Source: Deep Learning on Medium
People like to throw buzzwords like artificial intelligence, machine learning, and deep learning into conversations. I plead guilty. They accurately describe the work I do.
That does not offer an excuse to hide behind buzzwords without understanding what they mean. So, let’s go over what they mean so you know when to use each in conversation. Along the way, we’ll also see what they mean for healthcare and digital health innovation.
First, let’s explore how these three concepts relate to each other with a handy picture:
AI is the broadest category. This means that all machine learning and deep learning count as AI. And all deep learning counts as machine learning. However, this does not hold true vice-versa, ie. not all AI is machine learning. Look at the picture again to see what that means.
AI refers to any type of machine with intelligence. This does not mean the machine is self-aware or similar to human intelligence; it only means that the machine is capable of solving a specific problem.
Machine learning refers to a particular type of AI that learns by itself. And as it gets more data, it gets better at learning.
Deep learning refers to a particular type of machine learning that uses neural networks. It typically gives the best results by far out of any type of machine learning.
While neural networks were invented in the 1980s, they only became popular in the 2010s due to several technical breakthroughs. Basically, neural networks became faster, cheaper, and more accurate. And hence deep learning was born as a clever marketing term to describe these recent breakthroughs.
So what is an example of AI that is not machine learning? “Expert systems” basically set a number of “if this, then do that” statements. It does not learn by itself (so it is not machine learning), and it still can be very useful for use cases like medical diagnosis and treatment.
This decision tree becomes AI once it is put into a computer. The computer will need information to answer the questions, and then it will automatically give the treatment. For example, give the AI this information:
- Patient temperature is 102°F
- The patient is in pain
- The infection is bacterial
And the AI will automatically prescribe antibiotics for the patient.
As healthcare becomes more connected with sensors, we can automatically feed AI with more data to help it make decisions, such as for clinical decision support. This means that a healthcare use case may not need buzzy deep learning. It could only need a couple of connected sensors and new AI expert systems.
At the same time, countless opportunities lie in machine learning and deep learning. Take annotating radiology images. Machine learning learns how to spot tumors by itself, and deep learning in particular provides excellent results. Deep learning already outperforms human radiologists at a fraction of the time.
However, deep learning comes at a significant cost: we usually cannot say why it made a certain prediction. We gain accuracy at the cost of interpretability.
This presents a dilemma in healthcare. Let’s imagine deep learning predicted a tumor to be benign. It turns out the tumor was malignant, and the prognosis became much worse in the meantime. The patient, family, and doctors all would likely want to know why the neural network thought the tumor was benign.
The neural network shrugs in response.
In cases that need responsibility for a decision, interpretability matters. In such cases, it may be better to use other types of machine learning over deep learning. There will probably be an accuracy decrease at the benefit of being able to say why it made a certain diagnosis.
The overarching point is that there is a time for AI, a time for machine learning, and a time for deep learning. There are also times when we should not use any of these. Some things in healthcare should be left to humans for the foreseeable future.
Understanding the difference between common buzzwords allows for productive conversations. Here are some buzzwords that are commonly conflated:
Artificial Intelligence: A machine with any type of intelligent behavior
Machine Learning: A part of artificial intelligence that learns by itself
Deep Learning: A part of machine learning that uses neural networks
Interpretability matters when a decision requires responsibility. Here are core points:
- Interpretability is the ability to say why a model made a prediction
- AI and machine learning sometimes make interpretable predictions
- Deep learning is usually more accurate while not being interpretable
AI, machine learning, and deep learning can all be the right solution. It takes discernment to figure out which one is best for the problem at hand.