BRN Discussion Ongoing
- Thread starter TechGirl
- Start date
Tothemoon24
Top 20
BOSSA looks like one to keep a watch on in the hearing aid space
In a busy room full of talking people, most of us can still pick out one voice to focus on. This common yet complex task—known as the “cocktail party effect”—relies on the brain’s incredible ability to sort through sound. But for people with hearing loss, filtering out background noise can feel impossible. Even the most advanced hearing aids often struggle in these noisy environments.
Now, researchers at Boston University may have found a new way to help. They’ve developed a brain-inspired algorithm that allows hearing aids to better isolate individual voices in a crowd. When tested, this method boosted speech recognition by an impressive 40 percentage points, far outperforming current technologies.
Virginia Best, a speech and hearing researcher at BU, says this is the number one complaint among those with hearing loss. “These environments are very common in daily life,” Best explains, “and they tend to be really important to people.”
Traditional hearing aids often include tools like directional microphones—also called beamformers—that try to focus on sounds coming from one direction. But these tools have limitations. In complex environments with many voices, beamforming often fails. In fact, in tests conducted by the BU team, the standard industry algorithm didn’t help much—and sometimes made things worse.
That’s where the new technology, known as BOSSA, comes in. BOSSA stands for Biologically Oriented Sound Segregation Algorithm. It was developed by Kamal Sen, a biomedical engineering professor at BU’s College of Engineering. “We were extremely surprised and excited by the magnitude of the improvement in performance,” says Sen. “It’s pretty rare to find such big improvements.”
All subjects average word recognition scores. (CREDIT: Kamal Sen, et al.)
“You can think of it as a form of internal noise cancellation,” Sen says. Different neurons are tuned to respond to different directions and pitches. This lets your brain focus attention on one sound source while ignoring others.
BOSSA was built to mimic this process. The algorithm uses spatial cues—like how loud a sound is and how quickly it arrives in each ear—to pinpoint its location. It then filters sounds based on these cues, separating them like your brain would. “It’s basically a computational model that mimics what the brain does,” Sen says.
Each person completed the task under three different conditions: no algorithm, the standard beamforming algorithm used in current hearing aids, and BOSSA. The results were striking. BOSSA delivered a major improvement in speech recognition. The standard algorithm showed little or no improvement—and in some cases, performance dropped.
Speech reception thresholds (SRT) are shown as boxplots for each processing condition. (CREDIT: Kamal Sen, et al.)
Alexander Boyd, a BU PhD candidate in biomedical engineering, helped collect and analyze the data. He was also the lead author of the study, which was published in Communications Engineering, part of the Nature Portfolio.
Best, who formerly worked at Australia’s National Acoustic Laboratories, helped design the study. She says testing new technologies like BOSSA with real people is essential. “Ultimately, the only way to know if a benefit will translate to the listener is via behavioral studies,” Best says. “That requires scientists and clinicians who understand the target population.”
Sen has patented BOSSA and hopes to partner with companies that want to bring it to market. He believes that major tech players entering the hearing aid space—like Apple with its AirPod Pro 2, which includes hearing aid features—will drive innovation forward. “If hearing aid companies don’t start innovating fast, they’re going to get wiped out,” says Sen. “Apple and other start-ups are entering the market.”
Individual participant audiograms. The different curves show pure-tone thresholds for each of the eight participants (averaged over left and right ears). Unique symbols distinguish individual subjects. (CREDIT: Kamal Sen, et al.)
And the timing couldn’t be better. As hearing technology becomes more widely available and advanced, tools like BOSSA could help millions of people reconnect with the world around them. From social events to everyday conversations, better sound separation can mean a better life.
That’s why the team is now exploring how the same science could help people with ADHD or autism. These groups also struggle with multiple competing inputs—whether sounds, visuals, or tasks—and may benefit from tools that help guide attention.
They’re also testing a new version of BOSSA that adds eye-tracking. By following where someone looks, the device could better figure out who they’re trying to listen to. This could make the technology even more effective in fast-paced, real-world settings.
For many with hearing loss, this could change everything. Being able to join conversations, pick out voices, and stay connected socially are vital parts of daily life. With tools like BOSSA, those goals move a little closer. And as this technology continues to grow, its reach may extend beyond hearing loss, offering help with focus and attention challenges too.
What started as a solution for a noisy dinner party could one day reshape how we interact with the world.
Here, we present a system employing a novel strategy for stimulus reconstruction from neural spikes. Conceptually, this strategy uses time-frequency masking by computing a spike-based mask (spike-mask). We first consider the strategy for one-dimensional stimulus (e.g. sound waves). We show how this stimulus reconstruction method can be applied, using the cortical model as an example. We also show that this strategy produces reconstructions with intelligibility and quality higher than those reconstructed from the linear filtering method (table 1). Then we discuss how this strategy may be generalized for multi-dimensional stimulus (e.g. images and videos). The strategy presented here may be generalized to perform reconstruction on both artificial SNNs and neural models from experimental data as long as they satisfy the assumptions for our model.
