I’ve been trying to understand pleasure recently. I mean, subjectively we all know what it is, but I’ve been trying to figure out how it works. It’s kind of an esoteric-but-fascinating question, so I figured there might be somebody who has already obsessed over it and could maybe answer this question before I start obsessing myself.
Of course I’m already obsessing; that’s my nature.
I’ve been reading about brain physiology and neuroanatomy and dopamine uptake and neurotransmitters and varying degrees of resistance to alteration of activation thresholds and long-term and short-term neural plasticity which appear to operate on different principles, and about Hebbian theory and experiments on Hippocampuses (Hippocampi?) and in vitro and in vivo experiments with neurons…. and watched some truly disturbing videos starring wired-up rats, and read the associated papers … and I feel like I may be no closer to an answer yet.
It’s a disarmingly simple question.
What is the operating principle of pleasure?
What is it about the experience of enjoyment that results in the individual synaptic connections getting modified in a way that leads the experiencer to repeat the activity that brought them joy?
I’m familiar with the model of neural nets used in computer simulations, and until recently they’ve all been about minimizing a total error function. That is, every bit of correction they got was ‘negative’ feedback, about how much of a mistake they made and how to do better next time.
Recently we’ve made some advances in computer simulations allowing training of very deep networks by using autoencoding to build representational models, so that’s another basic principle that’s useful. Google’s Deep-Dream stuff used that technique heavily and the advance in methodology was really big news to me; I knew what there was to know about computer neural networks years ago but they did something I’d have sworn was impossible using a new principle that got discovered while I wasn’t paying attention. So I devoured that research and maybe I’ve caught up to knowing most of what there is to know about software neural networks again.
But our set of principles is still incomplete. We’ve got negative feedback which sort-of models pain, and we’ve got a representation building principle which sort-of models understanding, but …
So far there’s no such thing in our operating principles as positive feedback, something that would model pleasure. How do we train in response to a good outcome that’s analogous to “Oh, that went RIGHT, I want to repeat that behavior or experience that sensation again.” And the more I obsess and read neurological papers, the more I wonder if ANYBODY out there really understands how it works in biological brains.
However such a mechanism would work, I don’t think it’s one we use or even attempt to use in computer simulations – and understanding it is important. Aside from opening possible new capabilities for software neural networks, we really need to understand how this happens in our grandly more complex wetware neural networks.
Anybody got any relevant facts?
Anybody got any decent theories?
Anybody got a crazy idea for a potential model or experiment or something for me to see if I can figure out the math to make it work?
When you say negative feedback, and positive feedback, punishment and reward might fit better (though minimizing error really is negative feedback).
As far as I understand it the function of pleasure is “Do this a lot more” (and do other activities a little less). As long as artificial neural networks do not have a choice between activities so there is no no point to reward. As soon as you augment them with “attention” so they can choose what things to pay attention to, reward will be useful because we need to say “Yes pay attention to that”. The part that reward is meaningful to is the activity selector.
(I get the feeling that I might not understand the question, because you already know this better than me.)
Wow. I think maybe you got it exactly right, at least for a software model. Thank you! I’d been thinking in terms of finding a counterpart to the negative-feedback we use in training, but that was the wrong model. The idea of activity selection is actually a much better model for the role that ‘pain’ and ‘pleasure’ play in practice.
Of course this presupposes a feedback function which varies depending on how well each of a set of needs have been met; when we’re full, we stop getting pleasure from food – and hunger (error due to unmet need) rises again gradually until an activity resulting in satisfying it becomes enjoyable again.
And activity selection presupposes a variety of such needs – such that choosing which activity to do (or which to avoid) becomes a matter of allocating time and attention.
The more I think about this, the more I like it.
Well, off to a different round of obsession. And again, Thank You! That’s probably the most constructive and insightful response to this question that I’ve even considered, because I was diving in at the wrong level and sort-of asking the wrong question.