The hippocampus as a 'predictive map'
deepmind.com
deepmind.com
A compressed sensing perspective of hippocampal function (2014): Hippocampus is one of the most important information processing units in the brain. Input from the cortex passes through convergent axon pathways to the downstream hippocampal subregions and, after being appropriately processed, is fanned out back to the cortex. Here, we review evidence of the hypothesis that information flow and processing in the hippocampus complies with the principles of Compressed Sensing (CS). The CS theory comprises a mathematical framework that describes how and under which conditions, restricted sampling of information (data set) can lead to condensed, yet concise, forms of the initial, subsampled information entity (i.e., of the original data set). In this work, hippocampus related regions and their respective circuitry are presented as a CS-based system whose different components collaborate to realize efficient memory encoding and decoding processes. This proposition introduces a unifying mathematical framework for hippocampal function and opens new avenues for exploring coding and decoding strategies in the brain.
What about the neurotransmitters like dopamine or serotonine? Could they (or more specific: their effects) be key to an AI that feels more “natural“?
Take a look at this brain schematic here [0]. I vehemently disagree with some of the presented-as-fact ideological aspects of the article, but the depiction of neurotransmitters working differently in different brain regions is illustrative.
i.e., a dopamine molecule does not already constitute a meaningful coordinate system -- it doesn't have a meaning in and of itself, such as "the reward molecule", or euphoria/"feel-good"/"addiction trap"/excitement molecule) in the brain. Rather, the dopamine molecule has locally-interpreted semantics.
To make this more concrete with examples of different aspects of dopamine alone:
- "dopamine is known to encode the confidence level of motor predictions" - Scott Alexander, psychiatrist [1]
- saturating the synaptic cleft between neurons with released (or non-recycled) dopamine is a part some euphoric processes (e.g. drugs, orgasm, gambling) in the striatum negra
- 'excess' (not a simple thing to determine for individuals imo) dopamine is related with psychosis (not per se causally), and antipsychotics are often designed to blunt dopamine transmission
However,
- dopamine transmission is also fundamentally necessary for muscle movement, in the motor cortex
- additionally, the conscious sensation of dopamine is highly dependent upon the frequency of dopaminergic neuron spikes
- Parkinson's disease is (iirc, said to be,) the result of the degeneration of motor neurons; shakiness, problems sleeping;
What does that mean for AI? Well, if we want to go beyond network topology, and start to use new varieties of component substrates, the individual types of signals they represent must either a) be fairly meaningless outside of a specific context or b) not be suited for learning from evidence and generalization of data. Check out [2], [3], and [4] for more.
[0] http://www.nationalgeographic.com/magazine/2017/09/the-addic...
[1] http://slatestarcodex.com/2017/09/12/toward-a-predictive-the...
[2] https://www.edge.org/conversation/stephen_wolfram-ai-the-fut... (not that i am a huge fan of wolfram :P)
[3] https://www.edge.org/responses/what-scientific-concept-would..., ctrl-f "biological"
[4] https://en.wikipedia.org/wiki/Multi-state_modeling_of_biomol...
[0] I did my PhD on this
[1] https://en.wikipedia.org/wiki/Parkinson%27s_disease#Pathophy...
"Reading about rewards in AI context makes me wonder, why current AI only focuses on the network topology aspect of natural intelligence."
Well, it doesn't. Not at all. It looks as if you conflate one very specific AI techique (deep neural networks) with all of AI. Because that one technique is currently insanely popular and has been widely mainstream reported and hyped that makes it look as if your education in that field did not go beyond reading a couple of mainstream media articles. Also, there's a whole huge subfield in AI which is actually defined by it's focus on learning from rewards, which you seem not to be aware of.
"What about the neurotransmitters like dopamine or serotonine? Could they (or more specific: their effects) be key to an AI that feels more “natural“?"
Not applicable as written, too broad and unspecific. "Not even wrong." Looks like someone with superficial knowledge is just throwing a few sticks into the blue in the hope that someone might be aroused to a reply which competently tries to make actual sense out of it.
The topic reminds me of a book I read (and was fascinated by at the time), The Muse in the Machine by David Gelernter. It seems he has moved away from AI research and into political punditry, however, and recently made headlines denying that the moon landing ever happened. Nevertheless, that book contained a lot of exposition around the idea of creating AI with simulated emotion. This is simultaneously in line with the idea of simulating neurotransmitter function, but also has parallels with more general mathematical techniques that are in widespread use. Thus, the book (and perhaps the idea itself) are a bit dated, but I don't think it makes them less relevant.
Still, please note that the critiqued comment did not even go as far as to even suggest an idea about the mechanisms involved beyond (implicitly, not even explicitly) "somewhat like in the brain". There's just not enough meat there to even do a proper critique, and on this basis I pulled the "Not even wrong"-card, which is of course at bit harsh, but that was the whole point about trying to put myself into the position of a downvoter.
