Question: Do AI systems get smarter over time?
Answer: No!

AI systems are smart, but they do not learn after they are deployed.

The “learning” part of an AI system happens during training. Once trained, the model is static until you retrain. Retraining can be expensive, time-consuming, and difficult because of the need for new, labeled data.

What do you mean ‘the model is static’?

Once an AI model is created — that is, once we have an equation and the parameters needed for our task and we are using it operationally — the model is static. We can ask it to perform its task repeatedly, and it will do so, even on data it has never seen before. But it does not get any better at doing that task.

Let me give you an example. Suppose we have a facial matching system. Face matching is the basis for more complex facial recognition systems. A face matching system takes two pictures as input and tells us if they are of the same person. The model is really good at telling if two faces are the same person, but it won’t get any better over time. If I give it two pictures and it says they are not a match, I could give it the same two pictures a year later and I would get the same answer.

So how do we improve system performance?

The only way to make the model better is to retrain it with better data. One thing we notice with our systems in operation is that the type of data changes over time. This can make it difficult for the model to continue to make good decisions.

Let me give you a simple example. Suppose we trained a facial matching system with pictures of adults. Over time, the system starts to see more and more pictures of children, and is unable to identify the matches. We would need to retrain the system using pictures of both adults and children and let it “learn” to match faces, regardless of age.

The new model (equation plus configuration parameters) will be deployed to operations and will work, as is, until we need to retrain again.

Is this what people refer to as Biased AI?

Yes.

You might also hear this referred to as model drift.

The problem isn’t so much the algorithm as the data used to train the algorithm. If the training data is not diverse enough, results may not be reliable. Unfortunately, trying to find sufficient volumes of labeled and diverse data for training is hard. Big platforms are getting better at it – they have millions or billions of photos of people from around the world and can use that to train systems. It is more difficult for different fields of study like cancer cell identification; we just don’t have billions of images we can use to teach the system.

Why doesn’t the operational system learn from experience?

Three reasons:

1. Time – it takes a long time to retrain a system. Training often uses millions, or even billions of data points.
2. Money – it takes a lot of compute resources to retrain a system – see No. 1.
3. Lack of data – The more diverse the operational data, the more data we need for training. Suppose we have an animal recognition system. If all we need to do is say that the picture is either “dog” or “not dog” we would need a lot of dog pictures, plus some other animals to make sure. If we want to be able to classify animals as dog, cat, horse, cow, pig, etc., we need a lot of pictures of each.

Conceptually we could create systems that get smarter over time by continuously retraining the model every time new, labeled data is available. Practically this would be too time-consuming and costly with today’s technology.

---

---