BRN Discussion Ongoing

[Fullmoonfever]

From a few days ago out of the TCS Annual Report 22/23.

 

[Rach2512]

Morning IloveLamp

BrainChip and Socionext expect Akida silicon in Q3

BrainChip and Socionext expect engineering samples of their neural net processor Akida in Q3. They are being made on a TSMC MPW run. BrainChip has

www.electronicsweekly.com

This from Apr 2020, so could that mean that Socionext has something to offer with Akida already built in?

BrainChip and Socionext expect Akida silicon in Q3

BrainChip and Socionext expect engineering samples of their neural net processor Akida in Q3. They are being made on a TSMC MPW run.

BrainChip and Socionext expect Akida silicon in Q3

BrainChip and Socionext expect Akida silicon in Q3

brainchip.com

Morning everyone, have a great day ❤

Pom down under, I was merely trying to highlight the fact that BrainChip and Socionext have been working together for over 4 years and I was thinking that between them they may already have something to offer right now, not 2-3 years down the track. Would anyone else care to agree?

 

[skutza]

I know these guys have been discussed before, but interesting read. thoughts?

Hailo Closes New $120 Million Funding Round and Debuts Hailo-10, A New Powerful AI Accelerator Bringing Generative AI to Edge Devices

Hailo’s funding now exceeds $340 million as the company introduces its newest AI accelerator specifically designed to process LLMs at low power

www.edge-ai-vision.com

 

[Diogenese]

I knew TDK were looking into neuromorphic as posted last year but unfortunately not ours by looks.

Thoughts @Diogenese .

Maybe Anthony, Nandan, Anil or someone should get in touch haha

Excerpt start of blog. Rest in links.


https://product.tdk.com/en/techlibrary/developing/index.html

Solving AI Energy Problems with Neuromorphic Technology​

With the expansion of AI usage, the energy consumed by computers has been increasing explosively. TDK is developing a spin memristor, an analog memory element that electrically mimics the synapses in the human brain. Devices that utilize this technology, known as neuromorphic devices, are expected to be over 100 times more energy-efficient than conventional devices. This device technology can be manufactured using techniques similar to those currently used for MRAM*1 production. Leveraging its expertise in magnetic technology, developed through HDD heads and magnetic sensors, TDK aims to reduce the energy needed for AI and to discover new AI devices that can learn in real-time and adapt to their environment and users.
*1 MRAM：
Magnetoresistive Random Access Memory. Unlike conventional memory, data does not disappear (non-volatile). This allows for reduction in calculation and standby power as it can be quickly started even if the circuit power is turned off.

Contents​

• 1
  The rise of AI use and the energy problem
• 2
  Neuromorphic devices mimicking human brain functions
• 3
  Challenges with conventional memristor devices
• 4
  Features and advantages of TDK spin memristors
• 5
  The future of ultra-low power AI devices

Hi Fmf,

MRAM = analog.

 

[Frangipani]

A bit older 2022-02, but for me as a sceptical German it's nice to see any sign that Akida is known here. It's about a project funded by a federal state. And what do my tired eyes discover:

"PROJECT RESULTS​

In the course of the project, "event-based" cameras were used in parallel with classic "frame-based" cameras to create a data set with twelve different gestures/actions. This dataset was used to train both classical neural networks (CNN+LSTM) and spiking neural networks to classify the gestures shown. The networks trained in this way were run on corresponding hardware accelerators and the two resulting systems (classic and event-based) were compared in terms of prediction accuracy, electrical power consumption and generated data rate. The event-based system achieved a higher prediction accuracy with a significantly lower data rate and electrical power consumption. The figure shows a section of the demonstration system.final report can be extracted."
View attachment 36197


"The SNNs were also ported as far as possible to the neuromorphic hardware platform BrainChip Akida and evaluated there in terms of throughput and electrical power consumption. A comparison with the simulation of the SNN on a GPU-based computer and with the values of the KNN-based system can be seen in Table 3."

Means power consumption and frames per second
View attachment 36195
The entire final report translated see attached.


It is/was about:
View attachment 36196
The Germans can read the report here:
https://www.embeddedneurovision.de/...2-02-abschlussbericht-embeddedneurovision.pdf

https://www.embeddedneurovision.de/

https://www.fzi.de/

https://www.inferics.com/

https://www.hs-analysis.com/

Who knows, maybe one of the project participants is among us?


