ATS), are reported. The quality of service is evaluated by monitoring the buffer occupancy distribution, the packet time delay distribution, the packet loss, and the gap distribution of the consecutively lost packets. Experiments show that both time delay and buffer occupancy distributions of multiplexed video sources display a marked bimodal behavior, which does not seem to depend on the buffer size. The reliance of the network designer on traffic sources that do not exhibit substantial correlations can lead to implementations with serious congestion problems. For ATS-based networks with different traffic classes, the impact of a traffic class on the performance of the other classes tends to be diminished when compared with single-class-based ATM (asynchronous transfer mode) networks. >

The problem of optimum flow control of a class of queueing systems which appears as a model for virtual circuits in computer communication networks is investigated. It is shown that the optimal control that achieves the maximum throughput under a bounded average time delay criterion can be specified by a "window" flow control mechanism. The window size can be easily derived from the preassigned upper bound on the time delay.

The network architecture of NMAGNET, a testbed for integrated local area networks, is presented. The system architecture consists of integrated workstations, control and communication subsystems, and gateways. The network architecture incorporates this organization at the system level and represents an instance of the Integrated Reference Model previously proposed by the authors. It provides resource management, connection management and control, and user information flows in logically separate planes. Protocols supporting the control subsystem are organized in the (M)-Plane, while the protocols supporting the communication subsystem are embedded in the (C)-and (U)-Planes. The requlirements of the different traffic classes are quantified by a set of attributes to guarantee specified performance measures.