You need to train an AI application to make it work.
We use an example to explain the method. It is based on a Stanford case with the identification of skin cancer.
In this example, we would like to train our AI to be able to detect whether or not a patient has skin cancer. We don’t know how Stanford did their AI training, but it could have been something like So, let’s say we also would like to train our AI model to do the same as they did at Stanford. Namely, analyze birthmark photos, and use them to determine whether or not they could develop into cancer.
Before AI was able for this task, the usual way to identify skin cancer would have followed a process similar to this description:
You begin with a starting point. Probably one or more high-resolution photos of birthmarks from potential patients. This is our starting point in the illustration below.
From these photos, an action is made by an employee. In our example, this means that an expert will analyze the images. This analysis will be completed with a diagnosis. Either the patient is ok, or he needs further examination if they cannot rule out the risk of cancer by analyzing the pictures.
The process follows a flow which can be illustrated as follows:
It is the same process we want our AI to conduct for us as well. So, our AI can make the diagnosis instead of the expert.
This means that the starting point in the AI case is this one:
Our problem is that we would like to identify patients with skin cancer. Our goal is to be able to sort the patients in two groups: Potential skin cancer candidates, and the rest (the known healthy patients).
As you can see that the illustration above is a process with five different stages:
1. Define goals, start, action and success criteria
2. Collect data.
3. Train model
4. Let your AI make the diagnoses (what is OK and not OK)
5. Model feedback
What you do as you go through the five stages is to train your AI. You teach it how to separate the birthmarks that are ok, from those that are not. What you really do is train your AI to spot what differences in the data that may explain why one birthmark is OK and another one is not.
The AI algorithm does this by building a model for the “ok” and “not ok” case based on patterns in your data. When the AI analyzes a new birthmark photo, then it will try to compare it with the two output cases “ok” and “not ok”. It will then match the birthmark with the output case that holds patterns that are most similar to the photo it is currently looking at. This is how it will make its recommendation as to where to classify any given birthmark photo.
At some point, your AI reaches a point where the quality of its response matches your requirements. You now have an AI model that can perform the job it has trained itself to handle. Sounds easy?
It can be. However, you can expect a great number of elements to happen during the different training phases. The next couple of pages let you in on what to watch out for.
First stage: Define goals
The prerequisites for your AI to take on a task is that it can be described as a flow similar to what is as shown in this illustration. Thus, with a known starting point, an action that has two outcomes. One outcome that is the successful case we aim for, and one that is not.
In this example, there is only one goal, namely if the diagnosis is ok or not ok.
Second stage: Collect data.
Your AI learns from data, and the more data you have, the better prerequisites your AI has to learn. This one of the aspects that makes your AI is also fundamentally different from IT. AI does not relate to the amount of data and the structure of data in the same way an IT system does.
This fact is related to the knowledge pyramid. Remember the weather station from the first part of the book? It was only when the measurements change from data to information that you got value from the readings. Where ” 21″ was converted to “21 degrees”. Putting units on your data converted it to information, and this was what enabled you to extract the meaning of the content.
So, data turned into information. If you need to get results from an IT system, then it must work with information, so that you, as a user, know how to interpret the findings you get. So, IT needs data that has been converted to information to be able to work properly.
AI is not interested in the meanings of data or information as such. AI looks for patterns, and meanings in patterns of data. So, there is no need for AI to know the connection between data sources. Or to know the unit of the data it is looking at. What AI needs is the ability to find patterns in the available data.
So, the quality requirement for AI data is that it is reasonable to expect that it can contain patterns that can explain a behavior. So good quality data for AI is something other than good quality data for IT because the two technologies use data for different purposes.
Third stage: Feed your AI with data and answers
Now you start teaching your AI how to make decisions. You teach it to be able to tell you whether the diagnosis is possible skin cancer or not in our example; you do this by feeding your AI with outcomes, i.e. patient data. You then tell your AI whether the data it is looking at is in the “OK” or “not OK” category.
The purpose of this is to help your AI understand the two different categories “OK” and “not OK”. This exercise will help you AI identify the strongest patterns that best can explain why one photo should be in the “OK” category and why another should be in the “non-OK” group.
So, your AI will split your data into two groups. The data associated with the users with the result “OK” and the data related to the “not OK” users. Your AI will then look at the strongest differences between the two groups.
The AI will look for meaningful patterns that can explain the difference between the two groups. You will get feedback from your AI based on the quality of its findings. Usually, as a percentage describing the level of certainty and accuracy of its findings. If you are happy with that level, you go to the next stage. If not, it will require more or different data for you to be successful, and that means you need to go back to the second stage.
Fourth stage: Let your AI give you answers
At this stage you have an AI model that can give you answers. If you feed it with a photo, it will tell you if it thinks that it is either in the “ok” or “not-ok” category.
Now your AI begins to tell you what the outcomes are. Your AI will now compare the data it is processing from you and compare it to the two pattern groups it has already created, ie. the “OK” group and the “Not OK” group.
Your AI will then choose which of the two groups it believes has the strongest similarities to, based on the patterns in data that it has found in the third stage.
So, you now get answers from your AI. It starts giving you diagnoses based on the data from the cases it knows of.
So, your AI can produce an output. The next stage has to do with validations of the quality of the outputs.
Fifth Stage: Give your AI feedback
Here you tell your AI if the answers you got in the fourth stage were right or wrong. Remember that you feed your AI with test data, so this means that you know what the correct output should be.
This is another dimension in training the AI that has to do with output quality. You will tell the AI where on a matrix the answer would fit.
If you have worked on clinical trials, or in the field of statistical probability, then you may recognize this matrix as a false positive or a false negative outcome.
Remember the focus now is on whether the quality of the diagnosis is OK. It probably won’t be in the beginning. So, you continue the training. Either by running several cases through your AI system, i.e. repetitions of stages three, four, and five.
Your aim is quality improvement, and this means that you want to improve is your AI’s ability to make better decisions. So, it becomes better in understanding what defines the difference between “OK” and “not OK”. You will also be able to go back to the second stage and give your AI more or different data if your suspect inadequate data quality is the cause of low-quality output from your AI.
However, at some point, hopefully, you will reach a quality level of the answers that are so good that you are ready to unleash your AI and let it begin to make diagnoses. At this point, you go live with your solution.