Written by Retalon CEO, Mark Krupnik, originally published in Entrepreneur.
The commonly accepted definition of artificial intelligence (AI) is the capability of a computer to decide without human intervention. Let’s think about this for a moment: If I taught a computer that 2 plus 2 equals 4, then from that moment on this computer will be able to answer the question “How much is 2 plus 2?” without human intervention.
Does this mean it has AI capability? Well sort of, but it’s a bit like cheating because human intervention did take place, just in the past. The computer is only able to solve the specific questions it was instructed to answer.
Humans have spent the last 50 years providing computers with millions of instructions in the form of algorithms. As a result, today computers can act in numerous situations that are well-defined, such as calculating the trajectory of a spaceship, tax returns, accounting balances and so on.
So, is it AI? Formally yes, but that’s not really what people mean when they say AI these days.
The true magic everyone is talking about is the ability of a computer to learn how to think and solve the problem on its own — without being provided specific instructions.
This is the most intriguing part of AI, called Machine Learning (ML).
Just like people, computers need to be trained on how to learn new things and act in new situations. Similar to the human brain, the computer neural net needs to be trained to use some optimal path when making a decision. This is called training.
How do computers learn today?
Often, people hear about ML and imagine a plug-and-play approach. They try it, but it doesn’t work.
Well, let’s be fair to computers. Imagine teaching a child some complex concepts without first teaching them the basics like reading and writing. Even the smartest child would fail.
Unfortunately, this is what some people do to computers. We train computers on complicated (real) data, which confuses the computer. We don’t do this when teaching people — why then, are we simply throwing computers into the deep end?
One reason is that we need and want to see results immediately. What’s more, many companies simply don’t have the data, expertise or resources required to take a computer through all the steps needed for proper training.
Some companies have already taken a step in the right direction by building artificial data to train computers, including Generative Teaching Networks and Progressive Learning methodology. In Andrew Ng‘s recent session, the Google Brain co-founder spotted the benefits of putting more emphasis on data preparation.
Obviously, each industry and project within the field requires training. This layer of training can be likened to college-level knowledge and should lay on top of the earlier learned foundation.
So, what should ML training look like to actually work?
Step 1: Start with the basics.
Before a student can learn algebra, they need to be able to count.
Computer training ML should also start with simple data. Think computer daycare. Slowly, the complexity of the data is increased, working up to using real data.
Step 2: Building a foundation.
Children are commonly taught “stranger danger,” which encourages young people to trust their parents and certain authority figures over someone they don’t know. For machines, it’s key to understand that not all data is created equal. ML should be instructed on how to recognize trusted sources.
Step 3: Specialization.
Children entering daycare all come with a “blank slate”. They are all taught the same basic information: the letters of the alphabet, how to count and who to trust. By the time they go to school, they can pick up a simple book and perform basic arithmetic.
Slowly, throughout their educational journey, each child begins to develop more knowledge in a specific subject. One child is better at science, while another excels in the arts. At the college level, students are taught in-depth on a subject within a narrow field.
Similarly, this measured approach must be applied when training an ML application. To train computers in specific areas such as banking, retail or medicine, you first have to build a foundation on simple artificial data, adding new layers of information until the application is finally ready to be trained for the specific industry.
The prediction for the next trillion-dollar business.
While the first step in ML training seems clear, it is not easy to implement, since:
- We don’t have time
- We are not qualified to build early childhood education programs for computers
- It is a very expensive undertaking
What’s more, the complexity only increases with each additional step in the ML training process.
All this to say, I predict the next trillion-dollar company will be a system of public educational institutions for ML. This is where future computers will be trained to enter the workforce.