TR: Some people think young technologists need to spend more time learning how to market their ideas.
MEAD: Science is not just about self-expression; you have to be able to explain what you’re doing. Dick Feynman was one of the best marketers I have ever met. He never wanted to admit it – in his day, anything entrepreneurial was socially unacceptable for an academic – but he was able to position physics as something exciting, in a way that has survived to this day.
TR: You and Feynman were behind a big neuromorphic-computing project launched at Caltech in the ’80s. What happened?
MEAD: Part of the problem was the refusal of the CS [computer science] community to have a new thought – the fact that there might be inherently more powerful ways to do computing. People said, “Everything’s a Turing machine, and that’s that.” No matter that we already have a working example of a massively parallel machine – the animal brain. And meanwhile, now, the quantum computing guys have come along and showed yet another alternative model – one that in theory will solve problems that are exponentially unsolvable by a Turing machine. I’m making no statement about the realization of quantum computers – we still don’t know about that. I’m just talking about our understanding of computing in the abstract. You need a fundamentally new conception of that if you want to try to make a better machine.
TR: Another neurally inspired company you’ve been involved with makes advanced hearing aids, Sonic Innovations.
MEAD: The thought process there came from thinking about how human hearing works, but again the actual device is just a little digital signal processor. The same thing happened with the idea of neural networks, by the way. They became just another algorithm for existing computers.
TR: What about Foveon, the camera company you founded in 1997? Most people probably don’t realize that its roots are in studies of the eye.
MEAD: We started out making models of the retina, which by itself might make a big difference to a few people, but it’s not enough of a commercial opportunity to justify big investment. What we realized was that if you took what we were doing and strip out the retina part, that’s a really good image sensor – so let’s do that. Foveon technology captures light directly, consuming less power and requiring far less processing than the file captured by a conventional digital camera. But when we explain it today, we don’t have any reference to anything neural.
TR: So we’re still at square one with neuromorphic computing?
MEAD: Actually, quite a lot of progress has been made. One of the exciting things that grew out of neuromorphic thinking is Lloyd Watts’s company Audience. They’ve got a working cochlear model that builds a significant portion of the auditory pathway – including precision signal recovery and sophisticated analysis – into a chip-level component. It’s more than just a better microphone; think of it as the auditory front end for any device that wants to use sound as an input.
TR: Voice recognition lives!
MEAD: Voice recognition as we know it is really brain dead. I shouldn’t say brain dead – a lot of smart people have worked on it for many years. But it’s an old paradigm. It’s advancing logarithmically with processing power; that’s about it. And yet we have these incredible working models right here – our own eyes and ears. That’s where we want to be looking.
TR: Hearing, vision – the same problems you picked out nearly 20 years ago are still interesting problems.
MEAD: They’re even more interesting, because we’re starting to know enough about them to make some progress. It’s taken this long to get the engineering-oriented people talking to the physiology people. Lawyers talk about “Chinese walls” in organizations; well, the barriers between scientific disciplines have been fierce.
TR: Is it the inherent difficulty of adapting digital technologies to our mostly analog human world?
MEAD: Digital abstraction is a wonderful thing. It substitutes a very simple set of logic operations – “and,” “or,” and “not” – for an infinite set of physical things. Working in analog is much harder, because there are essentially countless ways for the thing to go wrong. You’re working with the physics itself, rather than with some very small set of circuits that have been crafted to show digital behavior.
TR: We can’t let you get away without asking about Moore’s Law. You get a lot of credit for its formulation.
MEAD: Gordon had observed what was happening and asked me how far things could go, how small you could make the transistors. We did some work in the lab, and the answer turned out to be .15 microns [150 nanometers], maybe smaller. That was shocking at the time, but it turns out to have been conservative.
TR: So how far can it go?
MEAD: I looked at things again a few years ago, and if you don’t do anything differently, you can get down to 30 nanometers – a factor of five from what we originally said was going to be easy, and still a long ways from where things are today. So it’s certainly not going to stop.
And at the same time, we don’t have to keep doing things exactly the way we are doing them today. I for one certainly hope we don’t.
Salisbury, CT-based writer Spencer Reiss likes to interview people smarter than he is. The last time he did it for TR was with venture capitalist Michael Moritz, the man behind Google (April 2004).