How is AI different from ML?

This is an excerpt from my ML course notes during my M.Tech.

Courtesy — Prof. Srinivasa Raghavan.

One way to look at AI is to build human look-alike and “intelligent” machines/robots. AI as used was practically top-down i.e., you make an algorithm for the machine to follow and expect it to behave intelligently.

Let’s take an example of Expert Systems:

An expert system is nothing but a rule engine that is supposed to help experts in their own domains. A rule engine is basically first-order logic (boolean algebra or predicate logic). We have a bunch of axioms and then for the reasoning to happen, we need inference rules which are also considered as self-evident as axioms. These two constitute making a ‘theory’.

What is a proof: when you arrive at a chain A->B->C->……->T where each of the arrows is inference rules and each alphabet is an axiom. This chain need not be straight. Finally, we arrive at a theorem (T)/ statement which is true under all axioms and inference rules.

Every statement with proof is a theorem.

Note: In first-order logic, we have a symbol for every statement which can be classified as True or False. Symbols and operators (and, or, not, implication) together form an expression. That expression in an expert system is called a rule.

Now, these “symbols” values are needed to check the validity of the expression. In first-order logic these “symbols” may not have a fixed value, they can be variable.

P(n) = (for all) m : (m=1 V m=n V n%m != 0)

To represent: “Two is the smallest prime number”:

P(m) = m is prime, => (for all)m (~P(m) V m>=2)

In expert systems, everything about the world can be represented in first-order logic. But an inference rule can be a dumb symbolic operation because we don’t need to know what A, B is to evaluate ‘A->B’ because if A is true, B will be true is the basic semantic of the inference rule. That is why to verify a proof, it’s a mechanical process and also it is meant to be.

How do Expert Systems work?

Let’s take the example of an expert system for a doctor. So I will make rules according to how doctors diagnose a patient according to symptoms and store all the rules. Now when a new patient comes with A, B, C as symptoms and I say he may have a disease D.

So I will plug in ‘A and B and C -> D’ as a theorem into the expert system and ask to prove this theorem using the stored results which we have.

This is the classical way of doing AI which is why we call it top-down.

In AI we tell each and everything (no matter how small it is) to the machine to do in certain situations, and in the other extreme is let the machine decide what to do in any situation and we don’t say anything to it, which is considered bottom-up (ML).

We need a blend of these to make things work, hence we provide both data and algorithm to the machine and learn things using both. We cannot rely entirely on data, so we do Knowledge Management — encapsulate, formalize, acceptable, and “of use” data for the machine/algorithm.

Data a few decades ago was archival data, but data today is real-time and dynamic, i.e. why now our algorithm should adapt to the dynamic data because data drives decision making.