What if we live in a Jeff Hawkins world?

Related: What is so cool about grid cells?

Jeff Hawkins from Numenta believes that he’s cracked the secret of intelligence. He recently complemented his earlier work on hierarchical temporal memory with “the thousand brains theory of intelligence”.

What if he is correct?

What are the implications for machine intelligence timelines?

Hawkins’s work points in two directions with respect to AGI.
On the one hand, he is quite confident that variations on current neural networks with point neurons and weights updated through backpropagation are insufficient for general intelligence. He believes that there is important computation that is done by the thousands of synapses that are too far from the soma to cause a neuron to fire and that the need for a large number (20-ish) of synapses for a firing is one of the reasons why the brain is so robust.

Moreover, he believes that embodied cognition is a crucial component of intelligence, as navigation and spatial reference frames are necessary for building cognitive maps in domains of higher cognition like math. This also points to longer timelines as progress in robotics is quite sluggish and roboticisists tend to see AGI as really far in the future.

On the other hand, he is quite confident that

(a) The key to general intelligence is the neocortex. Older parts of the brain are important for driving hard-wired behavior but it’s the neocortex that drives learned intelligent behavior.

(b) his team found the core principles of neocortical computation, which is the key to general intelligence

If this is true, and if it’s sufficiently simple to build machine learning systems that operate on these principles, general intelligence might be surprisingly near.

Some important background:

  • The neocortex emerged in mammals some 200M years ago.
  • In the human brain, neocortex occupies 70-75% of the mass.
  • The neocortex is remarkably uniform relative to older parts of the brain.
  • In 1957, Vernon Mountcastle’s classic paper put forth the hypothesis that based on the uniformity we should think that the neocortex does the same computations for all perceptual signals.
  • According to Hawkins, the only difference between areas is what inputs they’re connected to.
  • There is a lot of good descriptions of Hawkins’s theories. For hierarchical temporal memory, the book On Intelligence is a good introduction. Here are introductory videos to the Thousand brains hypothesis.
  • A paper by Marcus and Marblestone taking the opposite view that cortical computations are diverse and vary across regions of the brain. Marblestone has since changed his mind, Marcus has not.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s