There’s no denying that deep learning is a hot topic right now. But what does it really mean, and how should it be applied in practise? In this article, I’ll look at how deep learning is best used … and when it should be avoided.
What is deep learning?
To understand deep learning, you first need to understand artificial intelligence and machine learning:
- Artificial intelligence (AI) is centred around the concept of building machines that can think like humans. Since we’re (for now) some way off machines that can truly act and think like humans, AI can best be described as the ability of machines to interpret the world around them and make decisions.
- Machine learning is a subset of AI, or, rather, it’s the cutting edge of AI. Specifically, machine learning is about teaching computers to learn from data, and make intelligent decisions or take action based on what they’ve learned.
Deep learning represents the next level of machine learning – it’s the cutting edge of the cutting edge, if you like. With machine learning, the computer or machine ‘learns’ from the data it’s given. So a programmer could ‘teach’ a computer to recognise images of cats, by giving it a teaching set of images: some that contain cats (these images would have to be labelled as ‘cats’) and some that don’t contain any cat (labelled ‘not a cat’), and directing the algorithms to the different variables to distinguish cats and dogs. The computer then ‘learns’ from the training set and applies that knowledge to a new set of images, getting better over time as it successfully identifies more images of cats and adds to its teaching set. If the machines return wrong results a programmer would help to adjust the code.
But with deep learning, machines don’t require a human programmer to step in. The computer can determine by itself if its predictions are correct or not. It does this by continuously assessing data via layers of artificial neural networks that mimic the decision-making processes in our human brains. In order to work well, deep learning algorithms require much larger data sets than traditional machine learning applications. So in our cat example, there’s no need to tell the algorithms which variables to use to distinguish between images with cats and those without – you simply give the computer millions of images and let it work out for itself what a cat looks like.
Deep learning in action
Let’s look at two real-world examples of deep learning at work:
- Translation and language recognition: deep learning is making automatic translation much more accurate, and it’s also enabling machines to not just recognise which language is being spoken, but which dialect is being spoken. All without human involvement.
- Autonomous vehicles: as a driverless car makes its way down the road, it’s making decisions based different deep learning models. For example, one model will specialise in interpreting street signs, while another will recognise traffic lights, while another will recognise cyclists, and so on.
When not to use deep learning
Deep learning is certainly impressive and exciting, but it’s not automatically suitable for every situation. In fact, there are certain circumstances where deep learning is probably not the best solution.
For one thing, deep learning really needs Big Data to make accurate decisions. So if you haven’t got an extremely large dataset to learn from, a regular machine learning algorithm is likely to deliver more accurate results.
It’s also more expensive to implement because it takes a lot of computing power to run a deep learning network. While services and tools like IBM’s Watson are helping to lower the barrier to entry for deep learning, remember that deep learning is still at the very cutting edge. For the average (i.e. Non-Google) business on a budget, deep learning could be too expensive to be practical.
Where best to apply deep learning
Deep learning is ideal for predicting outcomes whenever you have a lot of data to learn from – ‘a lot’ being a huge dataset with hundreds of thousands or better millions of data points. Where you have a huge volume of data like this, the system has what it needs to train itself.
It’s also best when applied to complex problems and things that would be vastly expensive to solve with human decision making. Image processing is a great example of this. So, rather than YouTube paying an army of human workers to trawl through millions of videos and tag the ones with cats for our viewing pleasure, it makes much more sense to apply deep learning. It’s the same with translation and speech recognition.
And last but not least, deep learning is only appropriate if you have the high-end computing power to make it work, or are partnering with an analytics provider who has the infrastructure and skills that might be lacking in-house.
If your circumstances match these criteria, then deep learning might be the ideal solution to your business problem. For anything else, less cutting-edge analytics solutions might be a better route to success.
Where to go from here If you would like to know more about Deep Learning, Machine Learning, AI and Big Data, cheque out my articles on:
Or browse other related articles.