Is neuromorphic computing the only way we can actually achieve general artificial intelligence?
Very likely yes, according to Gordon Wilson, CEO of Rain Neuromorphics, who is trying to recreate the human brain in hardware and “give machines all of the capabilities that we recognize in ourselves.” In other words, an artificial brain that thinks like we do … or general AI.
Rain Neuromorphics has built a neuromorphic chip that is analog. In other words it does not simulate neural networks: it is a neural network in analog, not digital. It’s a physical collection of neurons and synapses, as opposed to an abstraction of neurons and synapses. That means no ones and zeroes of traditional computing but voltages and currents that represent the mathematical operations you want to perform. This is similar to what Intel has done in the Loihi chip, but potentially far more advanced.
Support TechFirst: Become a $SMRT stakeholder
Right now it’s 1000X more energy efficient than existing neural networks, Wilson says, because it doesn’t have to spend all those computing cycles simulating the brain. The circuit is the neural network, which leads to some extraordinary gains in both speed improvement and power reduction, according to Wilson.
All of which opens up shocking possibilities for better artificial intelligence, much faster machine learning, and far smarter autonomous machines.
Keep scrolling for the video, the podcast, and a full transcript … or check out the story on Forbes.
(Subscribe to my YouTube channel)
TechFirst podcast: building an artificial brain with a neuromorphic chip
Transcript: Rain Neuromorphics is building smarter chips for better AI
(This transcript has been lightly edited for length and clarity.)
John Koetsier: Can we build a computer chip that operates like the brain? We’ve seen neuromorphic computer chips from companies like Intel before, like the Loihi chip which Intel claimed had a sense of smell. Rain Neuromorphics, however, says it has an end-to-end analog chip, a neuromorphic processing unit, that they say is the world’s first end-to-end analog, trainable AI circuit. It’s a fully analog neural network and it’s a thousand times more energy efficient than today’s best processors.
Here to chat is Rain Neuromorphics CEO, Gordon Wilson. Welcome, Gordon!
Gordon Wilson: Thanks so much for having me, John. It’s good to be here.
John Koetsier: It’s good to have you. When I see ‘neuromorphics’ in the inbox, I gotta check it out, it’s a cool space. Let’s start here: what the heck is an analog chip?
Gordon Wilson: Yeah, great question and I think a great kicking off point, because it really allows us to frame kind of what we’re doing compared to what has been done for this past decade. And I think it’s easiest to understand what an analog chip is, you know, in contrast to the neural networks, the AI, that have defined this last decade of deep learning.
So, in 2012, we kind of had a big event that started this new era, this AI renaissance, and we are seeing these massive neural networks grow in size and grow in capabilities since then.
But all of those neural networks that we’ve seen in this deep learning world are neural simulations. They are abstractions that are written in software and the neurons and synapses that are defined in that software don’t physically exist, but rather they sit on this highest layer, usually written in Python, and they’re then translated through many layers of abstraction down until it gets to the digital circuit — most of the time, a GPU graphics processing unit — where it then processes that math that represents that neural network at the top.
An analog chip and a neuromorphic chip is different.
It’s not a neural simulation, but rather it’s a neural circuit. It’s a physical collection of neurons and synapses, as opposed to an abstraction of neurons and synapses. And this is very similar to the brain. The brain is a collection of physical neurons and synapses. It’s governed by the laws of physics and it achieves all that it does with extraordinary scale and extraordinary efficiency within the bounds of kind of the physical world.
So, an analog chip, as we’re building it, is trying to achieve something similar … trying to find ways to learn, find ways to scale, all within the physical domain.
John Koetsier: So, let’s dig into that just a little deeper. Great, great overview. What does that mean — typically a chip will think in terms of, or operate in terms of on or off, one or zero, right? Binary logic, right?
Gordon Wilson: Yes.
John Koetsier: What does an analog chip — how’s that work? What’s that look like?
Gordon Wilson: Yeah. So digital chips are, as you said, built on the very bottom on zeros and ones, on this Boolean logic of on or off, and all of the other logic is then constructed on top of that.