In a busy room full of talking people, most of us can still pick out one voice to focus on. This common yet complex task—known as the “cocktail party effect”—relies on the brain’s incredible ability to sort through sound. But for people with hearing loss, filtering out background noise can feel impossible. Even the most advanced hearing aids often struggle in these noisy environments.
Now, researchers at Boston University may have found a new way to help. They’ve developed a brain-inspired algorithm that allows hearing aids to better isolate individual voices in a crowd. When tested, this method boosted speech recognition by an impressive 40 percentage points, far outperforming current technologies.
A New Approach to an Old Problem
In crowded social settings like dinner parties or workplace meetings, conversations often overlap. For those with hearing loss, these situations can be frustrating. Even with hearing aids, voices blur together in a mess of sound. This makes it hard to follow conversations, stay engaged, or even participate at all.A New Approach to an Old Problem
In crowded social settings like dinner parties or workplace meetings, conversations often overlap. For those with hearing loss, these situations can be frustrating. Even with hearing aids, voices blur together in a mess of sound. This makes it hard to follow conversations, stay engaged, or even participate at all.A New Approach to an Old Problem
In crowded social settings like dinner parties or workplace meetings, conversations often overlap. For those with hearing loss, these situations can be frustrating. Even with hearing aids, voices blur together in a mess of sound. This makes it hard to follow conversations, stay engaged, or even participate at all.Virginia Best, a speech and hearing researcher at BU, says this is the number one complaint among those with hearing loss. “These environments are very common in daily life,” Best explains, “and they tend to be really important to people.”
Traditional hearing aids often include tools like directional microphones—also called beamformers—that try to focus on sounds coming from one direction. But these tools have limitations. In complex environments with many voices, beamforming often fails. In fact, in tests conducted by the BU team, the standard industry algorithm didn’t help much—and sometimes made things worse.
That’s where the new technology, known as BOSSA, comes in. BOSSA stands for Biologically Oriented Sound Segregation Algorithm. It was developed by Kamal Sen, a biomedical engineering professor at BU’s College of Engineering. “We were extremely surprised and excited by the magnitude of the improvement in performance,” says Sen. “It’s pretty rare to find such big improvements.”
Built on Brain Science
Sen has spent two decades exploring how the brain decodes sound. His work focuses on how sound signals travel from the ears to the brain and how certain neurons help identify or suppress sounds. One key finding? The brain uses “inhibitory neurons” to cancel out background noise and enhance the sounds we want to hear.
All subjects average word recognition scores. (CREDIT: Kamal Sen, et al.)
“You can think of it as a form of internal noise cancellation,” Sen says. Different neurons are tuned to respond to different directions and pitches. This lets your brain focus attention on one sound source while ignoring others.
BOSSA was built to mimic this process. The algorithm uses spatial cues—like how loud a sound is and how quickly it arrives in each ear—to pinpoint its location. It then filters sounds based on these cues, separating them like your brain would. “It’s basically a computational model that mimics what the brain does,” Sen says.
Testing BOSSA in Real-Life Situations
To find out if BOSSA really works, the BU team tested it in the lab. They recruited young adults with sensorineural hearing loss, the most common form, often caused by genetics or childhood illness. Participants wore headphones and listened to simulated conversations, with voices coming from different directions. They were asked to focus on one speaker while the algorithm worked in the background.Each person completed the task under three different conditions: no algorithm, the standard beamforming algorithm used in current hearing aids, and BOSSA. The results were striking. BOSSA delivered a major improvement in speech recognition. The standard algorithm showed little or no improvement—and in some cases, performance dropped.
Speech reception thresholds (SRT) are shown as boxplots for each processing condition. (CREDIT: Kamal Sen, et al.)
Alexander Boyd, a BU PhD candidate in biomedical engineering, helped collect and analyze the data. He was also the lead author of the study, which was published in Communications Engineering, part of the Nature Portfolio.
Best, who formerly worked at Australia’s National Acoustic Laboratories, helped design the study. She says testing new technologies like BOSSA with real people is essential. “Ultimately, the only way to know if a benefit will translate to the listener is via behavioral studies,” Best says. “That requires scientists and clinicians who understand the target population.”
Big Potential for Hearing Technology
An estimated 50 million Americans live with hearing loss, and the World Health Organizationpredicts that by 2050, nearly 2.5 billion people worldwide will be affected. That makes the need for better hearing solutions urgent.Sen has patented BOSSA and hopes to partner with companies that want to bring it to market. He believes that major tech players entering the hearing aid space—like Apple with its AirPod Pro 2, which includes hearing aid features—will drive innovation forward. “If hearing aid companies don’t start innovating fast, they’re going to get wiped out,” says Sen. “Apple and other start-ups are entering the market.”
