The next generation of scientists may be coming from a different world to the one you’re accustomed to.
This is the world of data, artificial intelligence, artificial neural networks, and the cloud.
The question is, which one of those comes from you?
You might have heard about them before.
You might be familiar with them from your favorite television show.
You probably know the jargon: deep learning, neural networks.
And then you might not have heard the name Watson.
If you’re not familiar with the name, it stands for “digital neural network.”
In the first few years of Watson’s career, she was working with IBM Watson, the company’s Watson machine learning platform, and researchers at Carnegie Mellon University and Harvard University.
But Watson was far from Watson.
“She was just a really good friend,” says Chris Anderson, a professor at the University of Pennsylvania who helped develop Watson’s training data.
“She was very good at taking the time to work with people, and her approach was much more open-ended than Watson’s,” Anderson says.
At the time, Anderson says, “we were not ready for the scale of AI, the scale that the world is dealing with today.”
“Watson was not an overnight success,” Anderson adds.
“And she was a really, really good teacher.”
The future of artificial intelligence was on the horizon.
But the time had come to find out what Watson’s data would tell us about the future.
In the spring of 2018, Anderson and other researchers at MIT and Carnegie Mellon joined forces to build Watson’s first version of deep learning.
Their plan was to train Watson to do tasks that would be of interest to the rest of the world.
“We thought, maybe we can find a way to train her to do something that’s not really relevant for the rest, but we can learn something about it that she doesn’t have a clue about,” Anderson recalls.
Anderson and his colleagues were also trying to build a system that could learn from its own training data to improve its ability to predict things.
The challenge was, how could they train the system to do things that had never been done before?
So they started thinking about learning in a way that would give it some real-world examples.
For instance, in the future, there might be robots that are able to drive people around.
In the past, the best robots were used to move people around, like driving a car.
Anderson and his team had hoped that they could use a robot to learn how to navigate and operate a vehicle.
And they could learn to do it by using a few data sets, one from the future that they knew about, and another from the past.
The results were promising.
“This is a really cool result, because it gives us some interesting insights about how we can use this data to build an artificial system that can learn,” Anderson notes.
But the problem was that Watson was still learning.
She needed more data.
So Anderson and the rest built another system, called a neural network.
This was a much more sophisticated version of Watson, with an extra layer of layers of training data that was stored in a big data warehouse.
It was called a deep neural network (or DNN), and it allowed Watson to learn.
And the training data was now stored in Watson’s own database, called the Watson Cognition Warehouse.
It would allow the system the ability to learn and improve itself.
And it would help Watson to understand the world more.
Anderson says he is “pretty happy” with the way this worked out.
A big question is how the system learned.
The next big step was to find a model of how the neural network learned.
That’s where a computer scientist comes in.
Anderson had a few ideas.
“I knew there were a couple things I wanted to do that weren’t quite there,” Anderson explains.
One of those things was to create a model that would describe how the brain learned, something called a learning algorithm.
And Anderson and a colleague named Tod Kwon did just that.
Tod Kwan’s model of the human brain was the kind of model that is often called the human machine interface (HMI), a way of thinking about how computers interface with the world, or interact with it.
When Anderson and Kwon built the model, they realized that the problem wasn’t just one of training a neural system to learn, but also how it learned.
So they decided to create an algorithm called a recurrent neural network, or RNN, that would learn by learning from its training data, using the same kinds of models that were being used to train other types of systems.
As a result, the RNN could learn how the human brains learned.
And Kwon’s model was able to learn to understand more than just the world around it.
And in doing so, it was able learn how its neural network was able find things