I didn't mention variation in brightness though. Say you have an image of a cat upside down, and is darker than have seen last time! The same trick handles this, we start off small features and see both the times they should occur and the difference between both's brightness are not far off. For single pixel in layer 1 it is just 1 pixel comparison.
So in conclusion, continuous needs to be quantized, We can get slight variations in the arrangement of parts, like horizontal flip, rotation, scale, stretch, brightness. All these new many inputs are actually just the same thing, that's how we recognize them. If text prediction was discrete, there wouldn't be any problem, but there is. Rearranged words is as continuous as change in brightness because relatively there is no difference so much, each feature of the sentence or image is only a little off from what features it should be beside or away from in brightness. In life there's these sad inputs that are brighter, or have rearranged words just a bit, but we can recognize them. ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T181e21f4aba061f3-M554be66a9bee0badc1322099 Delivery options: https://agi.topicbox.com/groups/agi/subscription
