Space Cadet
Regular
Interesting comment on Akida by Sounak Dey
https://www.linkedin.com/posts/soun...he-morpheus-activity-7195181143552401409-7O0N
SC
https://www.linkedin.com/posts/soun...he-morpheus-activity-7195181143552401409-7O0N
SC
BRN Discussion Ongoing
- Thread starter TechGirl
- Start date
Smoothsailing
Regular
My sentiments exactly the uptake is slow due to models with Akida not being available from the get go , eg a partnership with say a Chiplets design would work imo.I recently had the thought again of what it would have been like if Brainchip had a simple ready-made product that people understood, cost little and only served to make NC known to the masses. In my view, GPT was a huge success not because it can be very useful, but because it was easily accessible and even people with no premonition used it as a crystal ball or just to play with. The world got to know it in a short time.
Examples are not so easy for me to think of, but with a lot of imagination maybe something in the size factor like a refillable vapour for testing liquids and that can be trained and then trains itself. Imagine if there was such a thing and you went into a bar, put it in a cocktail and it tells you what it is and what's in it. Or others would use it for their whisky and others for their coffee, beer tasting or green tea to determine the perfect temperature with little bitterness and so on? So it's more of a playful character and to show that there are also simple applications and that you don't have to be a rocket scientist.
Or a simple camera that does nothing more than tell you what it has just photographed.
My idea is probably too naive. But market acceptance comes after the perception that something like this exists. GPT was thrown onto the market that didn't existed and that was exactly the right tactic. Suddenly so many people are starting to think about how they can use it for their own companies.
NPUs in smartphones, for example, are nothing new. But very few people think about the benefits. A low-cost device for simple recognition and without a cloud would drastically demonstrate the possibilities in a playful way, which is fun for people and automatically ensures acceptance. The market would emerge automatically.
Perhaps a better idea for the bar counters.
I don't want to be a heretic.
____
Imagine that at the beginning there is this little cheap camera that can do nothing but take pictures to recognize what is seen and the user has to tell it what it is. Later it recognises it all by itself and gets better and better at it. Then others play around with it because they see a market that nobody else 'sees', blind people. A blind person points a well-trimmed camera at a scene and with SLMs this camera tells what the blind person could be looking at, or what a deaf person could be hearing and is now reading on the small display. Myriads of applications just in this small case.
~~The customer only can buy what he sees, I have often said to myself over the years, or so I think.
___
The exchange on Reddit, for example, about these simple models. One to the other: I found something brilliant here and loaded it onto my Akida....
GPT is hip and that opens up its market.
It was the same with the video cassette. No end user was interested in the technical realisation. It was purely about that thing that was new. ...nobody can buy what nobody can see.
Think of GPT and first would existing only in labs and what it is today.
Stable Genius
Regular
M Anthony Lewis on LinkedIn: #tenns #mamba #brainchip
Yes to Edge AI! BrainChip has wracked up an impressive number of beyond state of the art models for edge inference. These models are based on our own TENNs…
Pom down under
Top 20
Must be FF
DJM263
LTH - 2015
Thanks @IloveLamp
Dominic Pajak on LinkedIn: #ai #voiceai #edgeai #opensource #embeddedsystems #machinelearning #nlp…
🚨 Wow. This is running 100% on-device — no cloud, no offloading. Just a private, responsive, contextual AI assistant 🤯 Imagine any appliance — in home…www.linkedin.com
View attachment 73980 View attachment 73981
AI Everywhere 2024
aieverywhere.eetimes.com
This will be very interesting to see if Akida is a part of this! . I'm looking forward to see if Dominic lets the cat out of the bag during his breakdown on Thursday.
Dominic Pajak on LinkedIn: #ai #voiceai #edgeai #opensource #embeddedsystems #machinelearning #nlp…
🚨 Wow. This is running 100% on-device — no cloud, no offloading. Just a private, responsive, contextual AI assistant 🤯 Imagine any appliance — in home…
Smoothsailing
Regular
Brilliant to hear these models are now being developed with various levels of intellect and can be implemented to customers portfolios moving forward.
M Anthony Lewis on LinkedIn: #tenns #mamba #brainchip
Yes to Edge AI! BrainChip has wracked up an impressive number of beyond state of the art models for edge inference. These models are based on our own TENNs…
This is what will open big doors for BRN.
Last edited:
Bravo
If ARM was an arm, BRN would be its biceps💪!
SNNs Power Energy-Efficient Place Recognition
By Muhammad OsamaReviewed by Susha Cheriyedath, M.Sc.
Dec 9 2024
A research article recently posted on arXiv preprint* server explored the potential of spiking neural networks (SNNs) in improving visual place recognition (VPR) tasks within the field of robotics. The researchers introduced innovative methods that leverage the unique features of SNNs to address challenges in place recognition, especially in changing environments. They examined how SNNs could overcome the limitations of traditional navigation methods, particularly in energy-sensitive settings.
Advancements in Neuromorphic Technology
SNNs represent a significant advancement in neuromorphic computing by closely mimicking biological neural systems. Unlike traditional artificial neural networks (ANNs), which process information using continuous values, SNNs communicate through discrete spikes triggered when a neuron's activation crosses a threshold. This spike-based approach enables SNNs to achieve high energy efficiency and low-latency processing, making them ideal for real-time robotics applications. These benefits are advantageous when implemented on neuromorphic hardware that processes information similarly to the human brain.Despite these advantages, SNNs face challenges that limit their potential, particularly in training and implementation. The non-differentiable nature of spiking neuron activation functions complicates supervised training, requiring specialized methods. SNNs show significant promise for VPR, a key component of robotic navigation.
VPR is crucial for tasks like localization and mapping, where robots must identify places despite changes in their appearance. It helps robots recognize previously visited locations and update their maps, even under varying conditions. VPR plays a critical role in applications such as loop closure detection in simultaneous localization and mapping (SLAM) and global re-localization of mobile robots. However, challenges like appearance changes due to time of day, seasonal shifts, weather variations, and perceptual aliasing make this task particularly complex.
Modular Spiking Neural Networks for VPR
Robotics & Automation eBook
In this paper, the authors proposed three major advancements in applying SNNs for VPR. First, they introduced modular SNN architecture, where each module represents a set of geographically distinct, non-overlapping places. This design enables scalable networks for large environments. Each modular SNN consists of only 1,500 neurons and 474,000 synapses, making them compact and ideally suited for ensemble configurations.
Second, the study developed ensembles of modular SNNs, where multiple networks represent the same place. This approach significantly improved accuracy compared to single-network models. It demonstrated that ensemble members exhibit higher variations in their match predictions, resulting in significantly higher responsiveness to ensembling.
Third, the researchers explored sequence matching, a technique that uses consecutive images to refine place recognition. This method demonstrated better responsiveness to ensembling than traditional VPR techniques, enhancing recognition accuracy under varying conditions. The modular SNNs were independently trained on distinct dataset segments to specialize in specific places, utilizing the biologically inspired Spike-Time Dependent Plasticity (STDP) learning rule for unsupervised learning.
Key Findings of Using Presented Methodologies
The proposed methods were evaluated on benchmark datasets, including Nordland, Oxford RobotCar, and SFU-Mountain, and compared against conventional VPR techniques like Sum-of-Absolute Differences (SAD), DenseVLAD, and NetVLAD. Performance metrics, such as recall at one (R@1), were used to assess effectiveness, with the ensemble of modular SNNs achieving a recall rate of 36.9% on the Nordland SW dataset with sequence matching, surpassing conventional VPR methods.The outcomes highlighted the potential of SNNs in VPR tasks. Specifically, modular SNNs showed strong performance across diverse datasets, effectively handling significant appearance changes and environmental variations. Ensembling SNN modules notably improved accuracy, with ensembles showing greater adaptability to input variations than single-module models. The ensemble of modular SNNs consistently outperformed individual modules across all datasets, confirming the strength of this approach.
