MesoNet: predictive coding and graph-search questions

predictive coding tangent—what is the neuron actually predicting, and over what time window? Dijkstra vs A* too, need a heuristic + a way to keep distances useful while weights change

learning ideas: predictive coding and contrastive/self-supervised objectives as possible ways to form representations from unlabeled inputs. The notes leave open how to express the objective through spike-based, locally available signals.

graph work: compare Dijkstra and A for paths through network connectivity and avoiding redundant connections. An A route requires a meaningful heuristic; the entry does not specify one or demonstrate a speedup.

tangent: the notes also collect Fisher-information material. Treat it as a reading lead until a specific estimation quantity and model are defined.

References saved in this journal section

• https://www.v7labs.com/blog/self-supervised-learning-guide
• https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2022.1079357/full
• https://iopscience.iop.org/article/10.1088/2634-4386/ad3a95/pdf
• https://en.wikipedia.org/wiki/Fisher_information
• https://en.wikipedia.org/wiki/Multilayer_perceptron

• review questions: what target does a neuron predict, over what horizon, and how is prediction error represented? For graph search, can cached distances be reused while weights change, and how is their validity checked? Keep “a path-search method runs” separate from “the resulting learning rule improves representation.”

Graph-credit design notes