2020 BRN Discussion, page-18869

Well it is probably timely to look at Intel’s offering inthe world of neuromorphic computing. Tobe fair to Intel it makes clear in all its presentations that it is some yearsaway from producing a commercial neuromorphic chip and that its currentofferings are early stage research and development chips. The most advanced of which is PohoikiSprings. Pohoiki Springs was unveiled inMarch 2020 some 7 months ago now and is a system of 768 Loihi neuromorophicresearch chips in a chassis the size of five standard servers. Each of the 768Loihi chips contains 130,000 neurons and 130 million synapses.

(AKD1000 has1,200,000 neurons and 10 billion synapses on a 28nm chip measuring 15mm x 15mm.I don’t think you would need a server server chassis better to leave in youpocket because it does not get hot.)

In releasing the Pohoiki Springs neuromorophic processorIntel demonstrated how it was able to pick out 10 smellswith a small training sample, high accuracy, and in a system that consumes only300 Watts.

Below I have reproduced the fullarticle announcing this achievement and immediately after I have extracted fromthe original paper back in October, 2019 by Peter van der Made, Adam Osseiranand others announcing that utilising only the AKIDA Development Environment theyhad successfully identified the 20 gas data set with a latency of only 3seconds.

This paper and its findings with therelease of AKD1000 and Peter van der Made’s latest presentation is now out ofdate as the current performance of Brainchips e-nose is improved by beingimplemented on AKD1000 and is proving successful not just with the 20 gas dataset but also Covid-19, Malts, Volatile organic compounds of various diseasesetc;

Of course AKD1000 uses so littlepower compared with Pohoiki Springs 300 watts it seems a bit childish to point outAKD1000 is using a few thousands of a watt in power to out perform PohoikiSprings.

INTEL SMELLS NEUROMORPHICOPPORTUNITY

March 19, 2020 Nicole Hemsoth

A photo shows Intel’s latest neuromorphic system, Pohoiki Springs, andone of the rows within it. The system unveiled in March 2020 integrates 768Loihi neuromorphic research chips inside a chassis the size of five standardservers. (Credit: Intel Corporation)

Neuromorphiccomputing has a rather long way to go before it becomes an accepted part ofsystems. Like its quantum brethren, mapping problems to the architecture isstill a heady challenge even though a few use cases show remarkable promise.Further, just as with quantum computing, most major chip and systems playershave some interest in exploring the technical possibilities and Intel is noexception.

Since the scope ofproblems that can be tackled with neuromorphic chips is still limited, Intelfocused on one very specific use case to highlight the advance from its last64-chip system based on its “Loihi” architecture. Using a system called“Pohoiki Springs” that has scaled to 768 Loihi chips (100 million spikingneurons) sitting in a 5U rack-mount chassis Intel showed how a neuromorphicsystem can pick out smells with a small training sample, high accuracy, and ina system that consumes only 300 Watts.

When we firstdescribed “Pohoiki Beach” which is based on Intel’s “Nahuku” boards, each ofwhich contained eight to 32 Loihi processors, it was a 64-processor systembuilt from between two and eight boards (Intel did not offer details of theexact configuration, including how the chips and boards are networked together(and still has not). That same system has now been scaled to the aforementioned768 chips.

Loihi, which Intelunveiled in 2017, initially provided the equivalent of 130,000 neurons and 130million synapses, implemented as a manycore mesh, a dramatic increase now in2020 with over one million neurons. Each core contains a “learningengine” that can support different many types of AI models, includingsupervised, unsupervised, and reinforcement learning, among others. Accordingto Intel, Loihi is about 1,000 times faster and 10,000 times more efficientthan CPUs for applications like sparse coding, graph search andconstraint-satisfaction problems. The chip has been available to researchersthrough the Intel Neuromorphic Research Community (INRC) via a cloud serviceand as the Kapoho Bay platform, a Loihi-based USB form factor device. (And yes,if Intel wanted to work toward commercializing an esoteric technology, it couldhave done the world a favor and not done it all under names that aren’t stickyeither).

Like the brain,Loihi can process certain demanding workloads up to 1,000 times faster and10,000 times more efficiently than conventional processors. Pohoiki Springs isthe next step in scaling this architecture to assess its potential to solve notjust artificial intelligence problems, but a wide range of computationallydifficult problems. Intel researchers believe the extreme parallelism andasynchronous signaling of neuromorphic systems may provide significantperformance gains at dramatically reduced power levels compared with the mostadvanced conventional computers available today.

According to MikeDavies, who heads Intel’s neuromorphic computing program, these specialtysystems can be used by doctors to sniff out diseases, in airports to detectweapons, drugs, or bombs, or dangerous chemicals at manufacturing sites, forinstance.

Although it mightsound like a stretch to use a specialized neuromorphic architecture with afussy programming suite to do all of this when a neural network could also pickup similar patterns (as Google and others have shown) there are some features of a neuromorphic system that traditional deep learning models and machines can’t touch. The energy efficiency and time to result are the two most prominent.

The efficiencygains come from fully integrating compute and memory on a neuromorphic system.There is no separate memory that streaming instructions and data need to swingthrough. Everything is integrated into one distributed fabric of compute andmemory. As Davies explains, it all boils down to asymmetry. “Wanting the systemto communicate or not is a power expenditure question. If you don’t sendanything that’s a 0 binary value; not sending means not using energy. Codinginformation in this temporal way, sending at a point in time can encode is away to send info and you can compute with those codes in a way that allows youto prefer a “0” state.” The problem is that getting to that state requires arethink of algorithms. This is what the spikes are all about in a “spiking”neuromorphic system, he adds.

