BrainChip Holdings Ltd (BRN) Earnings Call Transcript & Summary

Good morning, good afternoon, good evening, depending from where you're joining us from. I'm Ira Feldman, and I'm the tinyML Foundation Managing Director. And I would like to welcome you to today's webinar as part of the tinyML Hackathon on Pedestrian Detection that we're running. Today, we're going to talk about the BrainChip Akida development kit. So thank you for joining us. Before we start, I would like to thank all of the tinyML strategic partners who make this and all our events possible. I'll tell you a little bit more about them at the end of today's webinar, but I'd like to thank all of them for helping out. And for those planning ahead, our next in-person event is June 26 to 28, our tinyML Europe, Middle East, Africa Innovation Forum being held June 26 to 28 in Amsterdam. Please register now. Advanced registration is until May 31, and we welcome everybody in person there. It will be a fabulous venue, and it's a great technical program with great networking opportunities. So please check it out and make plans to join us in Amsterdam. In terms of this tinyML Hackathon, we have resources in terms of dev kits. We've done 2 of these webinars. They are now available on our YouTube channel, youtube.com/tinyml. Sony talked about the AITRIOS Intelligent Vision System. Infineon talked about their extensive 60-gigahertz radar sensor. Today, we have BrainChip talking about the Akida platform. We will also have office hours and ask the expert sessions, and we will announce those via the contest website. In terms of schedule, we do have the optional checkpoints, including proposals due tomorrow for people who would like some feedback from the judges and then we will have data sets in June, models in July, and devices in August. And just a reminder, the final submissions are due on September 15. So please keep the schedule in mind and please follow along. In terms of the contest platform, you can go to this web page. It's being hosted by the United Nations International Technology [ UNU-ITU ] as part of their AI for Good initiative. So you can go here and this is where you put your teams together and have all the detailed information about the contest. So today, it is my pleasure to introduce Nikunj Kotecha. He is a Machine Learning Solutions Architect at BrainChip. He's spent the past few years immersed in the Akida Neuromorphic technology, and he's focused on bringing the power of their technology to assist implementing cutting-edge AI solutions. Prior to BrainChip, he worked at Oracle, where he was using new AI-driven processes to streamline clinical trials. And with that, I welcome Nikunj. Actually, before Nikunj starts, just a reminder, please use the Q&A window for your questions. You can put them in as he's presenting. And at the end, I will work through the questions. So go ahead and use the Q&A window for your questions. Nikunj?

Are you able to see my screen?

Yes.

