| I have thought of Personality Vectors similar to this concept but for a completely different purpose. If we had a Chatbot that would interact with someone hypothetically we could have this kind of distinction between Personality Vectors to mutate its own behavior and to interact with others. Its sort of the idea I had when I watched the movie Interstellar and TARS is told to decrease his own humor. So if the system would just know who its talking to it would create a preference vector for each person it interacts with. But you would need some kind of baseline so basically the Personality Vectors would be set to neutral. And each interaction with a Personality vector would be a system that stores a preference vector for each person. As the system time evolves (which is also in the article) you would basically end up with a set of graphs. There is a notion out there which is called a Dynamic Network analysis [1]. If you also stored the history of that you could then not just be able to rewind the system backwards to a previous checkpoint. The way you could start learning how to learn a personality vector is a reinforcement learning situation. If you say "bad AI I didn't like that" the system would then look at what it said and mutate the preference vector for that person. Or you could just change the vector manually by a technique that was illustrated above by just saying "reduce humor by 70%". Then you could have model which would do conditional generation [1] of text based upon interactions with others. It would be basically a reinforcement learning system paired with a generative text model. The reinforcement system would store the Personality and Preference vectors for each person. For each behavior and preference you would need a corpus to bootstrap the system. The reinforcement system would be a retrieval network and it would retrieve a pertained system based on user(s) input and get a generational text network to provide a response. This would be similar to Alpha Go's design. [3] [1] https://en.wikipedia.org/wiki/Dynamic_network_analysis
[2] https://github.com/salesforce/ctrl
[3] https://datascience.stackexchange.com/questions/10932/differ... |