Building AI using ML - Part 2¶
Artificial Intelligence — A Clear and Simple Introduction: Part 2¶
This is the second part of five part introduction series on “Artificial Intelligence”, an attempt to explain the fundamentals in simple language without much technical jargon.
Read the first part here.
What it takes a machine to become an Artificial Intelligence?¶
The honest answer is that we don’t know yet, because we have not made any AI that learn, think and act like human.
Let us understand this statement better.
When classifying tasks, we often assume that some necessitate Human Intelligence. However, this is not always the case.
Researches has revealed that many tasks we consider as requiring intelligence can actually be accomplished through straightforward mathematics or by utilizing lookup tables, especially when we translate perceptions into an entirely different domain.
Think of music. When we graph an audio waveform in the time domain, it appears chaotic and hard to understand. But when we convert it into frequency domain, you can do much more with it.
By extending the analysis of the historical data, and coming up with a relationship between the environmental perception (the context data) and the expected outcome (actuation), it is possible for a machine to perform the given task to a level comparable or sometimes better than human.
Finding the correct domain to represent the data, computing power, and storage requirements for performing such complex tasks, which we thought required intelligence, is the key.
The machine performs the given task excellently. But, we still can’t call it AI, as we have designed and trained the system for very specific task and domain we started with. There is knowledge, but it can’t be transferred from one domain to another or one task to another. These systems are called Weak AI or ANI (Artificial Narrow Intelligence).
There is no system exists today, that can be called AI or “Artificial General Intelligence” to behave like a human. The outcome of research in Artificial General Intelligence or the Real AI is unimpressive.
All we have today are Weak AI or ANI.
Recognizing that Narrow AI as a compute and storage system, which requires programming, can help us better grasp many of the AI systems we encounter today.
What it takes a machine to become an Artificial Narrow Intelligence?¶
Compute power and storage capacity.
There are many tasks that were initially thought as requiring intelligence were demystified and algorithms were developed, much earlier than we could think. However, Implementing them didn’t happen because of unavailability of compute power and storage capacity.
Advances in computer industry made it possible to develop many such system in recent decades.
Another ambiguity we encounter very frequently was about machine learning.
- Is machine learning synonymous to AI?
- Does machine learning a mandatory for AI ?
The answer to both are ‘No’.
AI requires programs and parameters, hence may be implemented either by logic and rules or as a knowledge store with smart lookup ability. Sometimes we need to cascade multiple systems to perform a single task.
There are many such system evolved over period. But the problem was complexity. It was near to impossible to identify all the states (perception, historical perception) and write a program that meets the expectation. Still, there are many attempt to create one.
Once these approaches exhausted and impossible to convert the algorithms into programs, a completely out of box thought, which is much inferior in accuracy, but requires less computation and less effort to create was explored. That is ‘Statistics’.
Instead of computing, what if we predict? So, prediction models were used instead of traditional programming. These models require collecting, organizing, summarizing, and interpreting data.
While this approach simplified the development and reduced the computational complexity and storage, prediction can never be 100% accurate. They fail. To reduce the failure rate, in another word, to improve the accuracy, various methods were explored. One such method that has shown acceptable out come was Machine Learning.
Machine Learning is not statistics, but together both statistical model and Machine learning, the development of Narrow Intelligence Systems was simplified.
It’s important to recognize that while Statistical Models and Machine Learning are widely used, it is possible to design a system that performs better than Statistical Models. It is also feasible to impart knowledge to a machine without relying on Machine Learning. However, such approaches require both additional time and effort.
How the ANI was implemented before Machine learning take over? Let us continue in the next part.