Right. Thank you, Ira, and thank you, tinyML Foundation for organizing this Hackathon challenge. It's been a really exciting way where you guys are driving the machine learning community forward and encouraging all these new developers and existing developers to drive the community forward. Welcome, everyone. I'm Nikunj, and welcome to this webinar. We are going to be talking about our development kits and talk about how we can help you guys in making sure this challenge is completed by you guys. Before that, let me tell you a little bit about me. I'm a Solutions Architect at BrainChip. And I've been with BrainChip for over 2 years. I've worked with the Akida technology, trying to develop new models for the neuromorphic [ world ] and kind of worked with industry leaders to make sure we can integrate their cutting-edge solutions alongside Akida. And for today's agenda, we'll be talking about the problem statement again, so I'll just highlight the problem statement on what this challenge is all about. I'll give a brief introduction about our company, who we are and where we are heading to. I'll introduce Akida platform and talk about our development kits, so you guys have an understanding on what our platform is all about. We'll talk about how do you guys send in proposals so that you guys can get access to these development kits and can get them started alongside this challenge. And then our external sources are support platforms, where can you reach us if you need any help from us. We'll also have some sessions for the selected teams who have our development kits. So I'll talk about a little bit more on that. And towards the end, we'll have some session for questions and answers. All right. So this is taken from the kickoff that from tinyML. And initially, the policy talks about there have been a lot of pedestrians that's throughout the country and throughout the world. So we're trying to solve this in a way that AI can kind of assist a lot of these vehicles to detect moving people around or moving bicycles or anything that can avoid these type of accidents and mainly you can have different types of visions, maybe coming from vehicles, maybe coming from a flagpole. So there are a variety of different solutions that can be implemented. And they were 2 kind of examples given from the kickoff deck where, one, you have some accidents seen in major intersections where there are multiple different crossings. And even between intersections, you see some fatal injuries going on here and there. So we're kind of trying to address all of these scenarios in a way that we can highlight these problems and at the end, try to help or assist them with AI models. And for that, we have some judging criteria for this challenge. So we definitely want to correctly classify how many pedestrians we are able to recognize properly. And it may be in different conditions, different lighting conditions, different weather. Depending on different scenarios, you still want to be able to recognize these pedestrians. We want to understand your flexibility of your solution. So where are you trying to install this? Is it on some sort of a flagpole? Is it some sort of a different angle view? So with that, it also goes on the cost factor, right? So how much are you reducing the cost with your innovative solution, not only in terms of just your hardware silicon area, which is a BOM cost, but also in terms of installation and maintenance, right, so that's also one of the key aspects where we want to do this starts, but also want to understand your cost ratios. And then what's going to be the response on how quickly can you interact? Typically, these have to be real time because you have to take actions immediately. So we want to also understand what's your response time based on your solution. With that, let me introduce BrainChip. BrainChip is focused on AI solutions, and we're trying to put these AI solutions that are today working on the cloud towards the edge. And we do that with our neuromorphic platform. And eventually, we are an IP-based company, which means that we license our technology to other customers who then go and integrate our solutions combined with their own silicon. We are first to commercialize that neuromorphic technology. We are worldwide leaders in development of these AI chip for event-based processing and learning at the edge. And we have 15 years of AI research amongst our cofounders. They are mainly focused on this neuromorphic [ world ]. And then we have Centers of Excellence in U.S., Australia, France and India. So we also have a global presence, we are globally spread out and we are working with many different solutions and partners. So some of them, we are mainly trusted by MegaChips, Renesas. You must have heard about Mercedes and NASA. All these guys have used our technology in some sort of a fashion. And we also partner with external AI solutions to kind of help drive this industry forward. So we've partnered with a lot of different companies. The big -- some of them are ARM, Intel. And we have partnered with SiFive as well because really, we are kind of independent of your whole system. It can work with Intel. It can work with ARM-based processor. It can work with [indiscernible]. And then other AI solutions as well, such as you have [indiscernible] Edge Impulse. We've also worked with solutions partners like emotion3D, NVISO, Teksun, and AI Labs. And some of our key markets, we -- with this solution, which we're really trying to target mainly edge-based devices, and they can be segmented in different areas. So industrial, automotive, health, wellness and home consumer. And based on where you are, because it's IP, we can really scale the performance criteria and constraints that each of these different applications might have. Let me talk to you a little bit more about Akida and the neuromorphic advantage that we see today in the industry. With Akida, we mainly try to do event-based processing on your neural networks. What that means is, we only look at activities within your neural networks and do not do any wasteful computations that's not required. With that, we also have advanced facial temporary capabilities. So that way, we can account for data where time is most valuable. In many scenarios, this can be used for video analytics, predictive analysis. So this is where we can have all these capabilities done inside our computation. And we support event-based processing. We also have an event-based communication. So between different layers that goes inside networks, right, we can make sure that we are only sending an event for newer networks. We're not sending in any wasteful information throughout our technology. We also have at-memory computation. So what that means is, it's very localized for every core within our technology. So that way you reduce your data movement and you are able to reduce power consumption at the edge. And then something this has been long at the edge, which is on shift kind of learning. So we, again, do event-based learning, which supports this behavior of learning at the edge. So you don't have to retrain any