When you zoom down to the bottom of an analog chip, you don’t have zeros or ones, you have gradients of information. You have voltages and currents and resistances. You have physical quantities you are measuring, that represent the mathematical operations you’re performing, and you’re exploiting the relationship between those physical quantities to then perform these very complex neural operations.
So an example of this, is like matrix-vector multiplication. This is the kind of backbone of most neural network math. And GPUs do this by parallelizing these multiplications across a lot of digital cores and doing precise, digital multiplication and addition. In an analog chip, as we are building it, we’re not doing this with highly precise, digital math, but instead we have the activations of the neurons represented by voltages. We have the weights of the synapses represented by resistances, which are held in components called memristors. And when that voltage passes across that resistance, you have a natural relationship between voltage and resistance that’s multiplicative. To receive a current, you read out a current and that’s your output.
So an analog chip works by kind of first understanding these physical relationships between electrical quantities and exploiting those to do the math — to make the physics do the math for us.
John Koetsier: That sounds super interesting and it sounds, at once very, very complex, and at the same time kind of simple, right? And is that one of the key reasons why your chip is so much more energy efficient? I mean, you’re claiming a thousand times more energy efficient.
Gordon Wilson: Yeah, exactly. That really is, I think, the most fundamental reason. And I think when you consider the comparison again to the neural simulation, the neural network and that simulation exists so many layers above, and requires to be translated and then performed on a circuit that is really not that well optimized for neural math.
But in our case, the circuit is the neural network. And because it exists in that very bottom layer, it is on the substrate of the chip itself, you can achieve some extraordinary gains in both speed improvement and power reduction, which of course gives you that energy efficiency gain.
So, so yeah, it’s because we’re building the physical thing on that bottom layer.
John Koetsier: That — I’m fascinated. And I’m wondering, with Boolean logic, old-school computer chips, you kind of have to simulate reality and go through those multiple layers of translation because you’re not on bare metal, you know, as the old-timers in computing would tell us, right? Gotta go through various layers of translation before you’re actually hitting computer machine instruction, machine language, but you’re kind of modeling ‘reality’ or what you’re computing in reality.
Is that an accurate way of thinking about it?
Gordon Wilson: Yes. That’s an exact way, a really great way to put it. You know, I sometimes use the metaphor, like what would be easier to assess your ability to kick a soccer ball on a field, right? Would you rather reconstruct all of the physics of your body and of the soccer field and of the ball and simulate that kick and observe that in the simulated world?
Or would you just rather walk onto the pitch and kick the ball? For us, it feels very obvious that the most natural way to build a neural network, to make it as efficient as the brain, is to build it physically, just as the brain does.
John Koetsier: Are you trying to build an artificial brain?
Gordon Wilson: Yes, we are. That is our goal. We have kind of two missions that are very complimentary. One of them is to build a brain and the other one is to actually understand it. We really love the, and believe the notion that you can’t fully understand a system until you have reconstructed it and built it, you know, this goes back to Feynman and many others.
But what we believe we’ve developed really are kind of the core technologies that allow us to first build just kind of unit level chips that address near-term problems.
But ultimately, we see these as kind of like Lego pieces that due to their low-power footprint, we’ll be able to concatenate them together using things like chiplet integration, advanced packaging, and ultimately scale out these systems to be brain scale — 86 billion neurons, 500 trillion synapses — and low-power enough that they can exist in autonomous devices.
Because today, training is so expensive, first, we consider training and inferences as these separate problems and I don’t think they really should be considered as these separate and distinct problems. But it’s so expensive that we can’t even conceive of putting training of natural language in an autonomous machine. And yet, humans do it all the time.
So, that is what we are trying to achieve to recapitulate the brain in hardware and ultimately give machines all of the capabilities that we recognize in ourselves.
John Koetsier: So, Gordon, you’re a pretty soft-spoken guy and you sound like a very thoughtful guy, and in that very soft-spoken way you’re saying absolutely … ginormous things [laughing]. I mean, like we’re talking Frankenstein level stuff, right? I mean, you understand the gravity of what you’re talking about, right?
