I know I've probably asked this before but we still have a problem. See the image attached. We have a tree at left, or better, a hierarchy on right. When you get a prompt/ sentence, you can find multiple matches of varying lengths in the tree or hierarchy ex. 'th' and 'the' and 'the ' and 'the c' and 'the ca' and 'the cat'. You combine the probabilities to get 1 set of probabilities for each letter that may come next. It gives ok predictions. The weights are just counts on the connections that you update based on accesses made.This algorithm is called Partial Prediction Match (PPM) which is like a markov chain that uses Backoff. It just tells you the probability of the next letter i.e. what it probably is.
So, why would you want to use a *backprop *NN? Gives better results? But why does it give better results? Is it doing semantics!? Or something else? What exactly is making it predict better? You can't just say cuz it does, there is an actual thing going on if it gives better results. I will be back later today. ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/Tdde7634514c58304-M4c14baa41fac8816b03fbac5 Delivery options: https://agi.topicbox.com/groups/agi/subscription