model at the cloud again. If you have a new class that you have identified, we have the ability to do all that training at the edge and it can be done in 1 shot or in a couple of shots. So when we look at this Akida technology, it's really meant as a neural processing unit, so it's really self-managed neural processing unit. But when you look at different sensors where you have vision, audio, olfaction, gustation, tactile or any other sensor per se. At the end, there's really a preprocessing state that goes into and you have a neural network that gives you a response out. It does not matter to us which sensor you're using because at the end of the day, we are -- we will be accelerating these neural networks inside Akida. And we will be able to support these -- any kinds of sensors that you're trying to run. Typically for this challenge, you might focus more on vision or any other sensors like LiDAR, radar base or different sensors that can help you solve your solution. And then we can handle complex networks. If they can run completely on Akida, we'll do that completely. If there are some custom layers per se, where Akida does not support some of these layers, we can offload those on the CPU and can get the assist from your host and can still run the supported layers on Akida. So that way, you're still using it as an accelerator or your coprocessor to accelerate your neural networks, maybe not entirely, but at least most of the time. And then we can operate in as a stand-alone when we process the neural network. So we don't really -- when we are doing the computation on Akida, we don't involve any host CPU purely because we want to limit the bandwidth requirements, limit the power consumption and give you the best level of performance. So your -- when you hear about neuromorphic, you typically hear SNNs, spiking neural networks. But I want to show you one classic demo where this is in-cabin monitoring, but here's our partner, NVISO, who's using 5 different variety of models, and they're kind of doing in-cabin monitoring where they're identifying -- let me actually start the video for you. They're kind of identifying face, different body points, your hand gestures. So it's very useful. And if you are familiar with this solution, it uses at least 4 to 5 different AI models, and they're all CNN-based at the end of the day. So what's unique about this was, they were able to use Akida where they were able to put all of their solutions to our development board, which you can see out here. And they're just using a regular CMOS-based camera to kind of influence throughout these different kinds of models. And they were initially using an NVIDIA Jetson device. So if they compare the performance, from an FPS standpoint, we were outperforming on their best model and also for combining all of their average models by a lot, right? And if you see the scale between different devices is also not equal where Akida is running at the least frequency, whereas the GPUs are running at the max frequency and the CPU is even more than that. And this is more on the Arm-based device. So this is kind of a good comparison to show you from a performance standpoint? Yes, we can get you the performance that you need, maybe you need some real-time FPS. We can definitely get you these real-time FPS with our neuromorphic technology. Why is that happening? Some key differentiation between traditional deep learning accelerators and Akida is when we perform our operations, see, for example, these convolutions, we do that in an event-based domain. Compared to the DLA, which typically does matrix multiplication, we're only focusing on event-based processing. Now when we are able to do that, we are able to utilize sparsity, not just wait sparsity of your layers, but also the activation match sparsities completely on the neural mesh. And I have an example to show how that is being done. It comes in the [ ring side ]. But we're able to utilize this sparsity throughout. So that -- if you're getting there, it will be helping you to reduce your overall MAC computations that goes on the technology. So we are able to run a full network without any CPU intervention. So that's also a bonus advantage to when you're talking about power and bandwidth constraints. We have a self-configuration DMA. So what that means is whenever it's kind of seeking network, it can self configure itself. So again, we don't require any host CPU system to kind of help us and assist us and let's configure this type of networks. And it can also -- if you have a large model and if you want to configure it from a chunk-by-chunk basis, again, a self-configuration DMA will completely manage it by itself, without any requirement to assist from the CPU. Again, as I said, it's at-memory computation. So that really optimizes for memory size and power and then on-chip learning. So we don't require any cloud retraining as it can completely learn new classes on the fly at the edge. So here is some examples about how event-based processing works. And this is an example for a convolution operation. So when you look at the top part, which is traditional convolutions, traditional frame-based convolutions, this is how a DLA computes, right? You have an activation map. In this case, it's a 5x5 matrix, then you have a kernel. In this case, it is a 3x3 kernel. Whenever you want to do convolution operations, you take this kernel, you put it on this central location and you get the result. You keep doing it row by row, column by column. And eventually, you get a result in matrix out of it. When we move to Akida, we only focus on events in the activation labs. In this case, there are 3 events, which is dispersions out here. And we take the kernels and move the kernels along these 3 events. So if you see we're giving 3 different computations, right? At the end, we still obtain the same result as a traditional frame-based convolution. The major difference here is in the computation. In this case, there's a 90% reduction. But if you think about an average model, you typically want to use BaaS normalizations, values, all of these focus on centering a lot of these values around 0. So our typical model on an average cases may have about 50% to 60% activation sparsity. This is the activation sparsity that I'm talking about. So we will be reducing 50% compute just by using or converting these models to Akida. Now if you are a step ahead in the game, you can penalize the model to be more fast enough, so say you want to be 80% fast, and that will reduce the computations further down and also reduce your power. Naturally, it also helps with your latency because you're not doing that many computations compared to traditional frame-based convolutions. So this is how we focus on -- we implement our neuromorphic design principles, trying to use today's existing solutions that are seen in base and more advanced in giving you the best levels of accuracy and apply that to our neuromorphic-based design. Now when we compare it, how it's reducing the CPU and memory pressure? Whenever we run neural networks on Akida, if you see the orange line out here, you