Neurala the company behind The Neurala Brain a deep learning neural network software that makes smart products like cameras, robots, drones, toys and self-driving cars more autonomous and useful.
Unlike traditional AI companies, which designed for super computers connected to the Internet, Neurala’s first project was for NASA to be used for autonomous planetary exploration. Super computers were not available. Battery life was limited. Fast internet access was impossible. Their deep learning neural networks had to be lightweight and perform in real-time without ground intervention. With these constraints, Neurala modeled the Neurala Brain on animal brains because animal brains are highly efficient “computers” that do more in less space and with less power consumption than the computers that are in use today. The brain also knows how to use eyes (cameras) and ears (microphones).
This approach worked and Neurala are now bringing The Neurala Brain to market. Neurala’s smart, fast, anywhere brain works on systems from single board computers to large servers.
We spoke to Massimiliano Versace, CEO of Neurala and here is what he had to say:
Neurala is a New England company and so we have strong ties with England. We are Boston based, and started back in 2006, as a container for intellectual property and patent, that Anatoly and myself, because there are Russian and Italian and American came up with, back when we were in the doing our phd’s in Boston university. So back then we were working on machine learning, neural networks, emulation of brain functioning software and we came up with an idea on how to accelerate this algorithm, so it can run briefly at scales and the speed that is useful for real time deployment in robotics, and so we came up with a patent to use GPUs and of course today GPUs are the main fuel in the hardware for the robotic community, and we decided to incorporate, and at that the time we were still in school and we kept Neurala as a sort of a fun side project, until it became a bigger side project. We started to work with Nasa and the Air Force and some private customers starting in 2013, when we left the university full time and we started to work in Neurala, to grow the company as a business. So since 2013, we have as a company, raised about sixty million dollars in venture funding work with a variety of companies and today we are as you know signing for the first time, large deals for large scale deployment of our technology to different devices from small toy robot all the way to industrial and automotive, passing through drones and consumer devices.
Can you take a moment to explain the applications of A I.?
In general the application of AI with a mathematical symbol of arrow to infinity. AI will be in every single software so that half of AI, and that’s why it’s so exciting for many people today it is that every software can be made smarter with the use of AI, so there are infinite. In particular in those applications that will benefit from human like perception or decision making applied to large volumes of data, so let’s make some examples. Driving a car or flying a drone or driving a train or a flying an airplane or looking at security videos or looking at emails or looking at website or judging beauty of a picture or a sound of a song so all the human like activity, to the limit will be subject to various degrees of automisation thanks to the fact that with a certain kind of AI, not all AI is made equal but with certain kinds of AI today we can begin to approximate skill set of humans or animals in software. So perceiving which was the sole province of human ability today is also shared with machines and perception of auditory signal, visual signal or more complex signal like financial data or a multidimensional data is to they are tackle-able with AI.
Can you just briefly explain the difference between strong strong AI and weak AI?.
I used to call it traditional AI versus brain based of AI, So brain based AI was for many years the joke of artificial intelligence. You should understand AI has become only recently a synonym with neural networks in the past it was the method by a completely different paradigm, which was the traditional artificial intelligence that has much to do with “if“ “there” and statements instead of neurons. In the past when we joined the community I started programming neural networks very early, twenty years ago. We were not taken seriously. Our discipline was almost ridiculed in the sense that it is only an exercise and these are toy models and they will never work. We can build in them because we had the faith that, If you approximate human cognition by building it from the ground up, you’re going to be much better than any other algorithm that is built out of high level assumptions, and they usually make an example of an apple. The strong AI and weak AI have a very different ways to code an apple, in that weak AI, they will say if an object is red and has a roundish shape and has a little thing sticking from the top ,and you can eat it, then perhaps it’s an apple with 95.3 percent probability. The brain Based AI or the strong AI learns that by example so the underlying mathematics of neural networks enables you to build a conceptual model of an Apple through exposure to the data. So you’re building from the bottom up rather than somebody imposing those constraints from the top down and so today the strong AI is dominating. We saw that coming which is why we have a bunch of patents around it, but that the world is a awakening and it’s obviously changed radically the way robotics is done today.
What are the difficulties of introducing AI into everyday objects?
