Recently there’s been several neuromorphic chips or other AI chips that have been released. So, how long before the no. of neurons on a single chip equals that within the human brain? Yes, I realize that our current artificial neurons are not as advanced as our biological ones, but I’ll use this as an approximation of the lower bound of my estimate. I’m going to focus on spiking neural networks since they’re the closes to biological neurons. Chips such as Intel’s Neural Compute Stick don’t feature here since they’re not for spiking neural networks.
Wikipedia has a list of the number of neurons in the brain of each type of animal. A cat has 760 million, a dog has about 2.253 billion, a chimpanzee has 28 billion, a human has 86 billion and an African elephant has 257 billion neurons. Surprising that cats can enslave humans so well, yet have 100x fewer neurons. Humans also have about 14 trillion synapses. These figures are for the whole nervous system.
Intel has their Loihi chip that was announced on 25 September 2017. It supports 131,072 neurons and about 130 million synapses. Back in 2009, there was Neurogrid with 16x256x256 neurons, or about 1 million neurons. Assuming Moore’s Law continues to hold indefinitely, how long will it take the existing architectures to get to the scale of a human brain? The formula for compound growth is , where is the no. of 2-year periods, is the current no. of neurons and is the future no. of neurons. Remember that Moore’s Law ” is the observation that the number of transistors in a dense integrated circuit doubles about every two years.” according to Wikipedia. So, scaling up Intel’s chip should reach the 86 billion neurons of the human mind in 39 years. Scaling up Neurogrid will allow us to equal the human brain in 33 years. We’ll have the silicon equivalent of a dog’s brain in 22 to 28 years, and an elephant’s brain in 36 to 41 years. The silicon cat’s brain might arrive in 19 to 25 years. The Intel Loihi design might reach the no. of synapses in the human brain in 33 to 34 years.
So, by 2044, I’ll have perfect virtual pets that are smart enough to enslave me, but roughly 10 years later I’ll either be fighting brilliant Terminators or I’ll have a human-level robotic butler to attend to me in my old age. So, Elon Musk might just live long enough to see the world-ending AI he fears. Perhaps cat and dog brain-level is where we should stop with AI chips, for our own good. Of course, this is meant half in jest. Who knows, perhaps quantum computers will dramatically cut these estimates, or we’ll hit a brick wall in making transistors smaller.
Finally, why aren’t elephants in charge of the world? I guess being slow, lacking fingers and walking on all fours really sucks. I do have to wonder though, what are elephants thinking about all day. They’ve got 3x as many neurons as us, and no Twitter or Youtube, so I wonder if they’re bored out of their minds? Ok, enough sarcasm for one blog post, although I do now wonder just how smart are elephants and what do they know about the world?
By the way, if you like the cover image, the original is here. It is a full HD wallpaper.