Let us say we have a problem to be solved.
Approach 1: Closed Form Solution
We may use, where possible, a algebraic approach to get a closed form solution. This approach is the most efficient to finding a solution. Unfortunately, may real world problems are not solvable in this manner.
Approach 2: Physical model
We may use physical model based simulation approach to solve our problem. This approach is close to reality but is very expensive and tedious.
Approach 3: Algorithmic approach
We may use a sequential algorithmic approach to find an optimal solution for the problem. But this consumes a lot of time. Image if I were to de-duplicate a list of 40 million name records in 2 hours.
Approach 4: A neural (Conventional) network based approach
We may use a conventional neural network (NN) based approach. The NN based approach is faster than approach 3 but we need special hardware to realize the solution.
Approach 4: A Deep Neural Network (DNN) based approach
We may use one of the many DNN networks, such as, CNN, RNN, Transformers, GAN, etc. in solving out problem.
We can use a GPU / TPU / FPGA and realize the final solution in hardware.
Please note that time taken by Approach 1 is the smallest, Approach 4 time > Approach 3 > Approach 2.
So, finally based on the problem on hand and the quality and speed required, we may choose one of the above approaches.
In every approach narration, I have intentionally used the word “can USE”.
So each of these approaches are an enabler in solving the problem on hand. AI-ML by itself is not a solution to any problem.