have data acquisition, you have preprocessing, but this is a portion where you run the neural networks. When it's flowing on to Akida, the CPU is not doing anything on that point in time. So you can really have a low-power state CPU since doing your pre and post base operations, whereas the major computation, which is happening millions and billions of time, you can really offload that onto Akida. If you see the blue line out here, this is an example of any computation that's actually happening on a CPU when you don't have Akida on it. So you're still consuming a lot of CPU power and there's a lot of memory pressure based on CPU. So developers have to write a lot of control logic, make sure the bandwidth is being maintained, so on and so forth. So here is a little example about efficiency. So here is a model, which is Akida [indiscernible], and this is again taken from another partner, Edge Impulse. And this is a lightweight object detection model that does real-time object detection. So it gets localization of your objects that you're aiming for. So if you see out here, we're trying to identify the schedules, which are the circular candies in this frame, and it's able to correctly localize where these schedules are against all different kinds of objects or candies that you see out here. So very useful for like a prescreening algorithm or maybe in the industrial use cases where it's going over [ conveyable ]. In your case, in your solutions, maybe you want to do this prescreening algorithm to detect if there's a pedestrian ever in this frame, right? Just get the location where it would be, then you may be like crop out that local location and then extract some more information with complex algorithms, right? The benefit of this is, if you see this kind of algorithms as well, we're able to run these models in real time and with extreme efficiency. And not just these models, it goes throughout different kinds of -- variety of models. In this case, it's just less than 1 millijoule for this example. But we typically focus on microjoules to millijoules for many of these models, even large-scale models such as mobile nets and [indiscernible] and so on and so forth. So here is some examples of working with other types of sensors. So here, we are working with point cloud-based data, so you can think about this is coming from maybe a LiDAR based frame and here you just have an RGB reference to see what that actual sensor is looking at. So if I play this video, you can see that it's -- we're working off this point cloud, which naturally is familiar. It already introduces a lot of activation sparsity because it's not like an RGB where you have values on every pixel. You just have pixels on things that are really being -- it's being focused on. So this, again, helps with sparsity. Typically, these kind of sensors, when you work and create your neural networks, you will have about 80% to 90% activation sparsity. So that's really huge for us because now you'll be able to reduce power by a lot and also gain a lot on your latency. Now on the right, you have -- this is an event-based sensor data. So if you think about [indiscernible] or [ DVS ] based camera. So you have a street view from them and though we are applying object detection models to kind of identify different -- this is a truck, this is a car, this is a bike. Maybe it's not coming very well in zoom, but event-based data are, again, just like point cloud where it doesn't give you pixels for everything, but it gives you event intensities and different polarities for things that are shifting from time to time. So it's a little different from neural network perspective, but they focus on what's difference between timestamp 1 and timestamp 2 and so on and so forth. Now let's talk about -- we talked about -- introduced Akida technology to you guys. So how can you use Akida today? So we have MetaTF, which is our Akida machine learning framework. And this is publicly available. You can even visit that today at doc.brainchipinc.com to kind of go through our documentations. We have a lot of examples to help you and guide you how can you convert your CNN models on to -- and put on to our Akida platform. If you're an advanced user, you can even look at advanced tutorials. We have tutorials on edge learning as well. And so we have some installation guidelines, user guidelines and some constraints that maybe go with our hardware. So you can take a look at those things. And on this MetaTF, we have 3 different Python packages. One is the Akida, which has our run-time engine, so it can do our model inferencing. But the nice part about this is, it also has a software backing. Along with its hardware, it can also do a software backing. So let's say, you don't have the kit right now and then you are -- you're just doing your development. It's an iterative development that you're doing. Without even deploying the model on the development hardware, you can actually use the software simulator to convert the model to Akida to run an inference on the software and check for accuracy or losses or how much hardware resources it would require. So you can use all these tricks before even you deploy the model on the hardware. And then once you deploy the model in the hardware, you can use the hardware backing, and it will be with hardware-level performance such as power and latency. The other Python package is Akida models, and there's some various APIs. So 1 of them is model zoo, which has a lot of different varieties of models for different solutions. So you can use this for transfer learning. If you are developing a solution, you like a particular model that's there, uses for transfer learning. So it helps you with faster training, faster convergence and really an optimized solution that runs on our hardware. There are other APIs in this such as knowledge distillation, pruning of models. So you can -- if you are really focused on the later criteria of the challenge where you want to focus on your BOM cost or you're going to focus on your real-time latency, you can use all these APIs as well. And then we have another package CNN2SNN, which is really like a conversion tool to convert your TensorFlow Akida's models, which are in CNN today and convert them to a spiking based or Akida-based format. So that it can run on our hardware. It also supports tools like quantization, so you can work with 4-bit quantization, 2 bits, 1 bit for all activations and weights. So you can help you with all these techniques. It also has extended support for quantization of AI training. All of this is extended from TensorFlow library. So whatever TensorFlow supports today, it's supports to 8 bits. We have gone around and extended to support 4 bits, 2 bits and 1 bit as well. So it's really a helpful tool to make sure that you can do the right development for your application. And about development kit. So again, we are an IP-based company, but we do have silicon, which we kind of use it as a reference based design. It's called AKD1000 and we use this for prototyping, demonstrating solutions, working for exactly these kind of problems where