A very interesting model. It aligns with that general principle of biology that dictates biological systems should do as little work as necessary. The robustness to cell death exhibited in figure B near the end is reminiscent of our own brain's robustness.
This theory is very intriguing and intuitively makes sense. I read Hawkins' book way back in 2004 but there haven't been any successful implementations by Numenta. Mostly toy examples AFAICT. I would love to be wrong and see some great applications come out of this (much like the AI spring 2011- using conv nets etc.)
> By firing earlier it inhibits neighboring cells, creating highly sparse patterns of activity for correctly predicted inputs.
This part is so vague. It seems to lack an explanation of how interneurons inhibit other neurons nearby. Also, wouldn‘t sparsity even occur without the early firing enabled by distal pattern matching?
> When relatively few neurons are active relative to the population, then such pattern recognition is robust.
It's true that the columnar inhibition is not the best supported part of the theory but seems a very strong deduction based on many papers that don't directly support but show the same behavior as this part of the theory. We know that minicolumns exist and cells in minicolumns share receptive fields. But they aren't always active together and HTM theory provides a hypothesis for why. This very recent preprint paper out of Michael Berry's lab at Princeton shows almost identical behavior that HTM sequence memory would predict and I'm not aware of other theories that would have predicted this behavior:
> When relatively few neurons are active relative to the population, then such pattern recognition is robust.
At a high level, it seems that the constraint of having only a few neurons active is equivalent to a simplicity constraint, thus implementing Occam's Razor and yielding generalization.
I'm not convinced that the function of the brain is prediction: it seems that prediction is subsidiary to defining an action policy.
Prediction is only useful for an organism inasmuch as it allows it to calculate expected future utility for actions. And these actions are the only things that lead to different outcomes in terms of evolutionary fitness, thus the actions that are ultimately output by the brain are the only things that determined its evolution.
If the purpose of prediction is to calculate expected future utility of different actions, then it does not follow that a general-purpose prediction device will be useful, because the prediction device might use all its energy predicting aspects of the environment along dimensions that are irrelevant to utility. A useful prediction device would only predict along useful dimensions, and may be very different in behavior from a generalized autoencoder. As an example of this difference as it comes up in the field of deep learning, consider the case of machine translation: you could either train a sequence model (LSTM or whatever) to autoencode sequences---i.e., to be able to predict them---and then use the resulting representations to do translation, or you could train end-to-end where the objective function is translation quality. It turns out the latter yields better results.
Maybe the brain, or part of the brain, is a prediction engine and another part does action selection based on the predictions. But then why would we identify intelligence with the prediction engine part rather than the two parts combined? Searching for an optimal action is a very different task than prediction; it seems you need both to have what anyone would call intelligence.
There's a danger here in getting too caught up in words (where what they're trying to capture is what matters), but can't "searching for an optional action" accurately be seen as a kind of prediction?
My current model of how the brain works comes from this helpful gloss of a very technical book, Surfing Uncertainty (http://slatestarcodex.com/2017/09/05/book-review-surfing-unc...), which lays out the Predictive Processing model of the brain that is (apparently) the working model for those at the vanguard of neurobiology presently.
I'll try to summarize (my own layman's understanding of) the Predictive Processing theory further here, for people who don't want to read the article:
As the PP theory goes, not only is the human brain a "general-purpose prediction device", but it has no other components. When we think, we're simply making predictions for the way the world will be a moment from now; and then we attempt to reconcile those predictions with the image we get of the world a moment later. In equal parts, depending on our confidence in our senses vs. our confidence in our model, we either do this by "learning" (i.e. correcting the predictions to match the signal) or by "acting" (i.e. correcting the signal to match the predictions.)
The latter, "acting", is done by simply playing out the error signal (i.e. the difference between "the way the world is" and "the way the world should've been") to motor neurons, which interpret those error signals as motor commands. We seek to be less confused by the world by forcing the world to become more like our model of it!
And (terminal) preferences—those are just persistent exogenous-to-the-predictive-process chemical messengers that bias neurons into outputting a different "world that should've been" than a disinterested observer, only interested in predicting, would've generated. This causes the brain's "acting" process (i.e. correcting error by changing the world) to change the world in the direction of one's preferences, rather than in the direction of one's model. In effect, brains minimize error between the world that they observe, and the world they'd like to observe.
PP theory doesn't say much about where those exogenous chemicals come from, but obviously the neuroendocrine system exists, and consists of a bunch of glands that throw chemicals at the brain.