Individual participant audiograms. The different curves show pure-tone thresholds for each of the eight participants (averaged over left and right ears). Unique symbols distinguish individual subjects. (CREDIT: Kamal Sen, et al.)
And the timing couldn’t be better. As hearing technology becomes more widely available and advanced, tools like BOSSA could help millions of people reconnect with the world around them. From social events to everyday conversations, better sound separation can mean a better life.
Beyond Hearing Loss: A Wider Application
BOSSA was built to help those with hearing difficulties, but its potential doesn’t end there. The way the brain focuses on sound—what researchers call “selective attention”—matters in many conditions. “The [neural] circuits we are studying are much more general purpose and much more fundamental,” Sen says. “It ultimately has to do with attention, where you want to focus.”That’s why the team is now exploring how the same science could help people with ADHD or autism. These groups also struggle with multiple competing inputs—whether sounds, visuals, or tasks—and may benefit from tools that help guide attention.
They’re also testing a new version of BOSSA that adds eye-tracking. By following where someone looks, the device could better figure out who they’re trying to listen to. This could make the technology even more effective in fast-paced, real-world settings.
Sharpening Sound, Changing Lives
The success of BOSSA offers real hope. It’s not just another upgrade in hearing tech—it’s a shift in how we approach sound processing. Instead of trying to boost all sound or block background noise blindly, it takes cues from biology, using the brain’s blueprint to help listeners find meaning in the noise.For many with hearing loss, this could change everything. Being able to join conversations, pick out voices, and stay connected socially are vital parts of daily life. With tools like BOSSA, those goals move a little closer. And as this technology continues to grow, its reach may extend beyond hearing loss, offering help with focus and attention challenges too.
What started as a solution for a noisy dinner party could one day reshape how we interact with the world.
Brain-inspired algorithm helps hearing aids cut through the noise
Boston researchers created a brain-inspired hearing algorithm that helps users tune into a single voice in noisy environments.
www.thebrighterside.news
Brain-inspired algorithm helps hearing aids cut through the noise
Boston researchers created a brain-inspired hearing algorithm that helps users tune into a single voice in noisy environments.
www.thebrighterside.news
Trouble hearing in noisy places and crowded spaces? Researchers say new algorithm could help hearing aid users
BU researchers develop a brain-inspired algorithm that can help people with hearing loss pick out conversations in noisy, crowded spaces.
www.sciencedaily.com
Here, we present a system employing a novel strategy for stimulus reconstruction from neural spikes. Conceptually, this strategy uses time-frequency masking by computing a spike-based mask (spike-mask). We first consider the strategy for one-dimensional stimulus (e.g. sound waves). We show how this stimulus reconstruction method can be applied, using the cortical model as an example. We also show that this strategy produces reconstructions with intelligibility and quality higher than those reconstructed from the linear filtering method (table 1). Then we discuss how this strategy may be generalized for multi-dimensional stimulus (e.g. images and videos). The strategy presented here may be generalized to perform reconstruction on both artificial SNNs and neural models from experimental data as long as they satisfy the assumptions for our model.
Attachments
"
I asked ChatGPT for a neutral take on this.
What I thought was interesting, was that it weaved in Anduril's headset into the response."
That's how ChatGPT works. It uses all your previous conversations with it to personalize your answer.
Tothemoon24
Top 20
We are pleased to welcome Dr. Sasskia Brüers Freyssinet, AI Research Scientist at BrainChip , for the next lecture in our Current Topics in Digital Neuroscience series:
From Neuroscience to Applied Artificial IntelligenceIn this talk, Dr. Brüers-Freyssinet will explore the path from neuroscience research to applied AI, showing how insights from brain science can inspire more efficient and adaptive artificial systems. Drawing from her own experience transitioning from academia to industry, she will discuss the technical challenges, emerging research directions, and interdisciplinary skills needed to bridge fundamental brain research with real-world AI applications.
About the speakerDr. Brüers-Freyssinet is an AI Research Scientist at Brainchip, where she develops efficient computer vision models for neuromorphic hardware using techniques such as pruning, quantization, and regularization. She completed her PhD in Neuroscience at Université Paul Sabatier Toulouse III in 2017, studying the oscillatory correlates of conscious perception. Her current work investigates state-space models for temporal problems like gesture recognition and EEG-based behavior prediction—advancing the frontier where neuroscience meets artificial intelligence.
Tuesday 28th of October 2025
Per21, Université de Fribourg - Universität Freiburg Pérolles Campus, Room E140.Follow the talk online: https://lnkd.in/eQPKxAUX
Join us for a fascinating discussion on how brain-inspired approaches are shaping the next generation of intelligent, energy-efficient AI systems.
