This paper shows that some human "gestalt" style visual processing can emerge in purely feedforward deep nets with the right inductive biases and fine-tuning: journals.plos.org/ploscompbiol... 🧠📈 🧪 #NeuroAI
Replies
It seems plausible that V1-like representations could emerge via training given that we saw something similar with predictive coding (Rao & Ballard, 1999), and the fact that oriented Gabor detectors are common in early layers of D/CNNs. Fine tuning on known physiology may regularize the phenomenon.
aug 26, 2025, 6:52 pm • 1 0 • view
Stu Geman has an old theory paper (binding-by-synchrony) on ventral visual stream. The DL literature is interesting if it can give insights into feedforward common input model. www.sciencedirect.com/science/arti...
aug 26, 2025, 5:55 pm • 1 0 • view
Sounds similar to the ideas in the 1992 Hummel & Biederman JIM model
aug 26, 2025, 6:46 pm • 2 0 • view
Thanks!
aug 26, 2025, 6:49 pm • 0 0 • view