BRN Discussion Ongoing

[Braintonic]

Hi Dolci,
Can't doubt that have been able to read the market very well with your charts. You have predicted the decline but interested to know when you think it will turn? What do you think will make this sky rocket? Or what indicators would you be looking for?

 

[Tothemoon24]

​

​

Audio for The Information - a first post. TENNs-PLEIADES: Building Temporal Kernels with Orthogonal Polynomials​

The Intelligence of Information Audio Podcast Review Research on Artificial Intelligence Research and interesting developments
ENGINEERING WORLD COMPANY
FEB 04, 2025
Start writing today. Use the button below to create a Substack of your own
Start a Substack
This is a first post I am writing here in Substack. This publication follows a former publication Blog of the same name: The Intelligence of Information. In that publication I’ve written extensively about recent developments in Artificial Intelligence, in particular artificial neural networks, convolutional neural networks and Deep Learning. I usually wrote about research and development papers and industry developments. Here this will continue the same type ofcontent (any changes will be notified if appropriate). We start today with a NoteboolLMgenerated review podcast for a paper I’ve found interesting going into the technical details. The paper is TENNs-PLEIADES: Building Temporal Kernels with Orthogonal Polynomials and obvioulsy I would suggest a detailed reading before listening to the generated podcast review. Read the paper: TENNs-PLEIADES. Consult the following GitHub resource: https://github.com/PeaBrane/Pleiades
Additional resources: Orthogonal Polynomialsand Jacobi Polynomials
The origin of this research paper is the public Neuromorphic company scale-up BrainChip Inc.
I encourage readers and listeners to leave comments in the comments sections of this post. It will be valuable for the publication and a way to improve future publications.

Audio for The Information - a first post. TENNs-PLEIADES: Building Temporal Kernels with Orthogonal Polynomials

The Intelligence of Information Audio Podcast Review Research on Artificial Intelligence Research and interesting developments

theintelligenceofinformation.substack.com

Next we publish a detailed NotebookLM generated research analysis Briefing review, which I also recommend for those interested in the technical details but haven’t the time to read the whole paper:

PLEIADES: Spatiotemporal Neural Networks with Orthogonal Polynomial Kernels​

This research paper introduces PLEIADES, a novel Temporal Neural Network (TENN) architecture designed for efficient spatiotemporal classification and object detection using event-based data. PLEIADES utilises orthogonal polynomial parameterisation of temporal kernels, enabling long-range temporal correlation capture with reduced memory and computational costs compared to traditional methods. The authors demonstrate state-of-the-art performance on three event-based benchmark datasets: DVS128 hand gesture recognition, AIS 2024 eye tracking, and the Prophesee GEN4 automotive detection dataset. The architecture’s modularity allows for easy adaptation to different sample rates without retraining, and the paper explores optimal computational strategies for efficient inference. The results highlight PLEIADES’ potential for low-latency, resource-efficient applications in event-based vision.

Briefing Document: PLEIADES - A Novel Temporal Neural Network
1. Introduction
This document summarises the key findings and concepts presented in the paper "Spatiotemporal_TENN.pdf," which introduces a new temporal neural network architecture called PLEIADES (PoLynomial Expansion In Adaptive Distributed Event-based Systems). PLEIADES is designed to process spatiotemporal data, especially event-based data, for online classification and detection tasks. The core innovation lies in using structured temporal kernels built from orthogonal polynomials, allowing for efficient computation, reduced memory usage, and adaptability to varying data sampling rates. PLEIADES is part of a broader class of networks known as TENNs (Temporal Neural Networks).
2. Key Themes & Concepts