you want to demonstrate something is working in neuromorphic [ world ], right? So it's a silicon, which has a lot of different varieties. You have an M-class CPU on it. We have an Akida mesh, which is our -- the neural network accelerator. You have other I/O ports on it for getting inputs from your host. And so with this chip reference design, we kind of have multiple different development environments, 1 of which is -- so we have this PCIe Board with our chip on it. So really, this PCIe Board can get input from your host. And right now, we have a development kit, which attaches with a Raspberry Pi OS -- Raspberry Pi development kit. It uses an ARM-based processor. And for you guys, when you say, submit your proposals, out of all the proposals, the best 3 ones, we will be providing them with our entire development kits, which is Raspberry Pi with this Board. And what this will give you is a setup with an entire Akida platform. We have some sample demos, so you can look at it how we do things. You can play with those demos as well. And your environment will be set up. So you'll have your drivers and everything will be set up on your system. You can just have your solution and then you're ready to deploy it on this development kit. And then, for the remaining proposals who do not get this development kit, we will be shipping just this PCIe card with our chip on it. Again, this is a AKD1000-mini PCIe development board. And so you can use this on any Linux-based system that has a PCIe interface support. We will give you guidelines on how to install their drivers. So you'll be able to install these drivers and you can play with those things. You can even use the same Raspberry Pi kit or other kits, but you will have to do all the setups on your -- basically on your own. Now when you think about these kits and how do we view our AI solutions, right? So from how do we develop -- you a concept and how do we develop? So first, you can evaluate your models using MetaTF. You can use our partner, which is Edge Impulse. We have integrated AI solutions on to Edge Impulse platform as well, including the development kits. So you can evaluate how your solution looks throughout. You can then go ahead and design your solution based on the development kit you're targeting for. And then you can develop it specifically for our AKD1000 chip. Again, you can develop and deploy it using MetaTF or you can use Edge Impulse as a development platform as well. And then you can -- if you are focusing on this IP, this is where you can scale your solution. So this is -- since this is a reference design, you can use this for understanding what your scale of the representation could be, right? You may have a model that does not use all of its course. This is a portion of it course. So we have a way to kind of shut down some of these portions and just use those portions so that way you can scale up and scale down based on your models and your application environment. And we can work with different kinds of sensors. So I already showed you camera-based sensors, showed you point cloud-based, event-based. But there are other sensors. You have radar. You have time of light. You have IR sensors as well. So all of this is possible to work with us. And then no matter the solution -- all these solutions will work in event-based process, right? Some might work better than the others. This is based on sparsity levels, based on your power and constraints as well, but some might require more accuracy. So you might want to even combine multiple sensors to 1 sample image and do those kind of inferences and then we have external resources. So we have a YouTube page, if you search for BrainChip Inc. We have a YouTube page which has a lot of workshops, a lot of different webinars that we have done in the past as well. We have demos that can help you understand, oh I can also do this kind of things with Akida. If you visit our website, we have a variety of different blogs, so we have blogs with -- some with benchmarking performances, some with explaining about neuromorphic a little bit more in detail by 1 of our cofounders. We have a variety of different blogs that we kind of focus on different topics that might interest you. And check out Edge Impulse as well. So we have integrated our Akida platform onto Edge Impulse. So that way you can build a solution, which is no-code ML design studio and can really deploy it onto one of these development kits. You can -- even before deploying, you can just get the model out of the studio as well. So you can use that platform additionally. And in terms of schedule, so today, we did the development kit webinar. The link for sending your proposals goes live on the tinyML website. So if you -- we'll probably post the link in the chat as well, but if you go to the tinyML website at tinyml.org/event/tinyml-hackathon-2023-pedestrian-detection, you'll see the link live, where it will be like a Google Form where you can just submit your proposal and we'll go through all your proposals. The last day of submission is 2nd June. So we'll go through all the proposals after that and we select the ones that we -- for the development kits with Raspberry Pi and PCIe Board and just the PCIe Boards as well. And we'll be able to ship them by 8th of June. So that way, we will contact individual teams and make sure we get all the right details and we'll be able to ship all of them in first week of June. After the shipments, we'll have training sessions. We'll have a 2-hour session. A little bit -- this will be a detailed session just explaining about MetaTF, our engine library, how to play with certain things. We have an expert -- AI expert on the call. We'll go over these training sessions, specifically the teams who we have given the development kits. So that way, you guys, if you have any questions, you can get it done with the AI expert right away and we'll make sure that you guys have the right level of support that's needed. And these are the optional deadlines from tinyML based on the challenge. And then again, you have other -- after deadline is passed, you have shortlisted winners and presentations going around in October timeframe. For any other external support, after the training session, right from today onwards, this e-mail will be active, which is [email protected]. So this will -- if you need any help with our development kits or any questions that you have for us, please email us at this e-mail ID. And one of our AI experts will get right away and help you resolve your questions and help you understand some concepts as well. It will be our 24/7 support, but based on different time zones and where the AI experts are, kind of expect a response in 24 to 48 hours from us. With that, thank you so much for attending. If you need more help, you can visit our website brainchip.com. Our MetaTF platform is doc.brainchipinc.com. Please submit your proposals if you're interested in getting one of our development kits for solving this challenge. And then if you need any support, we have a support email that you can always e-mail to and get back to us. Ira, over to you.

