1
$\begingroup$

I want to make an AI with deep learning which can adapt itself from user to user.

Let's say we have food combiner AI which suggests a food to eat with another food as you give as input. This is the most personalized case I found to ask here. For example the AI suggested a food for me. However, the food AI suggested for me might not be good choice for another person. So another person will let the AI know like "I don't like that food to eat with this. Etc. When the user let the AI know that, It should affect AI's further combination food suggestions.

How can I build that AI? Where should I start from? which area or topics should I research?

$\endgroup$
  • $\begingroup$ Maybe have a look at active learning and online machine learning, but maybe this isn't exactly what you are looking for. Maybe you want some kind of AI which is aware of different users. For example, have a look at recommender systems. $\endgroup$ – nbro Mar 13 '19 at 20:53
  • $\begingroup$ This link might help. $\endgroup$ – Ugnes Mar 13 '19 at 21:07
1
$\begingroup$

Assuming you have enough training data and representational capacity, you can give each user a unique identifier and concatenate that with the other inputs to the neural net. As users give more feedback, the network will learn in the usual way since each situation (e.g. food-user pair) is represented by a unique input.

I'd consider this a general, brute-force approach, and it may not be suited to your application. Others' links to recommendation systems might be more useful. It depends on what your task is and what your constraints are.

| improve this answer | |
$\endgroup$
  • $\begingroup$ Firstly thx for your reply.I think it would be really huge cost that I can't afford to concatenate each users information with other inputs and also it wouldnt work well for a new user, because those who affected the deep learning model before,would also affect the new user's combinations. So It might be produce combinations weirdly.If I had real statistics of what the user liked or did not and encode this info to an array with an algorithm,your method would be really useful.However, I have only images in my real situation(not food comb).So I can't encode how the user likes or not ina good way $\endgroup$ – Faruk Nane Mar 17 '19 at 14:16
  • $\begingroup$ I have a new idea. what if I do the AI thing in users' phones? So I can ask them a few pictures (combinations) if they liked or not. Then I take those and give to the Deep learning model. So each users can have different model. However, there is a big concern that I ve been thinking that if I do that, I wont be able to update the model via remotely. For example, I improved the model by training a lot data, but users have their own models and they are old models. So I want to change my new model with their old models. I just cant do that. $\endgroup$ – Faruk Nane Mar 17 '19 at 14:24
  • $\begingroup$ What I can do is training each model in each user's phone with "my new big data" and this would be really bad idea. So I got stuck $\endgroup$ – Faruk Nane Mar 17 '19 at 14:25

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.