Deep Dreaming Aloud

Extremely popular misconception. Wrote a decent comment explainer. #################### Commenting on https://www.bloomberg.com/opinion/articles/2020-10-15/essential-part-of-obamacare-needs-expanding   "if the companies can avoid anyone who might get sick, insurance ceases to be insurance" I find it appalling how 99% of commentators (even presumably math savvy ones like Cathy) don't distinguish insurance from redistribution. So let me recap. Insurance (in a narrow sense) is about dealing with unpredictable risks: if any one of our cars is equally likely to get into accident costing $10K in repairs and the chance for each of us getting into accident is 1%/year, we can each pay 1%*$10K = $100 (+tiny insurance company profit)/year in insurance for the peace of mind. Note $100 is exactly the expected yearly loss from car accident for each of us, so it's not like we're getting a bonanza of free or cheaper repairs: we are still paying our expected costs, but get valuab...

Data augmentation is crucial for our networks learning using relatively little data. But what about animal brains? Naturally, animals learn from video, while moving, while shaking at the micro-scale. What if we fixed their bodies to avoid the shaking? What if we fix their bodies and only show them pictures, and never anything moving? Would they be able to learn to see? How much longer will it take? We cannot measure individual neuron activity, but maybe we can understand animal brain learning better by presenting animal brains with much more limited training data, making it pictures instead of videos, and possibly limited class of pictures. We can combine various classes with positive/negative reinforcement (food, electric shocks) so that correctly classified pictures elicit sufficiently strong response from the brain to be measured.