Humans Programmed to Walk?

I heard Jeff Hawkins say in this podcast that humans are programmed to walk, and that we don’t learn to walk. The PI at my lab is trying to explain how children learn to walk using an exploration & reinforcement learning framework. Does anyone know any papers that would explain what Jeff meant in more detail?


I think he was talking about central pattern generators in the spinal cord. He talked about them in Sensory-motor Integration in HTM Theory, by Jeff Hawkins - YouTube a while ago.

These papers look relevant. It looks like there might be plasticity, but I’d guess not learning to model the world like cortex. In humans, there’s a lot we don’t know because experiments can’t cause harm.


My personal guess is it is more efficient from evolutionary perspective to have newborns programmed with a desire + few hints to acquire an ability than with the entire information skill set needed to implement it.
An imperative like “rise and walk!” is both more simple and flexible than all details about how to walk.

Thank you! I will read these :slight_smile:

That’s interesting to contemplate. I should have thought of this sooner, but a lot of animals just fall out of their momma and can walk very soon thereafter. It wouldn’t be too surprising to learn that humans are an exception to that, but I am not certain of course.

1 Like

These are mostly grazing herbivores which need to stand up in minutes after birth to avoid predation.
They still spend quite a while after birth to …fine-tune their walk, running, turning and jumping they-re quite fun to watch.
But compared to them kittens or puppies are quite noobs. Cute Newborn Kittens Learning To Walk Compilation || NEW HD - YouTube

1 Like

My (uninformed) understanding was that all mammals are born able to walk on 4 legs.
This is useful for other animals that normally walk on all fours, but less useful for humans where walking on four legs is crawling.

There is a built in walking mechanism and drive to walk. Humans have indications of this mechanism early as babies but disappears. Human walking is fully “learned” given a neural system which supports the process. Very interesting is to see the standing mechanism and drive to stand for deer and the like. No so much for machine learning though or cortical learning.

1 Like

Same, make a newborn child legs touch at treadmill while you hold him, and he will instinctively walk (generate walking patterns).

1 Like

I really am not into cute cat stuff on the Internet, but that video made my day.

May not be relevant, but chicken and ducks are bipedal creatures that can walk, run and jump perfectly just a few hours after birth.

1 Like

Yes, but mechanically they are more stable than humans. Definitely it is interesting than many animals have walking built in but require some fine tuning which occurs as they experience walking. Really all quite interesting.

Pure conjecture here… We have the wiring to be able to walk but the signal conflict between cerebral development and cortical development means we need more time to resolve the process as to which sequence and where takes priority. The larger cortex in humans creates more noise early in life ?

Quadrapeds may also have hard wired basic (cerebellum) movement sequencing (walk, trot, run) at birth and the “fun” is learning excess/variants modes (cortex sequences) of these patterns that result in jumping, play, etc. and as RBo states, you get a lot more stability with 4 legs that does not have to rely on a gyro stability feedback loop to perform basic movement…

Again, all a guess…

Try not to introspect…

A horse can stand and walk on day one, and rapidly develops a set of (usually) 4 natural gaits with trade-offs between speed, energy usage, stability on various surfaces and so on.

Then a horse can be trained to add gaits: pacing, ambling, racking and so on. They get better with practice, and eventually become automatic. There is no reason to believe continuous cortical attention (but this can be tested).

A gait is a sequence of movements, with motor output and sensory input feedback that is highly time constrained. It sounds a lot like HTM, but sub-cortical. So one should be looking for a repeating brain structure (like columns), a data representation (like SDR), perhaps in the cerebellum, with both sensory and motor connections.

This article was a good read, thank you for posting it!

Frontiers | Central Pattern Generator for Locomotion: Anatomical, Physiological, and Pathophysiological Considerations | Neurology

"several decades of evidence has led to the conclusion that walk-
ing, flying, and swimming are largely controlled by a network of spinal neurons generally
referred to as the central pattern generator (CPG) for locomotion. It has been subsequently
demonstrated across all vertebrate species examined, from lampreys to humans, that this
CPG is capable, under some conditions, to self-produce, even in absence of descending
or peripheral inputs, basic rhythmic, and coordinated locomotor movements.

1 Like


That’s interesting then perhaps what needs to be learn is some kind of synchronize/align (at least) two sensory-motor streams.

Something like that happens with binocular vision too? Does the infant sees some strange complex image until the two foveas align on the same spot then “whoa! how neat is the whole world now!”

Does this alignment thing apply else/everywhere? When I look at a hard-to-guess MNIST digit I try to see how well it overlaps with what I imagine a “4” or “9” might look. Instead of moving fovea I stretch the inner models of the two digits to check which one stretches less to better overlap over the strange sign that I see.

Ok, maybe not a guess…

Take a look at the signal sequencing for micro expressions in body language (timing) as these I believe originate in conflict to subsequent HTM sequencing that then overrides the initial “animal” response.

The HTM network takes several more cycles/iterations to vote and override to take priority due to the way that HTM / cortical columns are meshed together. To have micro expressions originate (only) in the cortex would imply that the cortex can act (ungated) on outputs that are still not fully evaluated, which would then imply that speech would be incoherent.

Tourettes is a possible example case of sequencing override outside HTM / cortex.
“Tourette’s revealed significantly more gray matter in the thalamus, the hypothalamus and the midbrain than in those without the disorder.”

1 Like