Great, Nikunj. Thank you. We've been answering some of the questions via written answers, but let's talk about some of the more detailed ones. So one of the people ask, if you can also couple the Akida with the Jetson TX1 as one can benefit from a hybrid system, Akida on PCIe and the GPU. Could you maybe talk about that a little bit?

Yes, that's a great question. Yes, so you can couple that in a way such that -- because it's a PCIe Board, it still requires a host to kind of give you where your models are stored and you kind of give your inputs. So we can -- if you -- if the Jetson board has the PCIe slot, which I believe the one that you described has it, you can hook up the PCIe Board. Our drivers are Linux-based supported. So I think the Jetson has a Linux-based platform. So we can install those drivers, and we can help you support if there are any problems with installation. In terms of the model inference, a lot of the networks do run entirely on Akida. But in case if you write any custom layers and you require GPU help or if you require GPU help for your preprocessing or postprocessing, we can couple them and you can really tie up a nice application where a portion of it is running or accelerating on GPU for your pre and post and then your neural network is being accelerated on Akida. So that way you can use GPU only when it's needed and you can reduce your power consumption and Akida when you're accelerating your neural networks and reduce your overall system power consumption from an application standpoint. That's a good question.

Yes. Great. Okay. So that answered that one. Okay. So you talked about people submitting proposals for getting free dev kits. One of the people in the audience asked whether they could just purchase a dev kit. Is there a link or some place or should they just use the support e-mail to get a link if they just want to buy one or...