Gordon Wilson: Absolutely. No, this is … the scope and impact of our work is not lost on me. I mean, this is something we’ve been working on for nearly five years, and we recognize that if we are successful in achieving this, this will be historic and massively consequential. But, what motivates us is of course the impact on positively improving human life.
You know, when we can have personalized medicine, personalized education, we can automate all of labor, I mean, this is the world that we want to realize.
I think many people are already in this consensus that artificial intelligence is going to be kind of the defining technology of this century, but most people don’t know that the hardware that supports it is the bottleneck right now.
So, yes, we recognize the scope of this and it’s a big mission and a big task. But, I think that we’re approaching it the right way and we’re also very conscientious of the fact that we don’t want — that there are good ways of implementing AI and there are also not good ways of implementing AI. And we don’t think AI should be used everywhere for every purpose, but there are guidelines and ethics that should really direct how we build and implement these systems.
John Koetsier: And it’s controversial. It’s also political in the AI space, right, in terms of what you’re building, what you’re looking at. People who claim to be working on general AI, for instance, you know there’s a lot of scrutiny.
There was a, I believe a Google engineer who last week said, ‘I think that some of our systems are approaching consciousness,’ and there were a ton of people jumping all over him — probably correctly, but you know, perhaps in a bit of a mob mentality — for daring to suggest that and saying, ‘No, we’re so far out there.’ Maybe let’s come back here, because you and your co-founder have studied the human brain significantly.
What do we learn about AI from the way that our brains work?
Gordon Wilson: That’s a great question, John, and I’d say that there are really kind of two categories of clues that we look at from the brain that then inform our hardware. The first is, how does the brain learn so efficiently?
You know, the brain trains and learns with both very few examples. We learn with one example or two examples, one-shot learning, two-shot learning, and we can generalize extraordinarily well. So learning/ training happens very, very efficiently.
And there are — the learning rule of the brain, the algorithm the brain uses is not fully understood or identified, but we do know that there are certain requirements that that algorithm must have. So one of those is called a local learning rule.
So this, for the listeners who are familiar with backpropagation which is the industry standard algorithm for digital AI, this has what’s called a global learning rule. So, first let’s say, well, what is a learning rule? A learning rule is how any given synapse in a neural network knows, should it become stronger or weaker for the whole system to become smarter — better at its assigned task.
And in backpropagation in the digital world, for you to calculate whether this one synapse should be stronger or weaker, you need to do a math equation that incorporates the entire network. You need to differentiate against the entire system. That’s really mathematically expensive.
But not only is it expensive, it’s also impossible for the brain to be doing the same thing. There’s no agent sitting in our brain observing the whole system and then doing a math problem to update every synapse. It’s impossible. The brain has to be operating with a local learning rule. And so, a local learning rule is so that that synapse can just observe what’s right nearby and still know, I become stronger or weaker. So that’s one of the clues in the algorithm side. And that’s one of the pieces that makes our learning algorithm so special that it is as smart as backpropagation, but it does so with a local learning rule only.
John Koetsier: Oh, you burst my bubble there, there’s no homunculus — I believe that’s how you pronounce it — I mean, there’s no little guy in a little control room in my brain [laughing]. Unfortunate.
Gordon Wilson: No, no. Yeah. And there’s another side of the clue. So I’d say the first is on like how it learns, the learning algorithm. The other is about scale.
How does the brain achieve its massive scale — 86 billion neurons, 500 trillion synapses — and still process and move information so quickly, so efficiently?
So, and in that case, we take clues from like the topology of the brain and the specific thing is called sparsity. The idea that you don’t have to connect every neuron to every other neuron in the system for it to be well connected.
Now, again, to contrast to the status quo, digital AI and deep learning is — was primarily built on what are called fully connected layers in neural networks where you have a layer of neurons and another layer of neurons, and you connect every neuron here to every neuron in the next layer.
And this was…
John Koetsier: Whoa!
Gordon Wilson: …that was the natural way to do it on GPUs, because GPUs do these dense matrix multiplications which corresponds to this fully connected, densely connected system.
