BRN Discussion Ongoing

[Vojnovic]

Many thanks Vojnovic greatly appreciated. It is a useful paper but what stood out for me was the following paragraphs which fits very nicely with the hypothesis that Valeo's new Lidar uses the AKIDA SNN technology:

"Point Cloud: Light Detection and Ranging (LiDAR)
sensors have recently gained prominence as state of
the art sensors in sensing the environment. They
produce a 3D representation of the objects in the field of
view as distances of points from the source. This collection
of points over a 3D space is called a 3D Point Cloud.
Though cameras have been used for a long time and
they provide a more direct representation of the surrounding,
LiDARs have gained ground because of some critical advantages
such as long range, robustness to ambient light conditions and
accurate localization of objects in 3D space. They produce sparse
data and hence suitable for SNNs....

We summarize the key benefits of SNN for automated driving:

• Event driven mechanism which brings adaptation for different scenarios.

• Low power consumption when realized as neuromorphic hardware.

• Simpler learning algorithm which leads to possibility of on-chip learning for longer term adaptation.

• Ability to integrate directly to analog signals leading to tightly integrated system.

• Lower latency in algorithm pipeline which is important for high speed braking and maneuvering.

4 Conclusion
Spiking Neural Networks (SNN) are biologically inspired
where the neuronal activity is sparse and event driven in order to
optimize power consumption. In this paper, we provide an overview of
SNN and compare it with CNN and argue how it can be useful in
automated driving systems. Overall power consumption over the driving
cycle is a critical constraint which has to be efficiently used especially for -(Remember what Mercedes said 6 to 10 times more efficient)
electric vehicles. Event driven architectures for various scenarios in
automated driving can also have accuracy advantages."

My opinion only DYOR
FF

AKIDA BALLISTA

Aside from the SNN aspect, I was also glad to see Brainchip explicitly mentioned in this Valeo paper. I knew you'd make great connections by reading this article. Thanks FF!

 

[Diogenese]

It was actually this reply from a moron on Twitter that got me thinking that there are probably a lot more people that haven’t quite grasped what we’re capable of….. So if it worked on him then maybe our competitors are in the same boat.

View attachment 1377

Sebastian seeems to have missed the whole point of EQXX.

You're just not getting it Sebastian!

 

[Fact Finder]

Many thanks Vojnovic greatly appreciated. It is a useful paper but what stood out for me was the following paragraphs which fits very nicely with the hypothesis that Valeo's new Lidar uses the AKIDA SNN technology:

"Point Cloud: Light Detection and Ranging (LiDAR)
sensors have recently gained prominence as state of
the art sensors in sensing the environment. They
produce a 3D representation of the objects in the field of
view as distances of points from the source. This collection
of points over a 3D space is called a 3D Point Cloud.
Though cameras have been used for a long time and
they provide a more direct representation of the surrounding,
LiDARs have gained ground because of some critical advantages
such as long range, robustness to ambient light conditions and
accurate localization of objects in 3D space. They produce sparse
data and hence suitable for SNNs....

We summarize the key benefits of SNN for automated driving:

• Event driven mechanism which brings adaptation for different scenarios.

• Low power consumption when realized as neuromorphic hardware.

• Simpler learning algorithm which leads to possibility of on-chip learning for longer term adaptation.

• Ability to integrate directly to analog signals leading to tightly integrated system.

• Lower latency in algorithm pipeline which is important for high speed braking and maneuvering.

4 Conclusion
Spiking Neural Networks (SNN) are biologically inspired
where the neuronal activity is sparse and event driven in order to
optimize power consumption. In this paper, we provide an overview of
SNN and compare it with CNN and argue how it can be useful in
automated driving systems. Overall power consumption over the driving
cycle is a critical constraint which has to be efficiently used especially for -(Remember what Mercedes said 6 to 10 times more efficient)
electric vehicles. Event driven architectures for various scenarios in
automated driving can also have accuracy advantages."

My opinion only DYOR
FF

AKIDA BALLISTA

As Vojnovic points out earlier in this paper Brainchip is acknowledged:

"Several implementations of deep SNNs on neuromorphic hardware
such as SpiNNaker and BrainChip have demonstrated sensor applications
that support this potential of SNNs."

I do not think I would be prepared to bet against Valeo being at least one of the Logo's that the new CEO obtains permission to put under
the "Early Adopters" heading right next to Mercedes.

Not sure this needs to be mentioned but Brainchip is the only one with a commercial chip and IP in the market place able to be purchased and implemented in Valeo's Lidar for ADAS and in fact the IP has been out there for years already.

My opinion only DYOR
FF

AKIDA BALLISTA

 

[butcherano]

Sebastian seeems to have missed the whole point of EQXX.

You're just not getting it Sebastian!

lol…I think he might be a distant German relative of our old mate Travis!..

 

[Fox151]

G’day @IndepthDiver . Great post. But I actually landed on only 2 real drivers behind Mercedes revealing themselves.