CHIPS
Regular
View attachment 92437
We are pleased to welcome Dr. Sasskia Brüers Freyssinet, AI Research Scientist at BrainChip , for the next lecture in our Current Topics in Digital Neuroscience series:
In this talk, Dr. Brüers-Freyssinet will explore the path from neuroscience research to applied AI, showing how insights from brain science can inspire more efficient and adaptive artificial systems. Drawing from her own experience transitioning from academia to industry, she will discuss the technical challenges, emerging research directions, and interdisciplinary skills needed to bridge fundamental brain research with real-world AI applications.
Dr. Brüers-Freyssinet is an AI Research Scientist at Brainchip, where she develops efficient computer vision models for neuromorphic hardware using techniques such as pruning, quantization, and regularization. She completed her PhD in Neuroscience at Université Paul Sabatier Toulouse III in 2017, studying the oscillatory correlates of conscious perception. Her current work investigates state-space models for temporal problems like gesture recognition and EEG-based behavior prediction—advancing the frontier where neuroscience meets artificial intelligence.
Follow the talk online: https://lnkd.in/eQPKxAUX
Join us for a fascinating discussion on how brain-inspired approaches are shaping the next generation of intelligent, energy-efficient AI systems.
View attachment 92436
Thank you, but where is the Link?
I asked Chat GPT5 . We already knew the answer but it's a good reminder that BRN should do very, very well.
What would be the likely result if the US rejected Neuromorphic Edge AI for Defense and the Golden Dome but it was adopted by China and Russia?
In short: If the U.S. rejected neuromorphic edge AI for defense while China and Russia embraced it, the likely result would be a significant strategic disadvantage for the U.S., with adversaries gaining faster, more adaptive, and more autonomous military capabilities.
Bottom line: Rejecting neuromorphic edge AI would not be a neutral choice for the U.S.—it would be a strategic self-handicap. In modern warfare, where milliseconds matter, adversaries with neuromorphic-enabled systems could seize the initiative, leaving the U.S. struggling to catch up.
Would you like me to sketch out a scenario analysis (e.g., Taiwan Strait conflict, NATO air defense, Middle East missile defense) to illustrate how this imbalance might play out in practice?
What would be the likely result if the US rejected Neuromorphic Edge AI for Defense and the Golden Dome but it was adopted by China and Russia?
In short: If the U.S. rejected neuromorphic edge AI for defense while China and Russia embraced it, the likely result would be a significant strategic disadvantage for the U.S., with adversaries gaining faster, more adaptive, and more autonomous military capabilities.
Key Consequences
- Loss of Technological Edge Neuromorphic AI mimics the brain’s neural architecture, enabling ultra-low power, real-time decision-making at the edge (without cloud reliance). If China and Russia deployed this in defense systems, they could achieve faster targeting, better situational awareness, and more resilient autonomous platforms than U.S. forces relying on conventional AI or human-in-the-loop systems.
- Erosion of Deterrence The U.S. military’s deterrence relies heavily on maintaining a qualitative technological advantage. Rejecting neuromorphic AI while adversaries adopt it could undermine U.S. credibility, especially in contested regions like the South China Sea or Eastern Europe.
- Vulnerability of Defensive Shields (e.g., “Golden Dome”) A neuromorphic-enabled missile defense system could process sensor data faster and adapt to novel attack patterns (like swarms of hypersonic or drone threats). If China and Russia fielded such systems while the U.S. stuck with slower, less adaptive defenses, American missile shields could be outpaced and overwhelmed.
- Acceleration of the AI Arms Race Analysts already describe the U.S.–China rivalry as an AI arms race. A U.S. decision to abstain from neuromorphic AI would not slow this race—it would simply cede leadership. China and Russia could then set the standards, export the technology to allies, and reshape the global defense balance.
- Operational Risks for U.S. Forces In battlefield scenarios, neuromorphic AI could allow adversaries to:
- Deploy autonomous drones with longer endurance and faster reaction times.
- Run electronic warfare systems that adapt in real time to U.S. countermeasures.
- Enhance command-and-control networks with near-instantaneous decision loops. This would leave U.S. forces reacting rather than dictating the tempo of conflict.
Strategic Outlook
- China and Russia’s Gain: Adoption would give them asymmetric advantages—low-power, adaptive AI that thrives in environments where cloud connectivity is denied (e.g., electronic warfare zones).
- U.S. Loss: By rejecting the technology, the U.S. would risk falling behind in both offensive and defensive capabilities, undermining its alliances and deterrence posture.
- Global Impact: Allies might gravitate toward Chinese or Russian defense ecosystems, fracturing U.S.-led security architectures like NATO or Indo-Pacific partnerships.