• Temporal Convolutional Networks (TCNs) Limitations: The paper highlights the limitations of traditional TCNs, which often employ short temporal kernels (e.g., size 3). These short kernels struggle to capture long-range temporal dependencies and can be unstable during training with larger kernels. Parameterizing the kernels with MLPs is a solution, but can increase computational load significantly.
• "Temporal convolutional networks (TCNs) [...] have been a staple for processing time series data...However, in most cases, the temporal kernel is very short (usually size of 3), making it difficult for the network to capture long-range temporal correlations."
• PLEIADES: Polynomial Parameterisation of Temporal Kernels:PLEIADES addresses these limitations by parameterising temporal kernels using a weighted sum of orthogonal polynomials (specifically Jacobi polynomials), offering a more stable and efficient alternative to direct parameterisation.
• "Here, we introduce a method of parameterization of temporal kernels, named PLEIADES (PoLynomial Expansion In Adaptive Distributed Event-based Systems), that can in many cases reduce the memory and computational costs compared to explicit convolutions."
• Event-Based Data Focus: While PLEIADES can be used for various spatiotemporal data, the paper focuses on its performance with event-based data from event cameras. These cameras produce sparse data responding to changes in luminance at a microsecond time scale, providing rich temporal features.
• "Even though our network can be used for any spatiotemporal data (e.g. videos captured with conventional cameras), in this work, we investigate mainly the performance of our network on event-based data (e.g. data captured by an event camera)."
• Adaptable Sample Rate: A significant advantage of PLEIADES is its ability to adapt to different data sampling rates without retraining. This is achieved by re-discretizing the polynomial-based temporal kernels according to the new rate.
• "using this form of implicit parameterization will allow natural resampling of the kernels during discretization, meaning that the network can interface with data sampled at different rates without additional fine-tuning..."
• Optimal Operation Ordering: The paper emphasizes using Einstein summation (einsum) notation to optimize the order of operations during kernel generation and convolution. By strategically contracting tensors, the system can minimise both memory and computational costs.

"All the einsum operations are associative and commuta-tive, so we have full freedom over the order of contractions...In practice, we select the contraction path based on optimizing memory or computational usage..."

• (1+2)D Spatiotemporal Convolution Block: The core network is composed of (1+2)D spatiotemporal convolution blocks, factorised into temporal and spatial convolutions, allowing both to be depthwise separable, further reducing costs.
• "The main network block is a spatiotemporal convolution block, factorized as a (1+2)D convolution. In other words, we perform a temporal convolution on each spatial pixel followed by a spatial convolution on each temporal frame..."
• Causal Design: The architecture is designed to be fully causal, meaning it uses only past information to generate future predictions, making it suitable for online inference with minimal latency.
• "We keep every operation in our network fully causal, such that the network can be easily adapted for online inference with minimal latency. Importantly, we perform only causal temporal convolutions."

3. Technical Details

• Jacobi Polynomials: The temporal kernels are built using weighted sums of Jacobi polynomials which offer mathematical orthogonality. The orthogonality of the polynomials allows better inductive biases during training.
• "Jacobi polynomials P (α,β) n (τ) are a class of polynomials that are orthogonal in the following sense: [...] A continuous function can be parameterized by taking the weighted sum of these polynomials up to a given degree N , where the weighting factors (or coefficients) {γ0, γ1, ..., γN} are trainable."
• Discretization: The continuous temporal kernels are discretized by evaluating the integrals of the polynomials over time bins of interest. This involves computing the difference between the anti-derivative of the function at the end and start points of a time bin. This discretization method allows the kernels to be easily resampled by re-calculating these values for different bin sizes.
• "One method is to take the integral of the temporal kernels over the time bins of interest... Under this discretization scheme, it is very easy to resample the temporal kernels (either downsampling or upsampling), to interface with data sampled at arbitrary rates..."
• Einsum Optimisation: The use of einsum notation facilitates the flexible optimisation of the order in which tensor contractions occur. By first projecting the input features to the polynomial basis, the order of operations can be optimised for lower memory and computation costs.
• "For example, we can first generate the temporal kernels from the orthogonal polynomials, then perform the convolutions with the input features...But equally valid, we can also first project the input features onto the basis polynomials separately, then weigh and accumulate these results using the polynomial coefficients."
• Network Architecture: The spatiotemporal block uses depthwise separable temporal and spatial convolutions, with Group Normalisation and Batch Normalisation used after temporal and spatial convolutions respectively. ReLU activations are used after every convolution. Detection tasks use a temporally smoothed CenterNet detection head.

