Let’s talk about whether or not machine learning is necessary to achieve AI.
This debate has been going on for years with little resolution. There are two reasons for this.
First, people have different definitions of what constitutes “machine learning”. Second, people don’t know how to assess machine learning algorithms versus other methods.
Let's first address the definition of “machine learning”. It is very easy today to produce data sets using entirely manual processes. However, it takes a skilled human to create the training set, the software design, and the process used to extract insights from the data.
There are many examples of companies that pay humans large sums of money to do these tasks. What they are paying people for is skill. Why would anyone expect a robot to have such skills at cheap prices?
Companies like AutoSumer make money by selling marketing lists. The more emails and names she gets, the more valuable her sales leads are. Does she need machine learning knowledge to organize her email messages effectively? No, she needs ability to read an email and put it into one folder or file. This is something any good employee should be able to do.
Although there is no real definition of artificial intelligence (AI), most people reference it as something that can learn from its environment and self-improve.
In our era, we can give computers far more information than they could ever consume. It becomes difficult to define AI in these circumstances because the machine might still choose an optimal solution even without interpreting the information in terms of intent or purpose.
For example, say you are teaching a computer how to swim. You provide instructions like “stand with your feet together” or “step on the edge of the pool.” The computer then performs each instruction until it learns to swim.
After it learns to swim, would it therefore be considered intelligent? Would it understand what you were trying to teach it? Or would it simply repeat actions to achieve a result?
We cannot just assume that the ability to learn defines AI; maybe learning is only a start towards understanding human intuition. In fact, research suggests that neural networks, which are used in almost all modern NN algorithms, develop meta-knowledge when exposed to large amounts of data.
This indicates that humans possess internal logic within their minds that helps them make decisions. An algorithm designed using this type of NN will not feel fatigue after working for hours because it doesn’t interpret information into concepts.
Though machine learning is an excellent algorithm for training, it is not necessary for creating artificial intelligence (AI).
There are several types of AIs that do not use machine learning : rule-based systems and expert systems. These two types of algorithms rely on coding explicit instructions to perform a task.
Rule-based systems are used in workflows where tasks are assigned rules or steps. For example, you could have a system with “if product X comes into contact with product Y, then trigger step Z.”
Expert systems require human knowledge of their domain to create a set of rules like those above. You can think about this type as something between fully automated machine learning AI and explicitly coded instruction sets.