"In addition, the system is connected to an Nvidia Jetson NX8 AI processor and a BrainChip Akida as a neuromorphic AI processor.
neural networks that the partners have developed as part of the project. developed by the partners as part of the project.
...
The development kits CeleX5_MP from CelePixel Technology Co. LTD and Gen 4.0 from Prophesee are used as hardware.
...
In the end, Norse was used in combination with Tonic as a framework, as these promised sufficient functionality, fast results and easy integration of event camera data. Recurrent, convolutional and fully networked variants were chosen as SNN models.
...
The best results were obtained with a combination of recurrent, fully wired and convolutional layers, which achieve a very high accuracy of about 98% for 3 classes with less than 400,000 parameters. A detailed list and comparison of the different networks and accuracies can be found in Table 2.
...
HS Analysis has experienced great interest from our existing customers as well as some potential new customers to apply SNNs in commercial products, especially in the area of process monitoring. The main interest here is in the reduced energy consumption and increased data protection of event-based cameras. This enables data protection-compliant process monitoring even in areas accessible to the public, and the reduced power requirement offers the possibility of operating on-edge devices via Power-over-Ethernet (PoE), which simplifies deployment. Our previous core product, the HSA KIT, which is a toolbox of different which includes a toolbox of various customised AI analyses, can also be extended with the SNN knowledge gained in the project. This means that advanced, minimalistic time series analyses are now possible, for which we have already started with the quotation and ordering process."


_____________________
I would also give a general tip not to always search in English. For the first time I had today the idea of searching in my own language ("gepulste neuronale Netze" ~ spiking neural networks).

Does FCAS mean anything to you? There is no entry here.
https://en.wikipedia.org/wiki/Future_Combat_Air_System
In addition to the companies mentioned above the Fraunhofer-Institut is deeply involved in the development of military equipment, as I heard on the radio yesterday. Was new to me. This is one of the top research institutions in our country.

Here is further evidence - postdating that of cosors (02/2022) - that AKD1000 has been used for research at FZI (Forschungszentrum Informatik / Research Centre for Information Technology) in Karlsruhe, Germany.

Julius von Egloffstein is presently both an M.Sc. student at Karlsruhe Institute of Technology (KIT) and a research assistant at FZI. The technical university and the non-profit research institute FZI collaborate closely, and hence quite a few KIT students will do research for their Bachelor’s or Master’s theses at FZI, like this young gentleman did two years ago, when he worked on his Bachelor’s thesis titled Application Specific Neural Branch Prediction with Sparse Encoding (for which he got a perfect score, by the way).

And look what kind of neuromorphic hardware he used for evaluation:

It is worth noting that FZI researchers are not working in an ivory tower - instead, their research is all about applied computer science and technology transfer to companies and public institutions:

FZI Forschungszentrum Informatik - Wikipedia

en.m.wikipedia.org

 

[Fullmoonfever]

Hi Fmf,

MRAM = analog.

Thanks D.

Thought as much by some of the wording in the blog but getting the confirm from you

 

[Frangipani]

Perplexingly, Perplexity AI completely omitted King’s College London in its list of academic institutions, although is very well-known for the neuromorphic research conducted by Prof. Osvaldo Simeone and Prof. Bipin Rajendran that aims to revolutionise wireless communications.

Osvaldo Simeone

Professor of Information Engineering

www.kcl.ac.uk

Bipin Rajendran

Professor of Intelligent Computing Systems

www.kcl.ac.uk

They recently co-authored a paper with researchers from Luxembourg, the UK and France (“Performance Evaluation of Neuromorphic Hardware for Onboard Satellite Communication Applications”), in which they had implemented SNNs on Loihi 2 - it was briefly touched on here on TSE the other day:
https://thestockexchange.com.au/threads/brn-discussion-ongoing.1/post-409255

An interesting collaboration between these three partner universities - King's, Imperial and UCL - is the London Centre for Nanotechnology (LCN), “a UK-based multidisciplinary enterprise operating at the forefront of science and technology. Our purpose is to solve global problems in information processing, healthcare, energy and the environment through the application of nanoscience and nanotechnology. Founded in 2003, the LCN began as a joint venture between University College London and Imperial College London, based at the Bloomsbury and South Kensington sites; from 2018 King's College London joined the collaboration from its base on The Strand.”

https://www.london-nano.com/about

By the way, Perplexity AI also failed to mention Brunel University London as yet another institution being involved in neuromorphic research in the British capital.

Hot off the press - Akida finally getting acknowledged alongside Loihi by three researchers from King’s College London… Prof. Simeone is also with Aalborg University, Denmark, where another co-author is from. The fifth co-author is from Princeton.

https://arxiv.org/pdf/2404.01815.pdf

 

[Fullmoonfever]

I see Anup has moved from Perth to California in the last couple of months.

Congrats to him.

Thought I'd find someone else than Rob for liking something.

We know much about Icetana?

Anup Vanarse​

AI Research Scientist | Advancing AI at Brainchip | Expert in Deep Learning | PhD in Advanced Machine Learning​

BrainChip Edith Cowan University​

Irvine, California, United States Contact Info

553 followers 500+ connections​

About
As an AI Research Scientist with a deep-rooted passion for advancing the frontiers of technology, I specialize in the intersection of deep learning and neuromorphic engineering. With a PhD in advanced machine learning, I bring extensive expertise in designing cutting-edge algorithms and models. Currently, I am at the forefront of innovation as a valuable member of Brainchip Research Institute, contributing to the development of event-based and novel AI solutions optimized for hardware deployment.

 

[IloveLamp]

Does not look like it connects to the internet, but it DOES. Where is the battery? Where is the engine? Makes no noise.