Bottom line: Rejecting neuromorphic edge AI would not be a neutral choice for the U.S.—it would be a strategic self-handicap. In modern warfare, where milliseconds matter, adversaries with neuromorphic-enabled systems could seize the initiative, leaving the U.S. struggling to catch up.
Would you like me to sketch out a scenario analysis (e.g., Taiwan Strait conflict, NATO air defense, Middle East missile defense) to illustrate how this imbalance might play out in practice?
Tothemoon24
Top 20
CHIPS
Regular
It was not even funny the first time!
If you can copy and paste the whole article, you must also be capable of copying and pasting the link. I judge you to be smart enough for that.
Everybody here posts the corresponding links; only you are so inconsiderate as not to do it. Why?
Agreed 100%!It was not even funny the first time!
If you can copy and paste the whole article, you must also be capable of copying and pasting the link. I judge you to be smart enough for that.
Everybody here posts the corresponding links; only you are so inconsiderate as not to do it. Why?
SharesForBrekky
Regular
Government’s science minister backs Aston University-led neuromorphic computing research
The UK...
pressat.co.uk
"The UK Multidisciplinary Centre for Neuromorphic Computing is headed by the Aston Institute of Photonic Technologies (AIPT) and was officially launched at the House of Lords on 21 October."
"Neuromorphic computing, inspired by the brain's really remarkable energy efficiency, could fundamentally transform how AI operate” - Lord Patrick Vallance"
"As well as speakers from politics and science, a message from the Minister for Science, Innovation, Research and Nuclear, Lord Patrick Vallance was delivered to the guests at the Houses of Parliament reception. Speaking via video he said that he believed the initiative could be key to tackling the issue of massive energy use by data centres. Current government figures show that they use about 2.5% of the UK's electricity, a figure projected to rise significantly with the further advances in AI."
"The centre will be led by the AIPT and will include world-leading researchers from Aston University, the University of Oxford, the University of Cambridge, the University of Southampton, Queen Mary University of London, Loughborough University and the University of Strathclyde. The centre will be supported by a broad network of industry partners to enhance the centre’s impact on society. These include Microsoft Research, Thales, BT, QinetiQ, Nokia Bell Labs, Hewlett Packard Labs, Leonardo, Northrop Grumman and a number of small to medium enterprises."
Fullmoonfever
Top 20
Tothemoon24
Top 20
Oh here we go !!!!It was not even funny the first time!
If you can copy and paste the whole article, you must also be capable of copying and pasting the link. I judge you to be smart enough for that.
Everybody here posts the corresponding links; only you are so inconsiderate as not to do it. Why?
Good old CHIP on the shoulder .
You said ; ‘ I judge you to be smart enough for that ‘ - wow what a wanker .
For your information I didn’t copy & paste & simply took a screen shot !!
You said ; everybody here post links , only you are so inconsiderate as to not do it .-
I post links where I see fit & will continue to do so if you care to take a look at the 1st post on this page as an example. Theirs plenty of links supplied for your pleasure.
Frangipani
Top 20
View attachment 92437
We are pleased to welcome Dr. Sasskia Brüers Freyssinet, AI Research Scientist at BrainChip , for the next lecture in our Current Topics in Digital Neuroscience series:
In this talk, Dr. Brüers-Freyssinet will explore the path from neuroscience research to applied AI, showing how insights from brain science can inspire more efficient and adaptive artificial systems. Drawing from her own experience transitioning from academia to industry, she will discuss the technical challenges, emerging research directions, and interdisciplinary skills needed to bridge fundamental brain research with real-world AI applications.
Dr. Brüers-Freyssinet is an AI Research Scientist at Brainchip, where she develops efficient computer vision models for neuromorphic hardware using techniques such as pruning, quantization, and regularization. She completed her PhD in Neuroscience at Université Paul Sabatier Toulouse III in 2017, studying the oscillatory correlates of conscious perception. Her current work investigates state-space models for temporal problems like gesture recognition and EEG-based behavior prediction—advancing the frontier where neuroscience meets artificial intelligence.
Follow the talk online: https://lnkd.in/eQPKxAUX
Join us for a fascinating discussion on how brain-inspired approaches are shaping the next generation of intelligent, energy-efficient AI systems.
View attachment 92436
Note, though, that Sasskia Brüers-Freyssinet has been signalling on LinkedIn that she is “Open To Work” ever since Olivier Coenen was sacked a month ago.
So it looks as if she is planning on leaving BrainChip sooner than later…
… just like her Toulouse office colleagues Matthieu Hernandez…
Matthieu HERNANDEZ - BrainChip | LinkedIn
Programmeur C++/C# passionné par l'intelligence artificielle , le deep learning et les… · Experience: BrainChip · Education: Université Paris Descartes · Location: Toulouse · 107 connections on LinkedIn. View Matthieu HERNANDEZ’s profile on LinkedIn, a professional community of 1 billion members.