The first is based on the assumption that it’s well known that we’re working with a number of other automotive companies. And Mercedes simply wanted to ensure that they become renowned as the innovators and the rest as just followers.

And the second hinges on the competitive advantage and they’re trying to throw other competitors off the scent (that aren’t woking with Brainchip) to give the impression that Akida is only being used for voice recognition or “Hey Mercedes“…and not the plethora of other applications that we know Akida can and most likely is doing in the EQXX.

Mercedes were never announced as an EAP but Ford were... I wonder if they were trying to outdo Ford?

 

[Fact Finder]

It was actually this reply from a moron on Twitter that got me thinking that there are probably a lot more people that haven’t quite grasped what we’re capable of….. So if it worked on him then maybe our competitors are in the same boat.

View attachment 1377

Just thinking about Sebastian Schmitt's question surely the answer is "Not anymore."
FF.

 

[butcherano]

Mercedes were never announced as an EAP but Ford were... I wonder if they were trying to outdo Ford?

I actually thought that Ford might have been put out by the ASX forcing us to disclose their name and that now they‘re going through the likes of Renesas to avoid having to disclose any direct association with Brainchip…not sure really though….

 

[Diogenese]

Aside from the SNN aspect, I was also glad to see Brainchip explicitly mentioned in this Valeo paper. I knew you'd make great connections by reading this article. Thanks FF!

This table is just above the reference to BrainChip:

The paper was dated 2019, so it was published before the commercial version of Akida with the 4-bit weights/activations option, and we know that the commercial version exceeded performance expectations. Thus it is probable that the accuracy of the production version of Akida 1000 SNN is significantly better than shown in the table.

So, where speed and power consumption are more important that accuracy, Akida can be used in the 1-bit weights/activations (pure spiking mode)*, and where more accuracy is required, the usr can select up to 4-bit mode.

*I'd like to know how our engineers have implemented the 4-bit weights/activations. For example, have they incorporated a miniature ALU in each NPU? ... or can it be done with comparators?

No doubt all will be revealed in an upcoming patent specification.

 

[butcherano]

Just thinking about Sebastian Schmitt's question surely the answer is "Not anymore."
FF.

That’s probably the polite answer. When I first read his post I was thinking more along the lines of asking him to have a good look at himself in the mirror….

 

[Diogenese]

This table is just above the reference to BrainChip:

View attachment 1378
The paper was dated 2019, so it was published before the commercial version of Akida with the 4-bit weights/activations option, and we know that the commercial version exceeded performance expectations. Thus it is probable that the accuracy of the production version of Akida 1000 SNN is significantly better than shown in the table.

So, where speed and power consumption are more important that accuracy, Akida can be used in the 1-bit weights/activations (pure spiking mode)*, and where more accuracy is required, the usr can select up to 4-bit mode.

*I'd like to know how our engineers have implemented the 4-bit weights/activations. For example, have they incorporated a miniature ALU in each NPU? ... or can it be done with comparators?

No doubt all will be revealed in an upcoming patent specification.

I'm guessing we've implemented the 4-bit version with comparators with weighted outputs, so the least significant bit comparator has an output weight/multiplier of 1, the second least significant bit (LSB) has an output weight/multiplier of 2, the next most significant bit comparator has an output weight /multiplier of 4, and the most significant bit (MSB) comparator has an output weight of 8.

In other words, when the LSB comparator "fires" it generates 1 spike, whereas, when the MSB comparator fires, it generates 8 spikes.

{Amended 16 to 8 & pro-rata}

 

[gex]

I'm guessing we've implemented the 4-bit version with comparators with weighted outputs, so the least significant bit comparator has an output weight/multiplier of 1, the second least significant bit (LSB) has an output weight/multiplier of 4, the next most significant bit comparator has an output weight /multiplier of 8, and the most significant bit (MSB) comparator has an output weight of 16.

In other words, when the LSB comparator "fires" it generates 1 spike, whereas, when the MSB comparator fires, it generates 16 spikes.

Yeah I like stuff too.

Way over my head mate

 

[Fact Finder]

Yeah I like stuff too.

Way over my head mate

The strange thing is when years ago BarrellSitter used to speak dirty they were words I knew but they did not seem like they should be together in the same sentence but as time has gone by when he speaks as Dio he is finally making sense. So glad he finally understands the technology. LOL

Many thanks Dio. I am a tad confused regarding the 1 bit to 4 bit becoming 8 then 16. So you are not referring to bits or are you?

Thanks FF

 

[Diogenese]

Yeah I like stuff too.

Way over my head mate

Hi gex,

It's good stuff.

In the diagram, X is the activation and Y is the weight.

The comparators produce an output when both inputs are equal.

The output ranges from 1 spike (1 binary digital bit) for the LSB to 8 spikes for the MSB comparator.

