2022 BRN Discussion, page-1204

For local technology investors, the most exciting thing to comeout of the Consumer Electronics Show was the news Mercedes-Benz will use amicrochip developed by ASX-listed BrainChip Holdings in its latest electricvehicle.

The Vision EQXX concept car, which claims to be able to travel 1000km on one charge, uses BrainChip’s proprietary neural processing hardware and software.

The explosion in artificial intelligence and the demand forenergy efficiency are playing into the hands of BrainChip. DavidRowe

The carcompany said it was attracted to the energy efficiency of the Akidaneuromorphic processor developed by the California-based company.

Co-founderand chief executive Peter van der Made, who is based in Perth, says the chipuses 10 times less power than power-efficient alternatives and 1000 times lesspower than standard data centre architecture.

BrainChiphas added about $1 billion to itsmarket capitalisation over the past three months thanks to a series of positive customer announcements, not all of which have been released to the ASX.

The company seems to take a liberal view of what is and whatisn’t price-sensitive information. If price sensitivity is correlated with theexcitement generated on social media and other stock chat platforms, BrainChiphas some work to do.

Theshares started to spike late last year following the November announcement thatit had entered into a licensing agreement with Japanese semiconductor manufacturerMegaChips.

Radar research

Theagreement, which runs for four years, grants MegaChips a non-exclusive,worldwide intellectual property licence for use in designing and manufacturingits Akida technology into external customers’ systems.

Thedecision by Mercedes to use BrainChip’s Akida processor in the EQXX becamepublic a week ago. The stock is up 42 per cent since then.

OnMonday BrainChip said US client Information Systems Laboratories was developing an AI-based radarresearch solution for the Air Force Research Laboratory based on its Akida™ neural networking processor.

Notwithstandingthe company’s apparent loose interpretation of continuous disclosureobligations, it is clearly a tech stock to watch in 2022 given it is achievingcommercial endorsement and is operating in one of the most prospective areas ofartificial intelligence.

In AIthere are three classes of machine learning: supervised learning, unsupervisedlearning and reinforcement learning.

Whenexperts talk about machine learning they usually do so from the perspective ofsupervised learning.

If youwant to predict someone’s exam score, you can ask things like how many hoursyou have studied, or how many hours you have slept, and then you can analysethat to get an idea of what the grades could be.

Torepresent that in machine learning the data is expressed in columns, with eachof the columns in the table representing different features or attributes.

Themathematical function that transforms this into the likely test grade is calledmatrix multiplications, whereby certain weights are given to each feature inthe table.

Greaterweight would be given to the time spent studying and less weight given to thetime the student slept.

Agraphics processing unit (GPU) does matrix multiplications very well. They havea lower processing speed, but can do things in parallel very fast.

BrainChip,Intel and IBM have been finding more efficient ways to design machine learningmodels using event-based sensors, which will become ubiquitous as the globaleconomy moves to the internet of things.

Event-based processing approach

Whenapplying machine learning to someone playing soccer, the classic machinelearning would be to process all the information around the ball, such as thegrass, the sky and other factors.

An event-basedprocessing approach saves energy because it only focuses on the moving parts,such as the ball.

At themoment most machine learning processes rely on convolutional neural networks,which is like a moving window that slides across the matrix. Essentially, itfinds the patterns that are spatially correlated.

TheBrainChip processor works on somethingcalled a spiking neural network, which only processes “events” or “spikes” that indicate useful information. This approach, similar to the way the human brain works, is not efficiently represented in GPUs.

Accordingto van der Made the Intel and IBM test chips, including Loihi, Loihi2 andTruenorth are not comparable to BrainChip’s AKD1000 chip.

He saysIBM’s Truenorth has no on-chip learning, is a “very large” and is notcost-effective.

Intel’sLoihi chip is comparable in chip size to the AKD1000, but is made in a costly7nm process while the BrainChip AKD1000 is using a 28 nm standard manufacturingtechnology, according to van der Made.

“AKD1000has on-chip convolution and on-chip learning and can be simply configured usingstandard TensorFlow tools,” he says.

“TheAKD1000 is in production and has many application examples for vision, voicerecognition, key word recognition and classification of odours and tastes.”