… and Fares Ernez, whose LinkedIn profile picture has shown for months that he is “Open To Work”.
In August, he even liked a critical LinkedIn comment by a BrainChip shareholder (and TSE forum member), which shows that even some BrainChip employees share the frustration about the lack of new IP licences…
BrainChip's Akida technology enables on-device eye-tracking with high accuracy and low power consumption. | BrainChip posted on the topic | LinkedIn
BrainChip’s Akida™ neuromorphic technology is enabling ultra-low-power, always-on eye-tracking, with high-level accuracy all on-device to keep data private. From driver safety systems to wearables and AR, this breakthrough makes real-time, gaze-aware experiences possible without the cloud...
Just a reminder that we already lost long-term employee Richard Chevalier from our Toulouse office last month…
Sadly, Richard Resseguie is not the only Richard to have left BrainChip this month (saying that, hopefully there won’t be any Richard III, other than in some of our bookshelves…):
Richard Chevalier, who had been with our Toulouse-based office since 2018 (!), has joined Nio Robotics (formerly known as Nimble One, not to be confused with Nimble AI) as their Vice President of Platform Engineering:
#niorobotics | Richard Chevalier | 17 comments
I’m excited to share that I will be starting a new chapter at Nio Robotics this Monday. I feel both thrilled and humbled by the maturity, technical expertise, and excellence of the teams, and I look forward to contributing to this journey. #NioRobotics | 17 comments on LinkedInwww.linkedin.com
View attachment 91392
Richard Chevalier - Nio Robotics (formerly Nimble One) | LinkedIn
¤ Management Skills - Multicultural mindset with experience in multisite… · Experience: Nio Robotics (formerly Nimble One) · Education: Ecole nationale supérieure des Télécommunications · Location: Greater Toulouse Metropolitan Area · 461 connections on LinkedIn. View Richard Chevalier’s...
View attachment 91394
View attachment 91393
One can only hope that he will spruik BrainChip’s technology to his new employer - as far as I can tell, there is no indication that Nio Robotics have already been exploring neuromorphic computing for their robots.
Nio Robotics is currently building Aru, a shape-shifting mobile robot for industrial environments, but also say in their self-description on LinkedIn that they are “reinventing movement to create the first robotic assistant for homes”.
Something our CTO Tony Lewis, who is also a robotics researcher, will likely be very intrigued about.
Robotic maintenance and inspection with interactive capabilities solutions - Nio
Nio created autonomous and polymorphic robot called Aru for industrial applications, providing advanced solutions to automate routine inspections and enhance operational efficiency and maintain industrial infrastructure.www.nio-robotics.com
View attachment 91396
View attachment 91398
View attachment 91400
View attachment 91401
View attachment 91402
Watch Aru in action below, how it climbs stairs, reshapes to avoid obstacles, opens doors etc.
… who followed another ex-BrainChip employee from the Toulouse office - Alexandre d’Alton - to Nio Robotics:
Alexandre d'Alton - Greater Toulouse Metropolitan Area | Professional Profile | LinkedIn
Location: Greater Toulouse Metropolitan Area · 500+ connections on LinkedIn. View Alexandre d'Alton’s profile on LinkedIn, a professional community of 1 billion members.
Interestingly, Alexandre d’Alton’s last five months at BrainChip overlapped with his first five months at Nio Robotics (or Nimble One as it was called at the time).
But as I said in my above post on Richard Chevalier’s change of employers:
“One can only hope that he will spruik BrainChip’s technology to his new employer - as far as I can tell, there is no indication that Nio Robotics have already been exploring neuromorphic computing for their robots.”
Maybe the above is a hint Nio Robotics were exploring Akida in the first half of 2024 after all? On the other hand, Alexandre D’Alton may have had two concurrent jobs for a while that were totally unrelated?
With two former BrainChip engineers now working for Nio Robotics, it’s definitely worth keeping an eye on this company, though.
Let’s keep our fingers crossed, then, that two other very talented Toulouse-based BrainChip researchers, namely Douglas McLelland and Gilles Bézard, will continue to contribute to what our company has to offer.
Frangipani
Top 20
According to Athos Silicon Co-Founder and former Mercedes-Benz North America mSoC Chief Architect François Piednoël, Akida does not pass the minimum requirements for Level 3 or Level 4 automated driving.
#ai #artificialintelligence #ainews #aichips | AI & Robotics | 20 comments
Mercedes-Benz AG has launched Athos Silicon, a new chip company focused on developing energy-efficient chips for autonomous vehicles. Athos Silicon, based in Santa Clara, California, is a spin-off from Mercedes-Benz's Silicon Valley research arm and aims to create safer and more power-efficient...