Now our brain … is very different. Our brain, of all the possible connections that could exist between neurons, it’s something like just a fraction of 1% of those connections exist. And yet at the same time, the path for information to travel from one neuron to any other neuron is very short. The average path length is about four jumps for information to traverse anywhere in the brain. It’s extremely well connected.
So there are these special patterns of connectivity. One of them is called a small-world network. And if you’ve ever heard of a small-world network, it’s a network pattern that also mirrors human social networks that gives rise to the six degrees of separation property and human connections. And the idea is you can have lots of local connections. You want to be connected to your neighbors — you’re likely to be connected to your neighbors and that’s a very short path to bridge — and then you want some long distance connections.
And you can create these very well-connected networks at very large scale when you implement these intelligent forms of sparsity.
So, sparsity is core to our scaling architecture and really, and just to kind of summarize there, so we have, these are the two core technologies that we’ve developed: a learning algorithm that learns at the local learning rule, and a scaling architecture that scales to massive sizes of neural networks using intelligence sparsity.
John Koetsier: I knew Kevin Bacon would come up some point, I mean, [laughing] inevitable. Last year, I believe you sort of taped a functional chip together, your first prototype. Where are you on the journey to shipping a fully functional chip?
Gordon Wilson: So we have, I’d say, in the last four years, we have done such expansive work that has been mostly, I would say, qualified as a research. We’ve been exploring different algorithms. We’ve been exploring different architectures. Originally, we had this concept using random nanowire meshes — turns out it’s not very manufacturable and better to build things that you can manufacture today easily.
But in this last year, we kind of crystallized kind of our two core technologies, like what is the learning algorithm and what is that scaling architecture, and then developed hardware prototypes of each. We still have a good amount of time to engineer this completely and to get to market.
We hope to get to market on the order of — full at-scale shipping — 2025. But that said, you know, building the world’s first analog neural network is not easy, and it takes time to iterate through this and get it fully up at scale.
John Koetsier: And you just got some help there, you raised some funds.
Gordon Wilson: Yeah, yeah. So we just closed a $25 million Series A, which is thrilling. We had used $8 million for the last four years and honestly had gone through some challenging times where we got really close. One of our first tape-out prototypes didn’t function precisely as designed — it was a very silly error — but in chips, it takes a long time to iterate and resolve. You can’t just debug like software. So we went through a lot, we learned a lot, and now we have $25 million in the bank and we’re hiring like crazy.
John Koetsier: Wonderful. It goes quickly, I know from personal experience [laughing]. So, talk about what this will enable … you’re shipping this chip, what will it enable?
Gordon Wilson: Yeah. So, we want to not enable just like an incremental improvement in AI. I think that there are a lot of folks — Nvidia included — on the digital hardware roadmap, and because digital hardware is so mature, it’s just, it’s kind of incremental gains that we’re getting at this point. And I think there is still more improvement to be eked out on that roadmap, but we’re trying to enable a new roadmap that is really a step change in performance improvement.
And so we want to enter the market really at 1000X energy efficiency improvement over status quo hardware. And that 1000X comes from about a 10X reduction in power alongside a 100X improvement in speed.
And when you can do that — and importantly not just for inference, we’re talking about training and inference in the same platform — it unlocks possibilities that have just been inconceivable until now. For one, you know, currently people consider training and inference as these kind of separate problems that we need separate platforms of hardware with. We train up in a GPU cloud system and then we might upload those weights onto a more efficient chip and deploy it out into the world. I think that will be the first step for kind of low-power inference and devices, but we don’t want devices just to be pre-programmed and just do what they do in the world.
We want devices to learn on their own. We want devices to have an adaptive brain that’s continuously learning from a changing environment and from a changing self.
So imagine a robot, right? We have, we’re eventually in our lifetime, we’re going to have robots for everything, you know, but maybe it’s a construction robot that’s helping repair our streets or build our homes. The joints on that robot are going to erode and face damage in unique ways, and it needs to learn how to adjust its own movement so it can maintain its performance based on its own kind of evolution and transition in its physical self. Also, the robots might be adjusting to new environments and we’d like that ability to be baked in so they can continuously and adaptively learn.