Integrating sequence matching further enhanced performance, particularly in scenarios with substantial appearance changes. Additionally, higher recall rates indicated improved recognition of previously visited locations under challenging conditions. The authors emphasized the scalability of SNNs, which efficiently managed large datasets without increasing computational demands. This efficiency is crucial for real-world applications requiring resource optimization, as modular SNNs scaled effectively to datasets containing up to 10,000 places.
Applications
The advancements in SNNs can enhance robotic navigation systems, especially in environments where energy efficiency and real-time processing are crucial, such as space exploration or disaster recovery. Additionally, the modular and ensemble approaches can help develop robust autonomous vehicles capable of navigating complex urban environments.By leveraging the strengths of SNNs, these vehicles can improve localization, boosting safety and operational efficiency. Furthermore, this study can guide future progress in neuromorphic computing, leading to more advanced algorithms that mimic biological processes. This could result in breakthroughs in artificial intelligence, particularly in areas that require adaptive learning and real-time decision-making.
Conclusion and Future Directions
In summary, SNNs proved effective for VPR tasks, addressing the limitations of conventional deep learning methods, especially in energy-sensitive applications. They introduced modular and ensemble-based strategies that enhanced the accuracy and robustness of place recognition, offering a scalable solution for real-world deployment.Moving forward, the authors highlighted the potential for further exploration of neuromorphic hardware, which could significantly improve the performance and efficiency of SNNs. Future work should also focus on enhancing the resilience of these systems to viewpoint changes, expanding their applicability in diverse environments. Overall, this study paves the way for integrating SNNs into various robotic applications, promising significant advancements in energy-efficient navigation technologies.
SNNs Power Energy-Efficient Place Recognition
Modular spiking neural networks enhance visual place recognition, improving robotic navigation accuracy in changing environments through innovative techniques.
sleepymonk
Regular
1.8 million air force contract wooooohooooo
buena suerte :-)
BOB Bank of Brainchip
Whoop whoop 
...Happy days !!!
...Happy days !!!
Hi Baisyet,Morning @Diogenese just a confused question
I think that Jason has/had a narrow view of what constitutes neuromorphic, being analog neurons.
It is true that analog neurons more closely resemble biological neurons by adding actual voltages and firing when a threshold voltage value is reached. A leaky-integrate-and-fire (LIF) neuron can be represented by a capacitor in parallel with a resistor. As the capacitor is charged by input current spikes from other neurons, it begins to discharge (leak) through the resistor, so the rate of incoming spikes is important. If the rate is too low, the capacitor voltage will drop faster than it is built up by the spikes and the neuron will not reach the threshold.
Akida is designed to produce a digital imitation of the function of biological neurons. Akida's neurons "count" digital bits and fire when the count reaches a numerical threshold. The digital output of Akida is equivalent to the result of passing the output of an analog neuron through an analog-to-digital converter, which is a necessary step with an analog neuron if its output is to be useful in a CPU/GPU. Akida uses a time window to replicate the leak by only counting the most recent block of N spikes.
Akida is suitable for edge devices because of its speed of response (low latency) and its low power requirements which make it ideal for battery operated devices.
DingoBorat
Slim
I wanted a bike for Christmas and I'm getting a ping pong set and cowboy hat 
..
..
I'll take anything at this point...anything to stop the share price slide of late. And more validation to prove to the wife that I'm not insane!I wanted a bike for Christmas and I'm getting a ping pong set and cowboy hat
TheDrooben
Pretty Pretty Pretty Pretty Good
Air Force approaches industry for artificial intelligence (AI) and machine learning for C4ISR applications
Project's focus is on nano-computing; neuromorphic computing and machine learning; embedded deep learning; and efficient computing architectures.www.militaryaerospace.com
Hopefully we are submitting tenders for some of these juicy contracts.......
View attachment 73739
View attachment 73740
Happy as Larry
Oh yeah.........Could lead to some of the more juicy contracts with AFRL that I posted about previously. Take that shorters!!!!!
Happy as Larry
Similar threads