4. Experimental Results
PLEIADES achieved state-of-the-art results on three event-based benchmarks:

• DVS128 Hand Gesture Recognition:Achieved 99.59% accuracy with 192K parameters (100% with additional output filtering). Performed better than a Conv(1+2)D network with a smaller parameter count. Performance under different step-sizes was stable after re-discretisation of the kernels.
• "We achieved: 1) 99.59% accuracy with 192K parameters on the DVS128 hand gesture recognition dataset and 100% with a small additional output filter..."
• AIS 2024 Event-Based Eye Tracking:Achieved 99.58% test accuracy with 277K parameters.
• "2) 99.58% test accuracy with 277K parameters on the AIS 2024 eye tracking challenge..."
• PROPHESEE 1 Megapixel Automotive Detection: Achieved 0.556 mAP with 576k parameters.
• "3) and 0.556 mAP with 576k parameters on the PROPHESEE 1 Megapixel Automotive Detection Dataset."

The results were obtained with significantly fewer parameters compared to other state of the art networks. Additionally, the paper explored the tradeoff between accuracy and latency for gesture recognition by masking the start of frames and changing step sizes, showing flexibility of the model to be used in low-latency systems.
5. Limitations & Future Directions

• Memory Cost: The network, like any convolutional network has a high memory cost for explicitly buffering the moving window of recent input features.
• Future Work: The paper suggests that a running projection of input data onto the polynomial basis can be used to compress the input buffer. Future work may explore converting PLEIADES into a recurrent network for better online inference efficiency, and adapting the architecture into a spiking system for efficient hardware implementations.
• "Currently, the network is configured as a standard neural network, which by itself is already ultra-light in memory and computational costs. To truly leverage the full advantage of event-based processing, we can consider using intermediate loss functions to promote activation sparsity. Another direction is to adapt/convert this architecture into a spiking system..."

6. Conclusion
PLEIADES presents a promising new approach to processing spatiotemporal data, offering state-of-the-art performance with low computational and memory footprints. It achieves these results through the novel use of structured temporal kernels based on orthogonal polynomials. It is highly adaptable to varying data sample rates and it can be easily adapted to online inference for fast, low-latency responses. It opens new avenues for event-based systems and other spatiotemporal applications where resources are limited.
NotebookLM can be inaccurate; please double-check its responses.

 

[Rach2512]

Sorry, but I think you lack a moral compass Dolci.

We all have a choice in how we communicate to one another, and you have consistently chosen to do so by laughing at others' misfortunes.

Given the wealth you’ve gained from this stock, wouldn’t it be more humble for you to step back and graciously refrain from mocking those who are far less fortunate?

Perhaps graciously deciding not to "LOL" at other people could be a start.

You know, my beautiful father died suddenly very recently, not much more than one week ago actually, on the 23rd January at about 1am in the morning. He has always been my strongest advocate, my biggest influence and my brightest guiding light. I feel like I don't know what I will do without his guidance, but I also feel extremely grateful to have had such an incredible individual in my life. His greatest legacy to me was in teaching me the power of love.

"Love is what makes the world go around", he would remind me time and time again.

Laughing at others when they are struggling is not a virtue. Lifting people through love is.

B x

Oh Bravo, I'm so sorry to hear that, I think your father would be very proud of you, we all know here what a wonderful person you are. We all love your very generous and clever contributions and especially your wit and kindness, your Dad did a great job passing on the love. Xx

 

[manny100]

Sorry, but I think you lack a moral compass Dolci.

We all have a choice in how we communicate to one another, and you have consistently chosen to do so by laughing at others' misfortunes.

Given the wealth you’ve gained from this stock, wouldn’t it be more humble for you to step back and graciously refrain from mocking those who are far less fortunate?

Perhaps graciously deciding not to "LOL" at other people could be a start.

You know, my beautiful father died suddenly very recently, not much more than one week ago actually, on the 23rd January at about 1am in the morning. He has always been my strongest advocate, my biggest influence and my brightest guiding light. I feel like I don't know what I will do without his guidance, but I also feel extremely grateful to have had such an incredible individual in my life. His greatest legacy to me was in teaching me the power of love.

"Love is what makes the world go around", he would remind me time and time again.

Laughing at others when they are struggling is not a virtue. Lifting people through love is.

B x

Saddened by your loss Bravo. Your words about your father were inspiring and you were fortunate to have such a strong bond.