“Pohoiki Springsscales up our Loihi neuromorphic research chip by more than 750 times, whileoperating at a power level of under 500 watts. The system enables our researchpartners to explore ways to accelerate workloads that run slowly today onconventional architectures, including high-performance computing (HPC)systems,” says Davies.

---------------------------------------------------------------------------------------------------------------------------------------------------------------------------

Article A Hardware-DeployableNeuromorphic Solution for Encoding and Classification of Electronic Nose Data

Anup Vanarse 1,* , Adam Osseiran 1, Alexander Rassau 1 and Peter van der Made 2 1 School of Engineering, EdithCowan University, 6027 Perth, Australia; [email protected] (A.O.);[email protected] (A.R.) 2 Brainchip Inc., Aliso Viejo, CA 92656, USA;[email protected] * Correspondence: [email protected] Received: 4October 2019; Accepted: 31 October 2019; Published: 6 November 2019

Abstract: In several application domains,electronic nose systems employing conventional data processing approaches incursubstantial power and computational costs and limitations, such as significantlatency and poor accuracy for classification. Recent developments inspike-based bio-inspired approaches have delivered solutions for the highlyaccurate classification of multivariate sensor data with minimizedcomputational and power requirements. Although these methods have addressedissues related to efficient data processing and classification accuracy, otherareas, such as reducing the processing latency to support real-time applicationand deploying spike-based solutions on supported hardware, have yet to bestudied in detail. Through this investigation, we proposed a spiking neuralnetwork (SNN)-based classifier, implemented in a chip-emulation-baseddevelopment environment, that can be seamlessly deployed on a neuromorphicsystem-on-a-chip (NSoC). Under three different scenarios of increasingcomplexity, the SNN was determined to be able to classify real-valued sensordata with greater than 90% accuracy and with a maximum latency of 3 s on thesoftware-based platform. Highlights of this work included the design andimplementation of a novel encoder for artificial olfactory systems,implementation of unsupervised spike-timing-dependent plasticity (STDP) forlearning, and a foundational study on early classification capability using theSNN-based classifier. Keywords: SNN-based classification; neuromorphicolfaction; bio-inspired electronic nose systems 1. Introduction The biologicalolfactory pathway is one of the most complex sensing systems in the naturalworld, mainly because of its ability to identify a wide range of odors byprocessing high-dimensional, multivariate information. This is achieved throughneural signal representation of odor information from millions of receptorneurons and the use of specialized biological neural networks enablingclassification of odors in real time [1–3]. In order to emulate theseprocessing and sensing capabilities, artificial olfactory systems, also knownas electronic noses (e-noses), were introduced in 1982 as chemosensinginstruments. Since their introduction [4], e-nose systems have foundapplications in various fields, including bio-security, environmentalmonitoring, food quality control, and medical diagnosis [5–10]. However, theimplementation of traditional processing methods carries with it severallimitations, such as poor classification accuracy, high memory and powerrequirements, and data-intensive complex processing, that have impeded theirapplication as real-time standalone systems [11]. Recent developments inchemical sensing materials along with advances in bio-inspired neuromorphicengineering have shown promising solutions to overcome the limitations ofelectronic nose systems based on traditional approaches. Neuromorphic olfactionaims to emulate the Sensors 2019, 19, 4831; doi:10.3390/s19224831www.mdpi.com/journal/sensors Sensors 2019, 19, 4831 2 of 13 neuro-computationalprinciples of the biological olfactory pathway in electronic olfactory systemsto develop reliable, robust, real-time, and low-power machine olfactionsolutions [2,12,13]. The analog very large-scale integration (VLSI) olfactionchip proposed by Koickal et al. in [14] was among the first neuromorphicolfactory systems, comprising of a polymer-based sensor array, a signalconditioning unit for data transformation, and a spiking neural network(SNN)-based classifier implementing spike-time-dependent plasticity (STDP) forlearning. Following this research, several other spike-based olfactory systems,such as [1,12,15,16], were proposed, mainly focusing on emulation of thebiological olfactory pathway, hence resulting in impractical designs thatprovide results only under certain operating constraints and, therefore, maynot be suitable for real-world applications. Other major contributions inneuromorphic olfaction were reviewed and tabulated in [11]. Recent developmentsin neuromorphic olfaction have focused on leveraging the sparse spike-baseddata format to implement practical solutions using novel data-to-spike encodingmethods, SNNs as pattern-recognition engines, and bio-inspired learning forclassifier training. However, these implementations have not been able toaddress issues such as computational power requirements, and processing latency[11,13,17]. Through this study, we present an SNN-based solution for electronicolfactory systems that can be implemented on a neuromorphic hardware system forthe rapid classification of odors with minimal processing latency. Furthermore,we also introduce AERO (address event representation for olfaction), an addressevent representation (AER)-based encoder that encodes sensor responses tomeaningful spiking data. The SNN was implemented using BrainChip’s AkidaDevelopment Environment (ADE) [18], a Python-based emulation of the Akidaneuromorphic system-on-chip (NSoC) [19]. The classification performance of theSNN was validated using the benchmark dataset [20] under different scenariosthat test the robustness of the network. Furthermore, we also tested the SNN ona partial dataset with reduced timepoints to discuss the potential ofimplementing the network for early classification results.

My opinion only DYOR