| blog
Brains and Machine Intelligence, A Long Time Coming
Recently one of our U.S. agencies, the IARPA (Intelligence Advanced Research Projects Activity), put out an RFI (request for information) stating they are interested in machine learning algorithms that are based on neuroanatomy. Although this RFI is only a few weeks old, I feel like I have been waiting for it for thirty years.
Here are a few quotes from the RFI.
“IARPA is interested in understanding the neural algorithms that form the basis of inference and recognition in the brain.”
“This RFI specifically addresses mathematical, computational, or otherwise executable models of cortical computing primitives that are supported by known neuroanatomy.”
“Information regarding models that are purely conceptual, schematic, or descriptive in nature (i.e. models that cannot be instantiated), or regarding models that are not grounded in neuroscience, is not sought at this time.”
I couldn’t have been happier than when I read these words. For thirty years I have been promoting the idea that the path to machine intelligence must start with a detailed understanding of how the brain works. Until recently this idea had few adherents. I vividly recall being told by faculty at the MIT AI Lab that studying brains was a waste of time! Prominent neuroscientists told me it will take hundreds of years to understand the neocortex. And many machine learning experts have told me that brains are, at best, good for inspiration but the details are not important. Of course not everyone felt this way, but for many years I could count on one hand the people I knew who studied the neocortex in order to build intelligent machines.
But times are changing. In the past year there has been a resurgence of interest in hierarchical memory models (newly renamed “deep learning”) which are loosely based on cortical principles. DARPA (another U.S. agency) is working on creating hardware that closely mimics cortical architecture. There is the European “Human Brain Project” and the recently announced U.S. “BRAIN Initiative”, both striving to understand how the neocortex works. And now there is the IARPA RFI, which is remarkable in that it explicitly says they are only interested in learning algorithms that are based on the detailed anatomy of the neocortex. Of course the Cortical Learning Algorithm developed here at Numenta and now part of the NuPIC open source project is a prime example of the “cortical computing principle” being sought by IARPA. We will probably respond to the RFI.
It is too early to claim this new interest in cortical computing is here to stay and whether it will continue to adhere to neuroanatomy. But it feels that the momentum of the machine intelligence world is shifting slowly towards making biology an equal partner and this is a good thing.
Here is a link to the IARPA RFI. Thanks to Kevin Archie who sent it to the NuPIC email group.
Back to BlogJeff Hawkins
Founder, Numenta Inc.