 

[CHIPS]

So yet another correction to an ASX submission. The continued mistakes/corrections is simply not good enough and does nothing to create an aura of professionalism around the company. Sean should be reviewing each and every submission before its goes to the regulator to ensure it is correct. Pick up your game SH!!

What did I miss? What do you mean?

 

[cosors]

Australia bans the DeepSeek application on government devices.

 

[BrainShit]

What did I miss? What do you mean?

The specific error corrected in the Director’s Interests Notice for Pia Turcinov was related to the “number of securities after change.” The original notice lodged on 04 January 2025 contained incorrect information regarding this figure, which has now been updated to accurately reflect her interests in the company's securities (site 3).

So, she did something useful and acquired shares as part of the update. Specifically, she acquired 204,813 Ordinary Shares and disposed of 204,813 Performance Rights.

 

[CHIPS]

The specific error corrected in the Director’s Interests Notice for Pia Turcinov was related to the “number of securities after change.” The original notice lodged on 04 January 2025 contained incorrect information regarding this figure, which has now been updated to accurately reflect her interests in the company's securities (site 3).

So, she did something useful and acquired shares as part of the update. Specifically, she acquired 204,813 Ordinary Shares and disposed of 204,813 Performance Rights.

Thank you @BrainShit
Those stupid mistakes again. Who is responsible for them? There should be a change in staff!

But the Notice itself is a good sign to me.

 

[IloveLamp]

Frontier Audio | LinkedIn

Frontier Audio | 1,371 followers on LinkedIn. Making critical frontline workers and warfighters safer, more comfortable, and more productive. | Frontier is on mission to make critical frontline workers and warfighters safer, more comfortable, and more productive.

www.linkedin.com

 

[IloveLamp]

Haven't come across this Brainchip employee before, she commented on Rudy's post, formerly....RTX

 

[DingoBorat]

Haven't come across this Brainchip employee before, she commented on Rudy's post, formerly....RTX View attachment 77225

"Ex" employee?..
Looks like she finished up in October?..

"Congrats Rudy, I need a new job too, nudge nudge, wink wink, say no more, say no more.."

 

[IloveLamp]

"Ex" employee?..
Looks like she finished up in October?..

"Congrats Rudy, I need a new job too, nudge nudge, wink wink, say no more, say no more.."

Where you see a glass half empty, i see a glass half full.

Formerly or not, it's not a bad thing

 

[SharesForBrekky]

Haven't come across this Brainchip employee before, she commented on Rudy's post, formerly....RTX View attachment 77225

Interesting find, after being at a company like Raytheon for over 3 years, Ruth decided to move to lil' ol' Brainchip.

I'm wondering if her moving in March 2023, gives any sign on timeframe of Brainchip's possible relationship with Raytheon.

 

[IloveLamp]

Interesting find, after being at a company like Raytheon for over 3 years, Ruth decided to move to lil' ol' Brainchip.

I'm wondering if her moving in March 2023, gives any sign on timeframe of Brainchip's possible relationship with Raytheon.

 

[DingoBorat]

Where you see a glass half empty, i see a glass half full.

Formerly or not, it's not a bad thing

I didn't say it was bad, but other than being a possible "previous" dot join to our relationship with RTX, who we are now partnered with, what benefit is it to us now, knowing she "was" an employee?

 

[IloveLamp]

I didn't say it was bad, but other than being a possible "previous" dot join to our relationship with RTX, who we are now partnered with, what benefit is it to us now, knowing she "was" an employee?

I think it adds more weight to or partnership with RTX

 

[DingoBorat]

Where you see a glass half empty, i see a glass half full.

Formerly or not, it's not a bad thing

The glass is always full of something!

 

[IloveLamp]

The glass is always full of something!

At least i provide research, all i see from you is your opinion everywhere

 

[DingoBorat]

At least i provide research, all i see is your opinion everywhere

Hey, I wasn't lambasting your research, I just didn't think "that" was significant.

But that's just my opinion.

Ohh ...

 

[IloveLamp]

Hey, I wasn't lambasting your research, I just didn't think "that" was significant.

But that's just my opinion.

Ohh ...