,
10 tweets,
4 min read
Read on Twitter
Day 1 at #ccn18 was quite stimulating. Short summary of interesting information ahead. All the papers can be found at - ccneuro.org/Papers/Accepte…
@AlisonGopnik argued that the "bugs" of childhood might actually be developmental features. The phase of exploring the world without many priors might be important in rapidly inferring and learning associations.
Ugurcan Mugan argued that long-range vision + sufficiently complex environments are important for the development of high-level planning. The argument was based on a simple predator-prey game where the prey's access to information and kinds of obstacles were manipulated
@thisismyhat presented his work on feedback alignment in training NNs. That the random feedback could be construed as the feedforward pass for another network was interesting. Still have to mull over the potential consequences 🤯
Had an interesting discussion with @KohitijKar about how much the representations in IT cortex are malleable. Apparently, in monkeys, it has been observed that IT representations don't change much, when they learn categories, indicating some downstream area is the decoder 🤷♂️
Aria Wang's work showed that object affordances could be computed in a bottom-up fashion using CNNs. Another step in qualifying the bottom-up computability of factors that are traditionally considered "conceptual".
Had a nice discussion with Ruyuan Zhang about their work on assessing the connection between the usual CNNs and the brain by considering adversarial examples. Adv. examples might provide windows into the nature of the features the network cares about in performing it's task.
Andrew Zaharia presented a cool, and seemingly informative, technique for visualizing representational spaces by establishing spherical categorical bounds and then reducing the dimensionality of the distances between the hyoerspheres. The results look beautiful. 
@rockNroll87q presented preliminary results about the nature of "predictive" feedback to EVC. It's interesting that the "inpainting" decoder's features are better matched to EVC features in the non-occluded region! 7T might shed more/better light (by assessing cortical layers)
