I posted on OpenAI's Slack channel: The article mentions Nick looking for curves in images and couldn't just choose 1 of 4 classifications for a given image, and says "We were surprised when we saw that activations fell naturally into different levels of activation.". Really? I knew that. I actually know something deeper. Check out this image: https://ibb.co/x5J7s9k
If my hierarchy schema reads "cat", it activates the "cat" node including other nodes ex. "cattle" and "tac" and of course parts of itself ex. "at" and "t" to variable degrees. Each node has predictions of what comes next. They combine predictions by shared parent nodes. I have a real algorithm that does this. Energy flows rightwards only, you can't repeat the alphabet backwards naturally. As shown in my image, the "cat" node activated will activate neighboring context nodes, and through these local channels and only through these channels will trigger/discover nodes like "dog" that share the same contexts - cats and dogs both eat, sleep, run, lick, etc. The cat, cattle, dog, etc nodes are activated by variables amounts and all mix predictions. It can recognize unseen sentences plus use many matches/judges for prediction. My hierarchy schema can also discover/trigger my=your if it stores "my dog" and "your cat", because the shared contexts are, while not exact, similar, hence leaking energy still! Further, if both rabbits and horses are dogs, animals, 4 legged, cute, and have 2 eyes, then rabbits triggers horses. This is getting deep now, very parallel, each node leaks energy and each of these then does too... ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/Tf06e133ecd7df7c9-Mf30e1c833474595f72111c99 Delivery options: https://agi.topicbox.com/groups/agi/subscription