Yes, that's a good question. So yes, we do have development kits for -- available for purchase. If you visit our website, brainchip.com, there'll be a section where you can look at our enablement platforms and a portfolio for purchasing development kits. Additionally, from Raspberry Pi, we also have an Intel-based development kit, which is Intel x86-based OS. So that's useful for any powerful applications that might be required, powerful CPUs, so you can purchase them. If you have any difficulties with that, e-mail us and we can give you the link for the purchase as well.

Okay. Perfect. And then there is a question if you want to comment on -- Sony talked about their vision-based system and Infineon talked about the radar system. The question, how well could you support those on the platform and what level of integration is there? I mean we haven't done a close integration, but maybe you can talk about generalities there?

Yes. So definitely, we have not done a close integration with the other partners for this challenge. But being Akida, we support Linux-based OS, right? So whatever post these sensors can work with Infineon ones and Sony ones, we can plug in our board for the neural network activation portion. So you can really take in those inputs from those sensors on the provided host platforms and then run the neural network inferencing on Akida. So that way, yes, you can combine -- in fact, you can combine both sensors from Infineon and Sony and create a nice fusion sample before you submit it to your neural network for acceleration. So that may even give you a better level of accuracy if that's what you're looking for. But yes, you can combine them and use it with Akida. We are not -- we don't restrict inputs coming from a variety of different samples because at the end of the day, wherever the raw image or sample comes from, there's going to be some preprocessing, maybe a downsizing, maybe normalization, maybe some other preprocessing that might -- that you might use for your application. Good question.

Okay. This one may be a little off topic, but in the samples and in the doc, they asked whether there is an example to do a speech recognition using the Akida platform? Is there -- probably not directly relevant to the content, but is there a place people should go to see examples and demos?

So -- good question. So on the MetaTF platform, we do have a lot of examples. We have examples on the vision that -- object detection that will help you for this challenge. But we have other examples as well, such as speech recognition. We talked about audio classification. We have examples for vibration recognition. And we -- in our model zoo, we also have models for all of these different tasks as well. So for this challenge, you probably will be looking at vision and object detection-based models. So we have a zoo for that. And then if you're looking for, outside this challenge, other use cases, we have a model zoo for that as well. So that's a right appropriate website to look for examples and what you can do with different kinds of solutions.

Right. Okay. And Regina just posted the corrective link forms that GLE, so Google Form or dev kits or you can access it through the tinyML website. I think that's all the questions. Obviously, people can ask directly or through the form. So thank you, Nikunj. Let me take the screen back. Okay. And so we will be posting this information on the contest platform. We will also be posting the video on our YouTube channel in the next day or so, youtube.com/tinyml. So this will be archived there. In general, if you have questions about the Hackathon or you want to discuss about other participants, please use the tinyML form, just go to forms.tinyml.org. This is the specific topic, Fred, but you can just find that from the homepage. If you have general inquiries about tinyML or organizationally in terms of the challenge, please reach out to Regina. And once again, I would like to thank all the tinyML strategic partners who make this and all our other activities possible. In particular, I'd like to thank the executive strategic partners, including Edge Impulse, the leading development platform from Edge ML, Qualcomm AI Research, Advancing AI Research to make efficient AI ubiquities and Syntiant accelerate your edge compute making edge AI a reality. Our Platinum Strategic Partners include Renesas, Sony, AITRIOS and our Gold Strategic Partners are Analog Devices, Arduino, Arm, Infineon, [indiscernible], Microsoft, SensiML, STMicroelectronics and Synaptics. And then I'd also like to thank our Silver Strategic Partners, Aizip, BrainChip, Greenwaves, [indiscernible], IBM, Imagimob, Nota AI, NXP, Polyn, Qeexo, Schneider Electric, and Silicon Labs. Most importantly, I'd like to thank you all for joining us today and look forward to having you participate in the future tinyML program or event. And I wish everybody a good remainder of your day, and thank you again.