I am working on a hierarchical predictive spatio-spectro-temporal pattern model, which uses convolution-like patterns that use time and frequency of spikes, where each spike indicates a pattern 'match'. Patterns evolve over time from weak, non-specific (in space, time and frequency) preferences to strong, specific preferences. I think this approach has the potential to become the substrate on which we could build general AI.
I am very curious to know what labs/companies are using this approach.
No papers yet, still in (early) implementation/experimentation phase. Currently surveying research about minimal sentience/consciousness exhibited by animals and corresponding neural correlates. If there is enough data to isolate functional basics of minimal consciousness, I would like to implement that as the first "function" of the HTM model.
One of the key points of this theory is that its learning rule is hebbian. A variation of "If it fires together it wires together." Individual synapses make or break connections based on if the input neuron firing immediately proceeds the firing of the next neuron in the chain.
This is distinctly different than say a recurrent neural network where all neurons have outputs at each timestep and weights are updated based on the derived contribution of each weight to the quality of a final output.
The first rule is very local, both in time and space, and the second is in some senses "global."
Backprop has obviously had much more success in ML than any Hebbian learning rule. However we're pretty sure biological learning is essentially Hebbian. The merger / reconciliation of both these rules is one of the most interesting areas of research right now.
I might start with this paper from Bengio et al. and then read more recent papers which cite it. The problem hasn't been solved though so don't be surprised if you're not fully convinced by the arguments :)
Small correction: Learning algorithms that determine the contribution of a unit to the result over time can also be local in time, namely by tracking the contributions online (for example by eligibility traces or RTRL).
It's possible that backprop has more success for narrower bounded problems and is "faster," but maybe because it's greedy. Biological learning might benefit from being less greedy.
> The neuronal model used in most artificial intelligence networks contains few synapses and no dendrites.
Does anyone understand why this is true?
In a typical multi-layer network, doesn't each node in a lower layer connect to every node in a higher layer? All the {L(i, n-1), L(a, n)} edges going from the nodes in layer n-1 to to a particular node (a) in layer n would constitute a dendrite.
In this article they're using the word "dendrite" to mean a bit more than just an input to a neuron (which does exist in NN models) - one of the often criticized simplifications of most neural net models is that they don't account for the structure and function of the entire dendritic tree. Specifically, in the image you're referencing in the article they point out that synapses very near the soma directly cause action potentials, whereas further out they will cause some depolarization but not actually cause the cell to fire.
The way some authors explain it is that you can think of a neuron as really having an entire neural net embedded inside of it (based on the structure and function of the dendritic tree), that does some non-trivial amount of information processing even before you consider the connectivity of neurons to each other. Exactly how non-trivial that is is a deeper question, but it's worth noting that these structures are extremely plastic, and change over timescales of seconds to minutes, so it's not hard to imagine that these details are significant.
Ah, thanks. So another way of wording it would be to say that the activation function of nodes in a neural network is a gross oversimplification of the amount of logic that happens via dendrites in actual neurons.
I think that Hawkins has a great summary on some things that are necessary for intelligence [0]. Personally I think the tie between motor actuation and sensory detection is one of the most important -- did a project on that in 2010. Motor-sensory link is essentially how all intelligence evolved. Even c.elegans has neurons to determine whether it should wiggle one way or another.
On the surface he seems to be cramming a whole lot of "solutions" into his HTM (hierarchical temporal memory). HTM is an interesting implementation and the sparse coding is definitely a benefit. However I think he is focused too much on his baby and not on other techniques that may fulfill the necessary components more efficiently.
That is just on the surface. With a product / research balance maybe we just aren't seeing all the cool things going on underneath in research, but it does seem like that research will be shoe-horned into HTM whether or not it is the best architecture.
30 comments
[ 2.8 ms ] story [ 73.1 ms ] threadI also read Hawkin's book and loved it. I'm still hoping that they'll get something going.
This part is so vague. It seems to lack an explanation of how interneurons inhibit other neurons nearby. Also, wouldn‘t sparsity even occur without the early firing enabled by distal pattern matching?
> When relatively few neurons are active relative to the population, then such pattern recognition is robust.
Why?
https://www.biorxiv.org/content/early/2017/10/03/197608
As far as sparsity supporting robust pattern recognition, this paper details the math that shows this:
https://arxiv.org/abs/1503.07469
At a high level, it seems that the constraint of having only a few neurons active is equivalent to a simplicity constraint, thus implementing Occam's Razor and yielding generalization.