View attachment 92002
Good question. I suppose you would need to ask François Piednoël directly whether what he means is possibly that there is an Akida-inherent problem, which would also concern the IP. (The BRN shareholder on LinkedIn was asking “Why isn’t Mercedes using Akida technology, for example?”, referring to Akida technology in general, not only to physical chips such as AKD1000 or AKD1500).
Apart from that, since Athos Silicon has so far not signed any IP deal with us, we can’t be in the first mSoC silicon codenamed Polaris anyway that he was referring to in this recent video:
ipXchange Interview with Athos Silicon Chief mSoC™ Architect Francois Piednoel
IpXchange spotlights Athos Silicon's mSoC™ as a safety-first unified control fabric that combines integrated redundancy, real-time voting, and energy-efficient execution to advance certified autonomy across vehicles, robotics, and mission-critical systemswww.athossilicon.com
View attachment 92176
Athos Silicon: Multiple System-on-Chip for Safe Autonomy
Athos Silicon’s Multiple System-on-Chip (mSoC) delivers functionally safe chiplet-based compute for autonomous driving and robotics.ipxchange.tech
Athos Silicon: Multiple System-on-Chip for Safe Autonomy
By Luke Forster
Published
1 October 2025
Written by Luke Forster
Building functional safety into compute for autonomy
Athos Silicon, a spin-out from Mercedes-Benz, is addressing one of the most pressing challenges in autonomous systems: how to eliminate the single point of failure in compute architectures. Unlike traditional monolithic processors that can collapse if a single component fails, Athos Silicon’s Multiple System-on-Chip (mSoC) integrates redundancy directly into silicon.
The result is a functionally safe Processor platform designed to meet ISO 26262 and other standards required for safety-critical applications.
Why safety-first design is essential
Conventional computing platforms – with a CPU, GPU, and NPU working together – were never built for Automotive safety. If a processor crashes or a transient error occurs, the entire system may fail. In a consumer PC this means a reboot; in a self-driving vehicle or industrial robot, it could mean disaster.
Athos Silicon has rethought this architecture from the ground up. By focusing on functional safety as a primary design constraint, its mSoC avoids the patchwork redundancy of external systems and instead bakes resilience into the hardware itself.
The mSoC architecture explained
Athos Silicon’s mSoC integrates multiple chiplets into one package, each containing CPUs, controllers, and memory. Instead of a single supervisor chip that itself could fail, the mSoC operates through a voting mechanism — what Athos calls a “silicon democracy.”
Each chiplet executes tasks in parallel, and their outputs are compared in real time. If one diverges from the others, it is overruled and reset. This ensures continuous operation without interruption and prevents cascading system failures.
By embedding this redundancy, Athos Silicon enables High Reliability computing suitable for Level 3 and Level 4 autonomy while maintaining predictable performance.
Power efficiency for EVs and robotics
Safety is not the only benefit. In electric vehicles, compute power directly affects range. Athos Silicon highlights that every 100 watts of compute load can reduce EV range by as much as 15 miles. By designing a chiplet system optimised for Low Power efficiency, the mSoC reduces unnecessary energy consumption and makes autonomy more practical for battery-powered platforms.
From Mercedes-Benz R&D to startup scale
The technology behind Athos Silicon was incubated within Mercedes-Benz before the company was spun out to bring the platform to the wider market.
Its first silicon, codenamed Polaris, is designed to deliver Level 3 and Level 4 autonomous capability in a footprint comparable to current Level 2 hardware.
Working with chiplet-packaging partners, Athos Silicon has accelerated validation and plans to deliver silicon to early customers soon. With no competitors currently offering integrated voting redundancy in a chiplet-based compute platform, Athos Silicon is carving out a unique position in the AI ecosystem.
Applications beyond cars
While autonomous driving is the most visible use case, Athos Silicon’s architecture also applies to Robotics, avionics, and even Medical devices where safety and reliability are paramount. Any system requiring certifiable, functionally safe compute stands to benefit.
By combining chiplet redundancy, real-time voting, and safety-first design, Athos Silicon’s Multiple System-on-Chip may prove to be the missing hardware foundation for truly certifiable autonomy.
This is what the Polaris mSoC will roughly look like sizewise (watch from around 10:50 min):
View attachment 92177
According to François Piednoël, “Project mSoC” as such started in 2020 (still under Mercedes Benz North America R&D).
Not sure what exact date the interview was recorded, but given that Athos Silicon as a Mercedes-Benz spin-off has been around since April 2025, and in the video it is being said the company is about four months old, it must have been recorded sometime between late July and early September.