This also really enables personalization. It enables machines to get to know us and for us to have the assurance that that knowledge and data of ourself is in that robot and doesn’t leave that robot, I think is something will be very assuring. But that’s a huge piece, you know, training and inference in the same platform that is untethered.
John Koetsier: What’s really interesting about that, oh, there’s a number of things obviously, but I mean, I’m just thinking of a limping robot, for instance, you know we limp when we injure a joint and that is our adaptation to the limitation in a knee or a hip or something like that, and we function with that. And maybe we’ll repair the robot, but maybe the robot is inaccessible, or maybe the robot is on Mars or Pluto or who knows where, or maybe it’s too expensive, so limping and getting — you know, that’s interesting.
The other thing that you mentioned that’s super interesting is we want AI to make our lives better and we want robots to make our lives better, but that doesn’t mean that we want Amazon to know our deepest thoughts. That doesn’t mean that we want Google to know everything about our personal finances. If we can have AI that is personal, that you know, sure, it comes from somewhere and we’ve purchased it, or if you start to look at general AI you start thinking do you purchase that? Do you recruit that? Do you adopt that? Lots of questions there.
But it’s nice if you can get a system that learns you, understands you, and it stays in some sort of privacy corridor there. Really, really interesting stuff. How … it’s so hard to predict the future, you mentioned about speed, that’s always a moving target, right?
You want 100X speed, but you see what the speed is right now. In terms of adding speed right now, we seem to not be making chips much faster per se, we’re adding more chip sets on a dye or we’re creating chips that are more optimized for this task or for that task, or for energy efficiency, or for whatever and that’s how we’re making overall devices faster.
So are you keeping that sort of moving target in mind as you look at the performance levels that you want to hit?
Gordon Wilson: So we believe we can achieve that 100X improvement in speed and beyond, because we’re taking a different tack entirely, because we’re not on that roadmap of digital hardware that is very mature chips and they’re eking out performance from any number of, you know, kind of well-trod playbooks.
In the case of a fully analog neuromorphic chip, you have a neural network where you have analog neurons and synapses that are connected in sequence, and you can compare this, say, to other analog mixed-signal chips. So people are starting to get a lot of speed improvements by moving away from digital and doing, say, the synaptic operations either with photonic components or flash components.
But in all of those cases, they still actually have to translate between their kind of physical, their either physics or optical to digital for their neurons. And so they have these analog synapses, digital neurons, they kind of go back and forth. And that requires clocking, that requires slowing down the system.
So in a fully analog chip, when it’s designed well, you can have input to output and have the signal flow at wire speed from end to end. That is the full potential of an analog chip, and that’s why the speed here is so extraordinary, because we’re no longer working in this digitally clocked world, but we’re, again, exploiting the physics of the system, the physical nature of the system to do that math for us.
And you can perform inference as just a wave of electricity, from input to output.
John Koetsier: Amazing. I know that Tesla, for instance, is doing … we can debate whether the term ‘full’ is an appropriate modifier there, but full self-driving on Atom chips, right? I’m not saying that all the training and learning is happening there, but that’s what’s actively involved in the vehicle. So, just imagining what this could do for self-driving, for automated machinery, for robotics, for AI … it’s kind of mind blowing. Gordon, thank you for spending this time with us. I really do appreciate it.
Gordon Wilson: Thank you so much for the time, John. It was a pleasure.
TechFirst is about smart matter … drones, AI, robots, and other cutting-edge tech
Made it all the way down here? Wow!
The TechFirst with John Koetsier podcast is about tech that is changing the world, including wearable tech, and innovators who are shaping the future. Guests include former Apple CEO John Scully. The head of Facebook gaming. Amazon’s head of robotics. GitHub’s CTO. Twitter’s chief information security officer, and much more. Scientists inventing smart contact lenses. Startup entrepreneurs. Google executives. Former Microsoft CTO Nathan Myhrvold. And much, much more.
Consider supporting TechFirst by becoming a $SMRT stakeholder, connect to my YouTube channel, and subscribe on your podcast platform of choice: