How might neural networks use a symbolic A.I approach?

Question

I’m fairly new to this area so my understanding is pretty basic. Am I correct in saying that symbolic a.i would take data and use logic to search through and find the right answer. An example is a doctor machine. We feed the machine a ton of medical textbooks and it answers questions by looking up the answers from those textbooks. and The non symbolic approach is feeding raw information into the AI so that it can analyze and construct its own implicit knowledge about how to answer ?

But what I can’t understand is these approaches look completely different, so what similarities might it share ?

Answer 1

Yes, the symbolic AI operates on symbols, but your idea surpasses what it really can do. It was mainly used for the expert system and the input was in the form of symbols, not just any data.

To feed the textbooks into the AI would require Natural Language Processing at higher level than currently is achievable. The symbolic AI concept was created about seventy years ago, so it cannot operate as you imagine.

On the other hand feeding appropriate data into network to give prediction could be used, but here inherent problem is with lack of full medical knowledge to make it work.

If the goal is to check textbooks and return relevant information then the task is simpler, here dictionary based search would do the trick.

No, the raw data is being fed to NN to solve some optimization / classification / estimation problem based on provided data (unsupervised or supervised learning on the training set).

It doesn't create implicit knowledge, think about it more in the terms of regression or approximation. Currently it is not fully known what exactly is the state of the knowledge of trained network, we are not sure what it has learned.

If you compare the expert system and trained classifier, they do not have similarities. The expert system deduce based on given data, requests missing info, input is symbolic and extensible. For raw data approaches the network is not extensible in this sense, it requires new training samples and possibly new architecture if the task differs from the initial goal.

Answer 2

I haven't studied ai formally, but I can tell you that the two things you are comparing are actually examples of not simbolic ai, both of them. That is a medical textbook is actually considered "raw data" as you said. An example of simbolic ai would be a software that gets as input illneses symptoms from the doctor in a particular format that the machine can understand and then uses logical rules of inference to try to find the best diagnostic from the patient.

A no simbolic ai algorithm could be a neural network that looks at pictures of cats and other things and eventually learns to find cats in pictures that it hasn't seen before

As it happens with most things in AI, these definitions are not set in stone and there is lots of overlap between the two things

Answer 3

Symbolic AI covers a lot of different methods, including logic reasoning, but not limited to. The "doctor machine" you refer too has actually been tried in the 1980's, in the form of expert systems: broadly, a very large if-then-else tree. Problem is: it is very expensive (as in: time consuming) to turn expert knowledge into a program (you cannot feed medical textbooks, you have to extract knowledge from the expert). Symbolic AI is still around today, but not so popular as it was in the early decades of AI.

Sub-symbolic AI, which includes artificial neural networks, takes a different approach by training what is essentially a blackbox (e.g. a big mess of artificial neurons connected together). On the one hand, you may gain the ability to feed a lot of raw information to your learning blackbox. On the other hand, you may also loose the ability to explain what has been learned.

For example: a neural network can be trained to associate a description of symptoms with diagnosis (this is supervised learning with labelled data) and be efficient at playing doctor, but it will be very difficult to understand how it produces a specific diagnosis. In the same line of thought: Deepmind's Alpha Go or Alpha Zero can beat the best human player, but both Deepmind's programmers and the greatest masters struggle to understand how.