How does the human brain construct intelligent behavior? Computational models have proposed several mechanisms to accomplish this: the most well known is "Hebbian learning," a process mathematically similar to both principal components analysis and Bayesian statistics. But other neural learning algorithms must exist - how else could the brain disentangle mere correlations from true causation?
Temporal precedence helps to some extent - and does seem to play a large role in Hebbian learning (e.g., spike-timing dependent plasticity). But the smell of rain does not actually cause rain - although it can both precede and follow it - so temporal precedence is not a complete answer to disambiguating causality from correlation.
Environmental interaction is another possibility - and one function of children's exploratory play does appear to be the resolution of causally-ambiguous events.
But exploratory play is just a specific example of a larger class of learning situations, where the environment itself can provide an "error signal" in testing causal inferences - appropriately known as "error-driven learning" in computational models. Errors in predictions about the environment can be used to better extract causation from correlation.
Although the classic implementation of error driven learning (known as "backpropagation") has been criticized as biologically implausible from a cellular neuroscience perspective, cognitive and computational neuroscience has clearly converged on the idea that some learning of this sort must be occurring. Randy O'Reilly's GeneRec algorithm recasts backpropagation (and all other error-driven learning algorithms) in a more biologically-palatable form, but still requires discrete phases of "prediction" and "error signal" to accomplish learning.
A variety of other computational models, as described in a wonderful new paper by Kveraga, Ghuman & Bar, propose similar mechanisms in the domain of visual processing. According to these models, object recognition is facilitated by "top-down" predictions of what the eyes will see next - a streamlining process which not only limits the potential problem space of object recognition, but also allows for momentary experience to be used as an error signal in error-driven learning.
Kveraga et al review how this basic idea pops up again and again - from Adaptive Resonance Theory, to Ullman's "counter streams" framework, to Mumford's bidirectional recurrent connectivity framework, and recent Bayesian work by Friston et al. (There are countless more examples, including Churchland's new work as well as Jeff Hawkins' "memory prediction" framework.)
An interesting feature of the visual system is it's asymmetry: lower visual processing areas (V1) project only indirectly to the apex of the processing stream (IT), but higher-order areas do project directly back to the lower areas (V1). Kveraga et al. discuss how this asymmetry suggests that higher-order regions of the brain are "predicting" what they might see next, by biasing early visual cortex in that direction.
However, there's an interesting discrepancy between this "top-down" view of the visual hierarchy and a characteristic of known error-driven learning algorithms: error signals tend to become diluted as they pass through multiple layers of an artificial network, and so direct connections can be required to backpropagate a strong error signal. Thus, the most parsimonious explanation of the asymmetry in connectivity is that object recognition areas (IT) provide a corrective signal to primary visual cortex (V1) in order to tune-up feature detectors in those areas. Thus, these two pieces of evidence may be more compatible with "top-down" correction rather than "top-down" prediction.
On the other hand, Kveraga et al do review a variety of evidence showing activity in orbitofrontal regions (OFC) preceding that in object recognition (IT) and primary visual areas (V1) by up to 50-85msec - as though OFC is actually accomplishing the "top-down prediction" advocated by computational models (and curiously doing so via information from the magnocellular stream).
- Log in to post comments
Another nice post. Thanks for the pointer to the Kveraga et al. paper.
I've been thinking a lot about learning causality and correlation recently, so this stuff interests me quite a bit.
However, the smell is not the only input. Consider
Before it rains: [smell of rain] + [dry ground]
After it rains: [smell of rain] + [wet ground]
So one might learn a causal relationship:
[smell of rain + dry ground] ---> [rain]
One stimulus temporally preceding another within a given time window with statistical regularity is probably a sufficient definition of causality for me. I've tiptoed into some of the philosophy of causation, and have seen things like counterfactual definitions and notions of force. But you point out that with STDP all that is required is the correct temporal order of inputs within a given time window.
I'm thinking that in terms of sequence learning, it may be useful to talk about one moment in a visual array "causing" the next to occur, one word in a sentence "causing" the next, or one note in a song "causing" the next.
Causality is overrated. It is sufficient to model temporal sequence to get an advantage over agents that don't do this modeling. This advantage provided the evolutionary pressure to develop brains. Sub-logical prediction ("intuition") is the most important low level primitive in the brain and is biologically plausible, as opposed to logic, mathematics, or any other mechanism that insists on correctness or discovery of causality for its operation. Causality is "science", but brains are "sub-scientific".
Error backpropagation is also biologically implausible and quite unnecessary.
For more detail go to http://artificial-intuition.com or look for my talk on Google Video.
- Monica
Monica, as discussed in the post, not all forms of backprop are implausible - see GeneRec. I also fail to see how intuition and causality are at odds - we have intuitions about causality, right?
And of course it's better to model temporal sequence than not to. The point is that temporal precedence is not always sufficient to establish causality, and that disambiguating causality does carry explanatory weight in psychology (e.g., chidren's exploratory play, as described in the post).
how else could the brain disentangle mere correlations from true causation?
Ummm... not very well? That does seem to be a basic weakness for both animals and humans, q.v. "superstition". (Humans are of course better at it, but even we sometimes have trouble!)
If cognition is downloaded on our genes before gestation, or is part of the genes at conception, then intuition and precognition are islands of in-sight that are secreted from non-conscious subliminal signals that we encounter and that effect us to our unknowing reception and reaction.
Interesting.
I have thought a while on this - see my blogspot if any of the following is of interest.
http://positonicspage.blogspot.com/
For myself I see intelligence as an elaboration of the mechanisms of perception - which must have preceded it in evolutionary time.
By perception I mean something like Monica's intuition -terminology is slippery here.
I think that the mechanisms of perception must involve an ecosystem of replicating virtual entities - which I call positons - within the mind/brain that feed on assonance to a prediction and that a food-chain of such entities consitute a mechanism whereby the mind can process the concrete but meaningless into the abstract and meaningful.
So this is a somewhat Darwinian picture of the inside of your head.
PP