SNN with low power would be perfect for Ebikes.

View attachment 60190

STMicroelectronics helps Panasonic Cycle Technology bring AI to e-assisted bikes for affordable safety boost - ST News

STMicroelectronics (NYSE: STM), a global semiconductor leader serving customers across the spectrum of electronics applications, has announced that...

newsroom.st.com

 

[Frangipani]

Looks like Nikunj Kotecha left Brainchip end of last year?

 

[Fullmoonfever]

Looks like Nikunj Kotecha left Brainchip end of last year?

View attachment 60234


View attachment 60235

Maybe they brought Anup over to replace the vacancy?

 

[IloveLamp]

makes sense in reference to Brainchip imo.

CEO of Google DeepMind Says AI Industry Full of Hype and Grifters

Google DeepMind CEO Demis Hassibis likened the hype shrouding the AI gold rush to that of the crypto industry's high-dollar funding race.

futurism.com

 

[Pom down under]

Pom down under, I was merely trying to highlight the fact that BrainChip and Socionext have been working together for over 4 years and I was thinking that between them they may already have something to offer right now, not 2-3 years down the track. Would anyone else care to agree?

I hardly read anything properly hence the confusion lol

 

[Humble Genius]

BRAINCHIP, MAKING AI UBIQUITOUS

Latest in professional industrial tech

easyengineering.eu

Sign Up | LinkedIn

500 million+ members | Manage your professional identity. Build and engage with your professional network. Access knowledge, insights and opportunities.

www.linkedin.com

 

[supersonic001]

Markus Schäfer on LinkedIn: #mercedesbenz #leadininnovation | 14 comments

“Powertrain Disruptor of the Year” and “Technology Rollout of the Year” – two titles awarded recently to Mercedes-Benz by renowned American publication… | 14 comments on LinkedIn

www.linkedin.com

 

[Pom down under]

 

[supersonic001]

Renesas Electronics on LinkedIn: #embeddedworld #demokits #mcu

Are you planning to attend embedded world next week? Don’t miss out on your opportunity to secure a demo kit of our upcoming, new ultra-low-power MCU series…

www.linkedin.com

 

[White Horse]

makes sense in reference to Brainchip imo.

CEO of Google DeepMind Says AI Industry Full of Hype and Grifters

Google DeepMind CEO Demis Hassibis likened the hype shrouding the AI gold rush to that of the crypto industry's high-dollar funding race.

futurism.com

You have a habit of making assumptions, that display a basic ignorance of what we do, and where we fit in the would of AI.

 

[IloveLamp]

You have a habit of making assumptions, that display a basic ignorance of what we do, and where we fit in the would of AI.

Ok

Feel better?

 

[Frangipani]

Here is another preprint published yesterday, in which Prof. Osvaldo Simeone from King’s College London is listed as one of the authors. His five co-authors are all from Berlin (Fraunhofer Heinrich Hertz Institute and Technical University of Berlin).

While no specific neuromorphic hardware is mentioned (although the paper’s first reference is to a 2021 paper by Intel researchers on Loihi and the penultimate one is to Prophesee’s EVK4 HD), it is nevertheless an interesting article as it shows that the wireless community is at the forefront of advancing cutting-edge technologies and thus confirms what ISL’s CEO Dr. Joe Guerci said in the recent From The Crow’s Nest podcast:

From the Crows' Nest | Transcript: Advancing Cognitive Warfare: Unveiling Neuromorphic Frontiers

In this episode of From the Crow’s Nest, host Ken Miller talks to author Dr. Joseph Guerci about the evolution and outlook for cognitive electronic warfare systems. Dr. Guerci is an internationally...

fromthecrowsnest.transistor.fm

“Well, and to amplify that point, all the advanced capabilities that we have in our RF systems, radar and EW, most of that is driven by the wireless community, the trillion-dollar wireless community compared to a paltry radar and EW ecosystem.”

Also note the funding supported by the governments of Germany and the UK as well as through the EU’s Horizon Europe funding programme (luckily for King’s College London, post-Brexit UK it is still an associated country…)

https://arxiv.org/pdf/2404.01804.pdf

VII. CONCLUSIONS

In this work, we introduced a new system solution for device-edge co-inference that targets energy efficiency at the end device using neuromorphic hardware and signal processing, while implementing conventional radio and computing technologies at the edge. The investigated communication scheme combines on-device SNN and server-based ANN, leveraging variational directed information bottleneck technique to perform inference tasks. From a deployment perspective, our model demonstrates superior performance, less need for communication overhead and robustness under time-varying channel conditions, implying promising potential in the further 6G work. Aspects of the proposed system solution were validated in a preliminary testbed setup that implements a wireless robotic control application based on gesture recognition via a neuromorphic sensor. The testbed setup is currently being expanded to integrate end-to-end learning via an impulse-radio communication link. The proposed architecture and the corresponding testbed setup are general in the sense that they can support the implementation of different semantic tasks. Besides applications in robotics, in the future, we will also consider bio-medical applications that leverage the energy and communication efficiency of this architecture.