[Version 2 with deliberate mistake removed]

 

[Equitable]

Regarding the topic of non-disclosure and 'customers' going through Renesas. Renesas are undoubtedly developing products which they will sell to a number of companies that might currently have NDAs with us. In addition, our customers will likely be developing their own systems to meet the specific needs of their particular products. They may be happy to buy 'off the shelf' products from Renesas when those products meet a need. However, when they have bespoke requirements I expect they will purchase our IP directly from us rather than paying Renesas a margin of 10 or 20% just to keep 'hiding' behind an NDA. When we consider the addressable markets and volumes of some of our potential customers, paying a fee to Renesas could become very expensive and affect their cost competitiveness.

I do understand the need for secrecy in the early days, but eventually Akida will be so widely used, and so widely regarded, that companies might actually prefer to say yes, we have Akida inside.

 

[Equitable]

That’s probably the polite answer. When I first read his post I was thinking more along the lines of asking him to have a good look at himself in the mirror….

Don't be too harsh on people who don't understand our technology yet. They do not have the benefit of our '1,000 eyes' and daily reading everything we have posted over the past few years.

The world has a lot of catching up to do.

 

[Diogenese]

The strange thing is when years ago BarrellSitter used to speak dirty they were words I knew but they did not seem like they should be together in the same sentence but as time has gone by when he speaks as Dio he is finally making sense. So glad he finally understands the technology. LOL

Many thanks Dio. I am a tad confused regarding the 1 bit to 4 bit becoming 8 then 16. So you are not referring to bits or are you?

Thanks FF

Hi FF,

In Akida's world, the world of digital spiking neural networks, the terms "bits" and "spikes" are interchangeable, each "spike" being represented by a single binary digital bit.

The diagram above represents a circuit for comparing 4-bit weights and activations. A 4-bit "byte" is a series of 4 digital bits, each of which can be a 1 or a zero:
0000
0001
0010
0011
0100
0101
0110
0111
1000
1001
1010
1011
...

The MSB is on the left, and the LSB is on the right. 0001 = 1, 0010 = 2, 0100 = 4, 1000 = 8.

The change from 1-bit to 4-bits in Akida meant that, instead of just multiplying single 1 bits of X and Y, the NPU now has to be able to multiply 4-bit "bytes".

This would normally be done in an arithmetic logic unit (ALU), or matrix algebra.

Using comparators with appropritely weighted outputs is an efficient method of converting the comparison of two "bytes" to a "spike" stream (a stream of digital 1s). For the 4-bit bytes to be recognized as being equal, each comparator must produce a 1 as an output, ie, the LSB X of the activation must equal the LSB Y of the weights, and so on up to the MSB X and Y.

 

[yogi]

As a enthusiast photographer on my google feed a news popped up about Sony Alpha mirrorless camera being used by Skyfish Drone and when i checked their website, was amazed about How Akida can be used by those guys

The Skyfish Platform - Skyfish

skyfish.ai

 

[Bravo]

Hi FF,

In Akida's world, the world of digital spiking neural networks, the terms "bits" and "spikes" are interchangeable, each "spike" being represented by a single binary digital bit.

The diagram above represents a circuit for comparing 4-bit weights and activations. A 4-bit "byte" is a series of 4 digital bits, each of which can be a 1 or a zero:
0000
0001
0010
0011
0100
0101
0110
0111
1000
1001
1010
1011
...

The MSB is on the left, and the LSB is on the right. 0001 = 1, 0010 = 2, 0100 = 4, 1000 = 8.

The change from 1-bit to 4-bits in Akida meant that, instead of just multiplying single 1 bits of X and Y, the NPU now has to be able to multiply 4-bit "bytes".

This would normally be done in an arithmetic logic unit (ALU), or matrix algebra.

Using comparators with appropritely weighted outputs is an efficient method of converting the comparison of two "bytes" to a "spike" stream (a stream of digital 1s). For the 4-bit bytes to be recognized as being equal, each comparator must produce a 1 as an output, ie, the LSB X of the activation must equal the LSB Y of the weights, and so on up to the MSB X and Y.

 

[Diogenese]

Thought I’d found a new discovery here, but upon a google search for more info it turns out our old mate @Fact Finder beat me to it and made the link back on 19/07/21 on HC. Damn he’s good

View attachment 1348

Anyway that predates my BRN days (unfortunately!) and I’d done the research already so thought I’d continue my post - for two reasons:

1. Share the knowledge with more recent holders who may not have seen on HC
2. Reignite the discussion, as I personally don’t know if it had/has legs. Hopefully FF has something more to add


It’s not a silver bullet but yet another interesting dot joined on our path to Brainchip glory
View attachment 1338

Milind Joshi is the Intellectual Property Officer at Brainchip - since May 2021

Prior to this Milind spent almost 7 years at Samsung in India - at their R&D Institute


View attachment 1340



View attachment 1341
View attachment 1343


9 months ago he asked on LinkedIn for:

recommendations for patent watch tool that sends email notification for each new patent publication or grant (USPTO and EPO at least) of a competing firm(s)?

View attachment 1344


More recently he liked this post by Mercedes about the Vision EQXX


View attachment 1345

Cheers
TLS

That's an impressive CV.

Maybe @uiux could apply for the patent watch job.

 

[Diogenese]

Slow day in the songwriters room?