Modelling individual aesthetic judgements over time

All right, hello and welcome everyone. It is July 8th, 2024, and we're in Active Inference Art Stream number 2.1, A Computational Model of Aesthetic Value with Anne E. Bremen, and we'll have a presentation followed by a discussion. So thank you very much for joining. To you for the presentation and looking forward to it. Thanks so much for having me, Daniel. I'm really, really thrilled to talk about a computational model of something that a lot of people still doubt that machine models can never attempt to model, namely aesthetic value. And that's been the work I've been doing for a couple of years. And today I don't necessarily want to talk so much about the concrete contents of this model, which I do a lot of my papers and it's just as important. I think what's also important and often gets left out a bit from the conversation is the question of why we should even bother developing and applying a computational model to something so complex as aesthetic valuation. So this is going to be a really high level overview and I'd be really stoked if we could have a discussion about this because that's really what I'm trying to investigate here. So stuff that I want to talk about aesthetic values, grandiose words with a rather intuitive, but also quite fuzzy meaning. What I'm talking about personally, when I talk about aesthetic value, something actually quite mundane. It's the pleasure that we derive from any sort of sensory experiences. And you all have felt that value and acted according to it. When you decided to go to a concert or listen to a song, when you went out into the mountains or any sort of landscape and just gazed at it. When you went out to a restaurant and decided what you want to eat and savored that taste. Or even when you touched something or someone and it just felt so good, you wanted that touch to linger. All these experiences, the senses of pleasure that are based on the sensory properties of the objects to interact with, I personally call aesthetic values. And my mission for a couple of years has now been to find a computational model that can tell us something about how these values emerge and what they do to our behavior. And to answer that question, the most important foundation to me personally is to ask the inherently computational question, if we take it to the original meaning of the word computational, of what this so-called aesthetic value is there for. What does it do? Why do we have an emotional effective reaction towards sensory properties? And in the past, people have come up with two very different families of theories to explain these origins. One of them, I will ensure that fluency theories are very much about what I would call lazy brain, saying that aesthetic value actually indicates to us that we are easily able to process the sensory stimuli. And our brain likes that because it needs to do less work, in a sense, to simplify it horribly. And examples where that has shown up in the data is when we look at symmetric stimuli, so only one half has to be processed, or when we look at the effects that we get from familiarity or exposure. So the more often we see something, the more we tend to like it, and that might, according to fluency, be because we get to be able to process it more easily, becoming more familiar with it. And this is very much an evaluation of the present moment. How good am I right now in dealing with my sensory environment? This theory has very often been criticised for being overly simplistic and not very much looking into the future, because it's so grounded in the present moment. And so it came at different friendly theories that I will summarise under an umbrella term of learning theories, that very much rest on the assumption that aesthetic value is actually a signal for learning. So not so much present moment adaptation, but rather a look into the future, gauging how well the current sensory environment can serve me in learning about the future sensory environment.…