The difficulty have changed over the years. In the ten years ago, when we started working for instance and talking to Irobot about working with Nasa and the Air Force. They were treating us as an experiment. So our main difficulty was to convince them to invest in the technology. So that is gone now. Who invests in other technologies stupid, whereas we were stupid. That difficulty has changed and the difficulty today is that we are facing in deployment in robotics is the issue of trust and transparency so let me explain what I mean. Let’s put the robotics application into two dimensions for this discussion, non-critical application and fun application and very critical application and not so fun applications, so one example is a toy sold by a major toy company that if things go well will have Neurala brain in it, so in this case the worse that the toy can do is to misclassify you for your sister. Big deal, maybe you get tired of the toy and throw it away, and the other use case is the high altitude drone that is misclassifying a friend for an enemy and so it’s looking at video stream and saying that this is an enemy, we should do something about it. Hopefully there is a human in the loop but take it a notch down to a self driving car which is taking a turn and is misclassifying a kid for leaves. So that the issue is that the in particular for those tough use cases where there is either life or money at stake. People want to be sure that the AI is impeccable. So the challenges are twofold. First you have to convince people that alternative solution is relying on people and people are not perfect right, so the first thing is to educate the customer, that what they’re looking at is not one hundred percent but ninety two percent, or whatever it is that a human can produce in terms of accuracy, and we just had a meeting where we talked about airports you know scanning luggage has the misclassification rate of seventy percent. It’s staggering, I can go through with a machete and I have a fifty fifty chance that it goes through. It’s scary but he also mentioned the standard, the expectation on AI are unrealistic. OK. So perfection is not of this world so that’s the first challenge and the second one is to explain AI, so with a strong AI, it was fairly easy to understand how the machine took the decision, despite the decision was most of the time wrong, you could understand why. People are comfortable and say “oh ok, if I change this threshold of redness it’s going to get it right”, so that gives me some sense of understanding of AI despite that it sucks. Now with machine learning with a neural networks, or the one the kind Neurala build, it’s much more similar to how a human takes a decision which is completely based on instinct. We take decisions all the time and then we rationalise them with words where we can, and so AI cannot speak yet, but they will, but so that we aren’t at that sweet spot where the AI is not completely explainable and there are techniques being developed but we are today in that valley.
Can you tell us a little bit more about your work from NASA?
Nasa engaged us that back in 2011.They wanted to change the paradigm on how large missions on Mars might be taking place in the near future, the usual paradigm is if you send a giant rover and it’s controlled by Earth step by step. What if you can send a swarm of independent robots where each robot might have a brain the size of a rat brain where they operate completely independently from the earth so they’re not bound to the thirty five minutes of the back and forth of calling home. So the goal was to develop a baseline technology that enables an operation of a brain like nature completely locally in a very low cost robot. So the robot will have a single camera a small processing power like a small GPU, no active sensors, very lean, package and we have to do all that the rover can do in terms of exploration like imaging like a rat or a small mammal exploring an area for food. So we have to reproduce that ability which is to navigate around collision free, perceive, locate, locate itself and go back to the base. and all in that single processing power. So that was our moment of creation.
How have you managed to model your AI on animals and what are the key differences between animals and A I?
So our training has been, when I say “our”, I mean the 3 co-founders, we were trained on mathematical modelling of brain functions, so Anatoly who is our Russian guy, did his phd on parallel processing and rat neurophysiology, so he built this model rat brain for his phd thesis which was using the brain like algorithms to navigate around in a maze and make decisions on where to turn, find food and then go back to where he found the food and locate itself. I built cortical models of visual perception. In a software, I emulated the way that the cerebral cortex learns “on the fly”, new information through exposure to stimuli and had to be the similar thesis but focused on speech perception and speech production so we spent many many years modelling the brain with neural networks, and basically got to the stage in which we believe that this is the right paradigm for machines. If you want to be robust in a rapidly changing world, you need to be able to learn “on the fly” about the information post as opposed to traditional machine learning, where you learn and you code and so it turns out that the paradigm is paying off slowly, the machine learning community is realising that the way that AI was devised, which is training the factor and then deploy is not really scalable, and so we are capitalising on this early lead in the technology maturation.
Is there a favorite project that you have worked on, what is it and why?
Our favorite project is ongoing and we are working with a top consumer electronics manufacturer to embed our AI into millions of devices. I cannot tell you the details but it’s exciting because if the time in which Neurala must master a transformative step in it’s career which is going from technology development to large scale technology fielding and execution (deployment). So eventually we are going to reach the next step of the company, each step has been important but the next one is having a customer which has fielded tens of thousands of millions, if possible, of our little brains into many devices that end up in consumer hands. so there will be the next step.
As an established player what advice would you give to new market entrants?
Don’t be a copycat. Neurala was ten years ahead of its time and I think that that’s why we are the leader. Today we see scores of AI companies, the self-proclaimed AI companies, but all they do is they download a package from tens or cafe, they get some and they apply to a special kind of data and they call themselves the AI are for X. These are short lived gigs I think that to be worthwhile as an AI company you need to actually have AI technology and you need to be a producer and at the forefront of the tech rather than a user of the Tech. There are very few of those companies so as you look at them, you know if you measure that there are a million flies in front of you as you apply this filter ninety nine percent of those fly will fall to the ground, and there will be a few still standing in the long run and I hope that Neurala will be one of those.
I will say that for many many years robotics and the AI have been running on parallel tracks. I always believe that the two have to be merged in order to work, as often times people build a robotic company just thinking about the body and they think of the software as an afterthought, and that, I think, is a capital sin of the robotic industry, and I think the successful robotics companies have to take very good care of the software and the AI in parralel if not before the hardware is built. The technologies should converge, that you should take care of your AI as much as you take care of your servers. S