On Thursday, January 07, 2021, at 2:29 PM, Matt Mahoney wrote: > No. Humans can see because of decades of training, a petabyte through our > optic nerves. Even then we are born knowing how to recognize or learn to > recognize things important to our survival. Things like faces and animals, > vs. barcodes or watermarks. Yes more images help but this is 1-shot learning.
Modern approaches rely on Data Augmentation, but for it to be truly helpful you'd have to store an enormous amount of variations. And I believe we store actually none for variations unless seen/thought of. Only syntax, and maybe semantic (maybe BY syntax). Vision nodes have many pixels, there may be ex. 10,000 vision nodes hence semantics only links top 1,000 but each 10,000 nodes each can variate thousands of times. Still thinking. Other approaches are complex and still not human level. https://www.cs.cmu.edu/~rsalakhu/papers/oneshot1.pdf ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T751e544d2713dc23-M744bb9cf9e80bfea5026fc23 Delivery options: https://agi.topicbox.com/groups/agi/subscription
