> we can progressively increase the numerical values of the weights—eventually in some sense “blowing the mind” of the network (and going a bit “psychedelic” in the process)
I wonder if there's a more exact analog of the action of psychedelics on the brain that could be performed on generative models?
Andrew Gallimore's online course on psychedelic neurochemistry[1] argues that psychedelics disrupt something like a constant fine-tuning cycle taking place in normal perceptual activity, causing new activations between cortical structures where connections had previously been dormant (see [2] for an in-at-the-deep-end account), causing the generative model in the higher layers of the cortex[3] to adjust in an effort to regain predictive efficacy.
I wonder if some such kind of "dynamic" generative mechanism might be needed to perturb an artificial model analogously?
I wonder if there's a more exact analog of the action of psychedelics on the brain that could be performed on generative models?
Andrew Gallimore's online course on psychedelic neurochemistry[1] argues that psychedelics disrupt something like a constant fine-tuning cycle taking place in normal perceptual activity, causing new activations between cortical structures where connections had previously been dormant (see [2] for an in-at-the-deep-end account), causing the generative model in the higher layers of the cortex[3] to adjust in an effort to regain predictive efficacy.
I wonder if some such kind of "dynamic" generative mechanism might be needed to perturb an artificial model analogously?
1: https://www.youtube.com/playlist?list=PLbqdD4EM-aEfmLvbWu8GQ...
2: https://www.youtube.com/watch?v=L45A6XlPRM0&list=PLbqdD4EM-a...
3: https://en.wikipedia.org/wiki/Predictive_coding