Using AI to interpret the nature a specific input and use the correct model

Question

How would one go about building an AI that is capable to look at any kind of input and then identify what is the nature of this data?

For example, an AI that is able to do image classification, NLP and react to some other sensors. Is it possible to build an AI that will be able to identify what kind of data it is seeing such that it can send the data to the correct model for it to be treated, similarly to how the human brain knows to send visual information to the visual cortex and auditory information elsewhere?

In a simple scenario, I think we can get very good performance by having a cascaded image classifier. For example 2 layers, the first layer identifies if the image contains a dog and a cat. The next layer, has two different CNNs, one trained to identify the breed of dog and the other one for cats. That way once we identify that we have a dog, the image can be sent to the correct CNN. A CNN that is trained specifically to detect the breed, thus being much more robust that a more generalized CNN. Kind of like a professional in the field. First, the human identifies that he is looking at a dog, then he consults a professional to ask him the breed.

I would like to extend this idea to be able to identify various kinds of data sources that do not resemble each other at all. Various input. Are there any models that can do this?

Answer 1

Because of all the inputs you could give are obviously a list of number. It will only require some CNNs and fully-conncted, that classify what kind of data is it and then pass the data to the right stack of layers. The only tiny problem is that all the the things you pass into the network must always be a matrix(to handle also images) with a certain size. So even if you pass only a tiny chunk of data you must transform it into a fixed-size matrix and fill all the empty places with 0s, and that could bring to an accuracy lack.

But to solve that you could also try to slide over your data a tiny CNN-RNN to handle different sizes .