Thursday, March 10, 2016

Brain Efficiency

I've been thinking a lot recently about how efficient the brain is. I like to spend time thinking about how neural interfaces will change the nature of humanity. Presumably, at some point, it will be possible to create computers that have intelligence that is on par with that of humans. Does that mean the end of humanity? Maybe, but maybe not.

Computers were designed to crunch numbers, and they are ruthlessly efficient at it. Unfortunately for them, most of the tasks we associate with "intelligence" are not associated with number crunching operations. Human intelligence is essentially a feat in pattern recognition - when we recognize patterns, we learn to predict the future based on previous experience. We can teach computers to perform pattern recognition tasks, but first we have to convert those tasks into number crunching operations. This is a pretty inefficient way of solving those problems, but we make up for that inefficiency by using super fast computers. Think of it as trying to drive a square peg into a round hole: its a bad idea from the start, but you might be able to make some progress if you just agree to use a humongous hammer.

So, number crunching machines are inherently inefficient at recognizing patterns. Is there another type of computing system that would be more efficient? Yes! Millions of years of evolution have placed a very efficient pattern recognition system right between your ears: your brain. Brains are insanely efficient at pattern recognition tasks. Lets see how efficient:

  • The average adult consumes about 2,000 calories per day
  • Of those, about 1,300 are the "resting metabolic rate" which is basically how much energy you'd burn if you just lay in bed all day and didn't move - its what you burn to keep your organs running to stay alive
  • Of those, about 20%, or 260 calories, are consumed by your brain
  • 260 calories in 24 hours converts to about 1.1 million joules per 86400 seconds, which reduces to 12.7 joules per second which is basically 13 watts.

That's right. 13 watts to keep the universe's most sophisticated intelligence machine operational. Astounding. By comparison, the fancypants laptop I'm using to type this blog post with consumes about 45W. The Watson computer that succeeded in playing Jeopardy reportedly uses something like 200,000W, a factor of over 15,000x more. Perhaps a more impressive feat than Watson beating Ken Jennings would have been Watson beating 15,000 Ken Jennings! And lets remember, Watson didn't 'have fun' playing Jeopardy, or parlay its experience into planning for its future: Ken did. Even super computers like Watson, with all their power, are inferior to the wonder of the human brain.

So, will a computer ever become as smart as a person? While it's hard to say, I believe that it will be damn near impossible for a computer to become as smart as a person using only 13 watts of power. I suspect that the only material that can be made to operate as efficiently as a human brain is ... a human brain. You'll never get down to 13 watts with transistors, memristors, or whatever the next great innovation is. Nothing beats neurons with respect to efficiency.

A separate question worth asking is whether a computer that can think as fast as a person (regardless of the wattage) can compete with humanity in terms of collective intelligence. I'll save that question for another day.

1 comment:

  1. A well thought-out post! The efficiency of cellular respiration in and of itself is something to be marveled at - something to the tune of 40%.

    While I agree that the disparity between the power usage of electronic computation and neural computation as a whole isn't something that can be bridged with the standard electronics paradigm, I do think that there are some ways it may be competitive.

    For starters, I would venture to assume that memory maintenance in a brain, as compared to a hard disk, requires more upkeep over time.

    Another thought would be that energy usage is task dependent. For example, we may require different sized cultures if we were to use an MEA and train the culture on an addition task versus xor tasks versus simple statistical tasks due to the architecture being biased towards different things. The same would be true if the computer was stripped of the OS and such and was being used to calculate those same things - especially so if we made custom circuits for each task. I'd still bet on neurons winning (and also they have much more flexibility in the tasks they can solve), but I'd like to think electronics will show some merits in energy consumption.

    At any rate, keep up the awesome posts! I look forward to reading more.