It really isn't; what's reasonable is to always assume whatever you thought of in 10 seconds, has been deeply researched regardless of whether you're aware of it or not. What's reasonable is to assume researchers are vast leagues ahead of whatever shallow thoughts you have on a subject since you "know" you have a shallow understanding. The reasonable person simply asks for links to read deeper about a subject, they don't criticize something they have a shallow understanding of and ask why no one is doing X.
Of course I made the mistake to use “only“ where it should have been “so much“ or something. Sorry again.
Inescapably, much of the high level ideas in ANN-based machine learning ("connect a bunch of nodes", "integrate and fire", "convolutional neurons/ receptive field") are cribbed directly from neuroscience.
Spiking Neural Networks (see SpiNNaker) and Hierarchical Temporal Memory (see numenta/nupic) aim for more biological plausibility, directly.
*ANNs == artificial neural networks
Historically researchers looked at activation functions that are at least somewhat justified by nature, like the sigmoid. More recently people have focused on the mathematical aspects and figured out that the rectified liner unit `ReLU(x) = max(x, 0)` works just as well if not better for most applications.
Maxout networks (https://arxiv.org/pdf/1302.4389.pdf) are another interesting development in that direction.
There's still lots we can learn from biology, of course; but the last decades rapid advances we saw from deep learning at most barely relate to real brains.
That happens fairly often in engineering: looking around enough one can often find natural precedents for our artifices. But directly trying to copy nature did not lead to the plane or helicopter.
[1] R Hahnloser, R. Sarpeshkar, M A Mahowald, R. J. Douglas, H.S. Seung (2000). Digital selection and analogue amplification coexist in a cortex-inspired silicon circuit.
Actually, the name is not so misleading because `ReLU(Wx + b)` is an approximation of the response of a large ensemble of integrate-and-fire neurons in terms of the firing rate [1] (it models populations of pyramidal neurons in the cortex). The ReLU activation function basically approximates a logarithmic activation function [2]. Obviously many kinds of computations are left out in this model, for example in pyramidal neurons, the integration of the incoming spikes is often sublinear rather than linear; the branches of the neurons that receive signals (dendrites) can themselves perform complex computations such as spatial clustering and logical operators basically sending small (dendritic) spikes forward and backward within an neuron [3]; short-term plasticity is left out entirely and so are countless types of interneurons which inhibit other neurons nearby and measure mean activity across populations of neurons etc.
[1] https://web.stanford.edu/group/brainsinsilicon/documents/Won...
[2] https://www.utc.fr/~bordesan/dokuwiki/_media/en/glorot10nips...
[3] https://neurophysics.ucsd.edu/courses/physics_171/annurev.ne...
(I only mention ReLU because they are definitely simpler in mathematical structure than ye olde sigmoid. Max out is another interesting activation function that's first and foremost inspired by the math / pragmatics of drop-out training, and were any relations to biology more likely to be subsequent discoveries and not intentional design features.)
This works just as well whether you are modelling biological synapses with plasticity, or compositions of functions with training via backpropagation.
The only real difference is that NNs aren't usually "physically embodied" (i.e. embedded in a metric space with radiative signalling to non-connected but "local" neighbours) in the way they'd have to be to get the "area-effect excitation" effects of neurotransmission.
By comparison, ANNs have many meta-parameters (parameters besides the actual weights), such as learning rate and momentum, affecting stability and performance. Being able to tune these on the fly would likely improve performance.
The "ANNs are not like the brain!" camp like to point out that the ANN is a mathematical model bearing (on one level) little similarity to the human brain, but I think this glosses over the fact that key features (e.g. layers of connections, CNNs) were directly biologically inspired, and we can still glean a lot by studying the best example in our world of a learning system.
Nonetheless, this kind of algorithm sounds like exactly what I expect from a predictive brain that cares fundamentally about trajectories.
* Firstly, you're reading a neuroscience paper. It's biology! Numenta is not biology.
* Secondly, Numenta makes a lot of big claims and then never publishes anything.
This has gotten as if we were responding, "Simpsons did it!" to jokes which, in fact, The Simpsons had never aired, based on the supposition that since they once did a joke with vaguely similar vocabulary, it must have been the same as this joke.
Nvm found it: https://www.biorxiv.org/content/early/2016/12/28/097170
usually I try digging the authors in case preprints are on their own webpage
and just in case people miss it, the latest version is also there https://www.biorxiv.org/content/early/2017/07/27/097170
https://sci-hub.cc/https://www.nature.com/neuro/journal/vaop...
That reminds me, I should donate.