What is Deep Learning?

Supervised learning

When a person is learning a task under supervision, they have a teacher or trainer present who can correct them when they make a mistake. For a computer, this means the data being processed is properly tagged, so the computer is made aware when it makes a mistake.

There are two types of supervised deep learning models — classification and regression.

Classification: A deep learning model could be trained using lots of images of animals. Then, once it sees a new image, it compares that to what it has ‘learned’ from its training examples before trying to predict if the new image contains an animal or not. The algorithm is then judged on how accurately it can recognise an animal in a random image.

Regression: Alternatively, the deep learning model might be fed many variables and then asked to predict a value. One example might be in banking, when customers apply for a loan. The model is trained by crunching through thousands of loan applications, each containing data such as the applicant’s income, credit history, bank balance, employment status and other details, and then told if those applicants were granted a loan or not. The model would then be given a new applicant's details and asked to judge if they qualify for a loan.

Supervised learning is used when there is a large set of labelled data points — such as the images of animals, or loan applicants’ details.

Unsupervised Learning

In this case, the deep learning model isn’t provided with any labels on the data it is processing, instead it is left to discover patterns on its own.

One example might be to feed thousands of songs into a deep learning model and then letting it decide how to group them. As humans, when we listen to music, we recognise its structure, instruments, tempo, vocals and bpm, and then categorise it in genres such as punk, classical or techno. The unsupervised deep learning model doesn’t know what a genre or instrument is, so it would use its mathematical ‘brain’ to find patterns in the tracks and create new categories.

Semi-supervised Learning

This is a combination of the first supervised and unsupervised learning. Deep learning models are fed a combination of labelled and unlabelled data. Some structure and guidance is provided but otherwise they are left to find their own patterns. In the example of music, in this case the songs would have some data labels for each track, maybe the date of first release, which could help guide the deep learning model’s pattern creation to the desired result.

Different forms of AI can be used in a diverse range of use cases. In fact, AI has been creeping into every aspect of our lives because this method of processing data makes computers more efficient  and better at interacting with humans.

In the real world, it wasn’t that long ago when spam filters were crude and discovered potential spam by blocking any messages that contained specific keywords. With the introduction of AI, the spam filters are able to look at different aspects of the email message — where it was sent from, whom it was sent to, the context of the message itself — before deciding whether to send it to the recipient’s inbox or spam folder.

In another example, imagine a deep learning model examining raw data from a live security camera pointed at the forecourt of a petrol station. The video camera produces about 25 separate still images every second, which are fed to the model using a mixture of the different learning systems described above.

Let’s assume it’s a rural gas station and is not very busy. Most of the time the images show an empty gas station with the pumps and the forecourt. After a while, the system can recognise that the ‘normal’ is the empty gas station. When a car pulls in, the deep learning model ‘sees’ the vehicle and knows that something has changed. After further refinement, the model can tell the difference between a car, a motorcycle and a truck. If the system is then connected to the cash register, it’s not a massive leap to be able to tell if the vehicle that just left the forecourt has paid or not. It’s in this way that more layers of data input and processing are added to make the system more ‘intelligent’ and useful in the real world.

Similar deep learning models are used in airports around the world to spot unattended luggage. The deep learning model examines video data from security cameras and recognises people walking around with their suitcases and bags. With many layers of refinement, the algorithm can be taught to send an alert to security when it notices a person has put down their bag and walked away for more than a few moments.

In the financial world, deep learning systems monitor markets and try to understand the movements of different commodities. They can make connections with a vast number of inputs in real time, and slowly understand how some movements in values affect others. The inputs can be diverse — like extreme weather warnings and political unrest, as well as actual market prices. Over time, they can recognise familiar patterns and predict what will happen next. For example, the deep learning model might realise that a massive cyclone connecting with a geographic area that has a significant number of banana plantations will likely result in much higher banana prices next season. This kind of insight could prove extremely valuable to traders.

Deep learning modelling is also used in nuclear power plants, factories and data centres. The systems monitor input from Internet of Things (IoT) sensors and cameras located around the facility. Once the model knows the ‘normal’ operation for the facility, it can send out alerts when it notices unusual behaviour. Unusual behaviour includes things like equipment not working properly, a violent storm outside or certain personnel not clocking in. Cumulatively, these small issues could add up to a bigger problem. Deep learning models are far more reliable than humans at being able to recognise a combination of small factors that could lead to a serious problem, helping engineers avoid major issues before they happen.

Quality of data:  Deep learning systems need to observe data, so they are limited by the quantity and quality of the data that they are fed. If there is not enough data, or the data contains a bias, the algorithm may not be accurate and will reproduce that bias in its results. For example, a deep learning model is fed thousands of photos of birds, and then told to look through a selection of new photos and find the ones containing birds. If the original data contained photos of birds in the wild, in a jungle, with trees in the background, unless the deep learning system was specifically programmed to ignore the background and surroundings, it might have trouble identifying a bird in a photo taken in an urban area, or inside an enclosure.

Inflexibility:  Once a deep learning model is trained, it’s usually only able to deliver accurate results to the same problem using the same type of data. In the above example, if the quality of bird photographs was changed then the system might not be able to compare them accurately. The model would also have to be retrained if instead of birds, the user wanted the system to identify photos of another animal. So even for similar tasks, deep learning models need specific training — they are not flexible.

Computing resource:  Deep learning models require vast amounts of computing power. The reason deep learning is more popular today than it was a decade ago is largely due to the availability of multiple-core processors and GPUs. Deep learning models are also resource hungry when it comes to RAM and storage — as data is moving so quickly, traditional hard drives might not be able to keep up, so SSDs are required.