What is deep learning? How does it work? What is the difference between machine learning and deep learning? We break down this branch of artificial intelligence in plain terms so you can explain it – even to non-techies
How deep learning works: Think of a child’s brain
Helping others understand deep learning can again be boosted by a comparison to the human brain – especially that of a child.“After enough feedback, the child eventually forms an internal mental model of how to label different objects in the world.”
“Look at how a child learns language: The child points at an object and says, ‘car.’ The child’s parent immediately provides feedback: ‘Right’ or ‘No, that’s a jar,’” explains Moshe Kranc, CTO at Ness Digital Engineering. “After enough feedback, the child eventually forms an internal mental model of how to label different objects in the world. How does the child’s brain organize its billions of neurons in order to deliver the right answer most of the time? Each neuron transmits signals to other neurons, in some sort of unexplainably complex hierarchy that is formed based on feedback.”
One way to think about the “deep” in deep learning is that, just like with the human brain, we won’t necessarily understand every factor in a decision or outcome because it involves increasingly complex layers. But that is essentially what deep learning is attempting to mimic: The powerful complexity of the human brain.
“We create a neural network of perceptrons, [or] digital neurons, arranged in layers, with each layer’s perceptrons interconnected to the next layer’s perceptrons,” Kranc says. “We then provide training data, [such as] pictures of objects, to the neural network, and have it try to guess what each object is. Of course, its initial guesses will be atrocious. But as we provide feedback, the neural network adjusts how its perceptrons are connected until it produces highly accurate guesses. At that point, we have a trained model that can label objects with great accuracy.”
That’s key to the long-term promise of deep learning: It improves itself over time, and there’s not a visible ceiling limiting that improvement. But like with human decision-making, there may be some opaqueness when it comes to deciphering those decisions at a granular level.
“Unfortunately, as with the child’s brain, it is difficult to explain what each layer or each perceptron ‘knows’ or ‘represents,’ because what matters is not individual perceptrons but how they are connected,” Kranc says.
Deep learning and bias: A danger zone
Like other branches of AI, deep learning is not without other important considerations, such as the possibility for bias. If the widest explanation of deep learning is that it’s an attempt to mirror the human brain with a machine, then it follows that the machine is susceptible to “human” error.Count that as another reason for non-technical folks to achieve at least a basic understanding of deep learning and how it works.
“It’s easy to blindly trust the results of a deep learning algorithm, but like any machine learning algorithm, the results are only as good as the data the algorithm is trained on,” Brock says. “If the data contains unconscious bias or issues with fairness, machine learning can replicate those issues. These ethical concerns are important to keep in mind as we use applications that leverage deep learning.”
Count that as another reason for non-technical folks to achieve at least a basic understanding of deep learning and how it works. Actually, Brock notes that as the applications of deep learning and other AI disciplines grow in our everyday lives – from customer service interactions to medical diagnoses, as examples – it will be important for just about everyone to have some understanding of how it works.
Similarly, like with other significant technological developments, no one should expect “easy” as the default setting.
“To be fair, it’s not a free lunch,” Wilde says. “Deep learning is a very smart ‘learner’ but it’s also a very slow ‘learner.’ It needs hundreds of thousands of examples to find the solution to the example problem a user poses. New breakthroughs with approaches like transfer learning enable the deep learning algorithms to learn with a very small fraction of the volume of training examples typically required. Regardless, we have for the first time moved from needing to learn to speak the computer’s language to it learning to speak ours.”
What is transfer learning?
Transfer learning is not a new concept – University of Wisconsin researchers wrote about it back in 2009, for example – but it appears likely to become an area of increasing interest in machine learning. In a Medium post, KC AI Lab founder Brian Curry defines transfer learning as “a machine learning method where a model developed for a task is reused as the starting point for a model on a second task. Transfer learning differs from traditional machine learning in that it is the use of pre-trained models that have been used for another task to jump-start the development process on a new task or problem.”
Put another way, it’s the pursuit of making machine learning – and subsequently the sub-discipline of deep learning – more reusable, as opposed to the more common single-use models that don’t shorten the learning curve on new tasks.
Again, there’s plenty of AI terminology out there – probably more than most people need to know. Transfer learning may be a subset that doesn’t need to be on everyone’s radar just yet unless they’re actually working on machine learning or other AI disciplines. Deep learning, on the other hand, is already worth explaining to a much broader audience in our organization.
“Deep learning is increasingly becoming part of our everyday life, and it’s important for [people] to understand the basics of how it works,” Brock says.