Prediction is only useful for an organism inasmuch as it allows it to calculate expected future utility for actions. And these actions are the only things that lead to different outcomes in terms of evolutionary fitness, thus the actions that are ultimately output by the brain are the only things that determined its evolution.
If the purpose of prediction is to calculate expected future utility of different actions, then it does not follow that a general-purpose prediction device will be useful, because the prediction device might use all its energy predicting aspects of the environment along dimensions that are irrelevant to utility. A useful prediction device would only predict along useful dimensions, and may be very different in behavior from a generalized autoencoder. As an example of this difference as it comes up in the field of deep learning, consider the case of machine translation: you could either train a sequence model (LSTM or whatever) to autoencode sequences---i.e., to be able to predict them---and then use the resulting representations to do translation, or you could train end-to-end where the objective function is translation quality. It turns out the latter yields better results.
Maybe the brain, or part of the brain, is a prediction engine and another part does action selection based on the predictions. But then why would we identify intelligence with the prediction engine part rather than the two parts combined? Searching for an optimal action is a very different task than prediction; it seems you need both to have what anyone would call intelligence.
I'll try to summarize (my own layman's understanding of) the Predictive Processing theory further here, for people who don't want to read the article:
As the PP theory goes, not only is the human brain a "general-purpose prediction device", but it has no other components. When we think, we're simply making predictions for the way the world will be a moment from now; and then we attempt to reconcile those predictions with the image we get of the world a moment later. In equal parts, depending on our confidence in our senses vs. our confidence in our model, we either do this by "learning" (i.e. correcting the predictions to match the signal) or by "acting" (i.e. correcting the signal to match the predictions.)
The latter, "acting", is done by simply playing out the error signal (i.e. the difference between "the way the world is" and "the way the world should've been") to motor neurons, which interpret those error signals as motor commands. We seek to be less confused by the world by forcing the world to become more like our model of it!
And (terminal) preferences—those are just persistent exogenous-to-the-predictive-process chemical messengers that bias neurons into outputting a different "world that should've been" than a disinterested observer, only interested in predicting, would've generated. This causes the brain's "acting" process (i.e. correcting error by changing the world) to change the world in the direction of one's preferences, rather than in the direction of one's model. In effect, brains minimize error between the world that they observe, and the world they'd like to observe.
PP theory doesn't say much about where those exogenous chemicals come from, but obviously the neuroendocrine system exists, and consists of a bunch of glands that throw chemicals at the brain.
I am very curious to know what labs/companies are using this approach.
See this blog post for why I chose consciousness as the primitive to build (instead of say demoing the model's characteristics on common tasks like mnist digit recognition) - https://medium.com/creating-artificial-consciousness/the-cas...
This is distinctly different than say a recurrent neural network where all neurons have outputs at each timestep and weights are updated based on the derived contribution of each weight to the quality of a final output.
The first rule is very local, both in time and space, and the second is in some senses "global."
Backprop has obviously had much more success in ML than any Hebbian learning rule. However we're pretty sure biological learning is essentially Hebbian. The merger / reconciliation of both these rules is one of the most interesting areas of research right now.
https://arxiv.org/pdf/1502.04156.pdf
Edit: linked to specific post in forum thread instead of the whole thread
I’m drawing a blank how you’d run a nn that way?
Does anyone understand why this is true?
In a typical multi-layer network, doesn't each node in a lower layer connect to every node in a higher layer? All the {L(i, n-1), L(a, n)} edges going from the nodes in layer n-1 to to a particular node (a) in layer n would constitute a dendrite.
The way some authors explain it is that you can think of a neuron as really having an entire neural net embedded inside of it (based on the structure and function of the dendritic tree), that does some non-trivial amount of information processing even before you consider the connectivity of neurons to each other. Exactly how non-trivial that is is a deeper question, but it's worth noting that these structures are extremely plastic, and change over timescales of seconds to minutes, so it's not hard to imagine that these details are significant.
On the surface he seems to be cramming a whole lot of "solutions" into his HTM (hierarchical temporal memory). HTM is an interesting implementation and the sparse coding is definitely a benefit. However I think he is focused too much on his baby and not on other techniques that may fulfill the necessary components more efficiently.
That is just on the surface. With a product / research balance maybe we just aren't seeing all the cool things going on underneath in research, but it does seem like that research will be shoe-horned into HTM whether or not it is the best architecture.
[0] https://www.youtube.com/watch?v=4y43qwS8fl4&app=desktop
Starting at ~8m20s