Jeff Hawkins Is Finally Ready to Explain His Brain Research

Excerpt: A simple thought experiment may be useful to understand our model. Imagine you reach your hand into a black box and try to determine what object is in the box, say a coffee cup. Using only one finger it is unlikely you could identify the object with a single touch. However, after making one contact with the cup, you move your finger and touch another location, and then another. After a few touches, you identify the object as a coffee cup. Recognizing the cup requires more than just the tactile sensation from the finger, the brain must also integrate knowledge of how the finger is moving, and hence where it is relative to the cup. Once you recognize the cup, each additional movement of the finger generates a prediction of where the finger will be on the cup after the movement, and what the finger will feel when it arrives at the new location. This is the first problem we wanted to address, how a small sensory array (e.g., the tip of a finger) can learn a predictive model of three dimensional objects by integrating sensation and movement-derived location information.

But the word “column” implies discontinuous modules. You can actually observe this in sensory cortex of mice and rats, where each whisker has a 1-1 relationship with a “column” of neurons that primarily respond to that whisker. This observation is what originally led neuroscientists to propose the concept of the column as a fundamental unit of organization of the cortex.

But it turned out that lots of other brain regions don’t have this kind of discontinuous organization. Neuroscientists still use the term “cortical column” but they’re really just referring to the fact that vertically aligned neurons seem to be closely connected.

When the Numenta people talk about “cortical columns” they seem to be describing the original idea of cortical columns that didn’t pan out. It’s really weird.

It's been a bit frustrating that after 15 years they still didn't make headlines with something more practical and groundbreaking, so I'm a bit excited to hear what will be unveiled at this conference tomorrow.

So it doesn’t seem to me that the problem is that Hawkins has been uncommunicative, it’s that the wider academic and research community just hasn’t been interested. Perhaps that’s simply because his theories don’t have merit, but it’s also hard to shake the feeling that’s it more because his career, and therefore his theories, hasn’t fit the standard academic mould.

It seems to me that, if you want to be an influential theorist, you’re expected to study at a prestigious university, work with an older, respected researcher while getting your PhD, then become a professor, setup a research lab, mentor some PhD students of your own, then finally maybe spin off your research into some startups, or take a high-paying research position at a big company.

Hawkins hasn’t followed this path, so he has no network of peers and former students to evangelise his work and carry it forward. He’s done research and published it, like the scientific ideal says you should, but he’s missing the structure necessary to make it land in the research world.

It’s like a sort-of reverse cargo cult. He’s done the work of constructing a real airstrip and working planes, but he can’t persuade anybody to fly, because they’ve never seen or heard of air travel before.

The ideas represented and level of detail in their theory is such that the most likely response from everyone else is just 'Nice theory' and a shrug.

Hawkins and Co. have been working on this for years. Numenta sends frequent newsletters and try to keep people interested, but there is little concrete there except some interesting ideas.

This is an exciting idea to me. I don’t fully understand it, but it resonates on some intuitive level. Makes me think of spatial audio, echolocation, our sense of balance, all these forms of intelligence that rely on understanding space.

Random trivia- I interviewed with Numenta back in 2007. Got an offer, was excited by the vision, but I didn’t see how they would have any kind of product within the next 5 years. 11 years later, they’re maybe about to launch something. Much respect for them pushing through on this wildly ambitious idea. I’m rooting for them.

In computer hardware and software terms: we don't necessarily need to clone an x86 CPU, glue and I/O chips, and run an actual copy of firmware, bootloader, and Windows. If we understood just one of these, how the x86-like CPU works, and could implement a functional simulator that works similarly enough that would be extraordinarily useful. ANNs are a primordial stab at this.

But this article is incredibly light on detail. I hope he does something like a follow up book to explain what he has learned in the time since.

> In the neocortex 6 layers can be recognized although many regions lack one or more layers, fewer layers are present in the archipallium and the paleopallium.

What this means in terms of optimal artificial neural network architecture and parameters will be interesting to learn about; in regards to logic, reasoning, and inference.

According to "Cliques of Neurons Bound into Cavities Provide a Missing Link between Structure and Function" https://www.frontiersin.org/articles/10.3389/fncom.2017.0004... , the human brain appears to be [at most] 11-dimensional (11D); in terms of algebraic topology https://en.wikipedia.org/wiki/Algebraic_topology

Relatedly,

"Study shows how memories ripple through the brain" https://www.ninds.nih.gov/News-Events/News-and-Press-Release...

> The [NeuroGrid] team was also surprised to find that the ripples in the association neocortex and hippocampus occurred at the same time, suggesting the two regions were communicating as the rats slept. Because the association neocortex is thought to be a storage location for memories, the researchers theorized that this neural dialogue could help the brain retain information.

  before the world can build artificial intelligence, it must
  explain human intelligence so it can create machines that
  genuinely work like the brain

The concept described in this article is the same one in his book "On Intelligence", which was written long ago. Before we knew the impact of neural network arrangement and architecture on its performance.

Examples like that paper that described a task-specific region of the brain that handles face recognition is not supported by his model.

What would interest me more than the one algorithm is to find out the loss functions that steer our brain into fast learning after birth. We have prior knowledge about the world encoded in the structure of the brain and our loss functions (what the brain optimises for). The brain has many such loss functions - it can't possibly do all it does with just one like neural networks, and they are genetically evolved. These priors could be readily transferred to AI models.

And perhaps someone will be able to correct me on this, but didn't Kant treat spatial awareness as a priori and presumably rather fundamental?

Nonsense.

What % parts of a bird were necessary to understand for the first successful aircraft?

It may be very helpful all the way to unnecessary, no one knows that yet.