So when François Piednoël says “In fact, there is silicon coming back shortly. By the end of the summer we’ll have the chiplets in hands” (from 9:06 min), this means they would have received them by now, if everything went according to plan. (“We think we are in good shape for a startup - getting your silicon after, you know, five six months (…) With no visible competition, by the way.”)
He also says, they invented the IP.
Reality check continued with regard to the question whether or not Akida might be integrated into Athos Silicon’s mSoC codenamed Polaris - you’ve now heard it twice
from the horse’s mouth…3AM , goals of the day done, chalenge of the day was to only add data in the occupation delta cubes when they are occupied, and not when they are not, massive reduction in complexity, only processing… | François Piednoel de Normandie
3AM , goals of the day done, chalenge of the day was to only add data in the occupation delta cubes when they are occupied, and not when they are not, massive reduction in complexity, only processing occupation is a lot faster.
Last edited:
Fullmoonfever
Top 20
Fortiss just uploaded the baseline pipeline on GitHub for the ESA Anomaly Detection Hackathon 2025.
Set up for deployment on Akida.
https://github.com/enterprise
nc-fortiss/2025_ESA_anomaly
Public
Baseline pipeline for ESA spacecraft telemetry anomaly detection. Includes full dataset loader, Akida-compatible CNN model, caching, normalization, and TensorBoard visualization.
Official baseline for ESA spacecraft telemetry anomaly classification, developed as an open starting point for the ESA Anomaly Detection Hackathon 2025.
This repository provides everything you need to get started — from data loading and preprocessing to model training, evaluation, and Akida-ready deployment.
github.com
cd ESA-Anomaly-Baseline
bash setup_akida_env.sh
Place the ESA-Mission1 dataset in the folder python train.py tensorboard --logdir logs/fit --port 6006
Contact Priya Kannan, Kannan@fortiss.org Neuromorphic Computing Group, fortiss GmbH (Munich)
Set up for deployment on Akida.
Navigation Menu
https://github.com/enterprise
nc-fortiss/2025_ESA_anomaly
Public
Baseline pipeline for ESA spacecraft telemetry anomaly detection. Includes full dataset loader, Akida-compatible CNN model, caching, normalization, and TensorBoard visualization.
nc-fortiss/2025_ESA_anomaly
| Name | |
|---|---|
|
2 hours ago | |
| .gitignore | 3 hours ago |
| Dataloader.py | 2 hours ago |
| LICENSE | 3 hours ago |
| README.md | 2 hours ago |
| plot.py | 2 hours ago |
| preview.py | 2 hours ago |
| setup_akida_env.sh | 2 hours ago |
| train.py | 2 hours ago |
Repository files navigation
2025_ESA_anomaly
Hackathon NCOfficial baseline for ESA spacecraft telemetry anomaly classification, developed as an open starting point for the ESA Anomaly Detection Hackathon 2025.
This repository provides everything you need to get started — from data loading and preprocessing to model training, evaluation, and Akida-ready deployment.
Overview
Spacecraft generate enormous amounts of multichannel telemetry data — detecting anomalies in these signals early can prevent mission failures. This baseline uses deep learning on time-series data to detect anomalies in the ESA-Mission1 dataset, and is designed for:- Satellite telemetry analysis
- Real-time anomaly detection
- BrainChip Akida neuromorphic deployment
Features
dataset loader (dataloader_tf.py) – reads 76 channel zip files, merges, resamples, labels, and windows- Akida-friendly CNN model (TensorFlow 2.15)
- Automatic caching and normalization
- 1-minute sampling rate and 14-year coverage
- Class balancing and focal loss ready
- TensorBoard visualizations
Project structure
ESA-Anomaly-Baseline/ │ ├── dataloader_tf.py # Dataset loader & preprocessor ├── train.py # Main training script ├── requirements.txt ├── LICENSE # MIT License ├── logs/ # TensorBoard runs ├── ESA-Mission1/ # Dataset root │ ├── channels/ │ ├── labels.csv │ ├── anomaly_types.csv │ ├── telecommands.csv │ └── channels.csv └── norm_stats.npzSetup
1. Clone the repository
GitHub - nc-fortiss/2025_ESA_anomaly: Baseline pipeline for ESA spacecraft telemetry anomaly detection. Includes full dataset loader, Akida-compatible CNN model, caching, normalization, and TensorBoard visualization.
Baseline pipeline for ESA spacecraft telemetry anomaly detection. Includes full dataset loader, Akida-compatible CNN model, caching, normalization, and TensorBoard visualization. - nc-fortiss/2025_...
github.com
bash setup_akida_env.sh
Place the ESA-Mission1 dataset in the folder python train.py tensorboard --logdir logs/fit --port 6006
Contact Priya Kannan, Kannan@fortiss.org Neuromorphic Computing Group, fortiss GmbH (Munich)
Similar threads
