Deep Learning
What it is and why it matters
Deep learning is a subset of machine learning that trains a computer to perform human-like tasks, such as speech recognition, image identification and prediction making. It improves the ability to classify, recognise, detect and describe using data. The current interest in deep learning is due, in part, to the buzz surrounding artificial intelligence (AI).
Think about everyday systems like Siri and Cortana. These are powered, in part, by deep learning.
How deep learning works
Deep learning changes how you think about representing the problems that you’re solving. With deep learning, data trains the computer, through deep algorithms, to learn on its own by recognising patterns using layers of processing. If you’re someone who’s never heard of deep learning, you may be wondering how all this works.
Let’s say you create a tool that differentiates parsley from cilantro. Two totally different herbs. With machine learning, you’d need to tell the tool which characteristics can be differentiated. In this case, it would be the leaf type. But with deep learning, the features are already singled out by neural networks. This is all done through an automated and unsupervised learning process.
The promise of deep learning is that it can lead to predictive systems that generalise well, adapt well, continuously improve as new data arrives.
Deep Learning for Animal Conservation
How does a computer “see” an image? Jared Peterson, Senior Manager of SAS Advanced Analytics R&D, shows how deep learning neural networks are the science behind computer vision.
In this deep learning example, the computer program is learning to interpret animal tracks to help with animal conservation.
The evolution of deep learning
Over the years, human-to-machine interfaces have evolved greatly. The mouse and the keyboard are being replaced with gesture, swipe, touch and speech, ushering in a renewed interest in AI and deep learning.
Several developments are now advancing deep learning:
- Analytical improvements have boosted the performance of deep-learning methods.
- New machine learning approaches have improved accuracy of deep learning models.
- New classes of neural networks have been developed that fit well for applications like text translation and image classification.
- A lot more data is available to build neural networks with many deep layers, including streaming data from the Internet of Things, textual data from social media, physician’s notes and investigative transcripts.
- Computational advances of distributed cloud computing and graphic processing units have put incredible computing power at our disposal. This level of computing power is necessary to train deep algorithms through deep learning.
Deep learning opportunities and applications
A lot of computational power is needed to solve deep-learning problems because of the iterative nature of deep-learning algorithms, their complexity as the number of layers increase and the large volumes of data needed to train the networks.
The ability of deep-learning methods to continuously improve and adapt to changes in the underlying information pattern present an opportunity to introduce more dynamic behaviour into data analytics.
One of those possibilities is the greater personalisation of customer analytics. Your favourite streaming service creating a customised list of shows you might like based on your previous watch history would be an example.
While the current focus of deep-learning techniques is in applications of cognitive computing, there is also great potential in more traditional data analytics applications. For example, time-series analysis.
Another way deep learning can be applied is to simply be more efficient and streamlined in existing analytical operations. Recently, SAS experimented with deep neural networks in speech-to-text transcription problems. Compared to the standard techniques, the word-error rate decreased by more than 10% when deep neural networks were applied. Neural networks also eliminated about 10 steps of data pre-processing, feature engineering and modelling. The impressive performance gains and the time savings signify a paradigm shift.
Deep learning in today's world
The impact of deep learning is significant – and it’s only getting started.
Deep learning and GANs: How they relate
A generative adversarial network (GAN) is a type of machine learning algorithm. GANs help data scientists create synthetic data for data-hungry deep learning models. That’s important because using synthetic data allows for the creation of deep learning models not previously possible.
Discover how SAS uses deep learning
This step-by-step guide compares multiple neural network models and explains how to use them. You'll get an introduction to deep learning techniques and applications and learn how SAS supports the creation of deep neural network models.
How to use deep learning for embedding images
Embedding models reduce the dimensionality of input data, such as images. With an embedding model, input images are converted into low-dimensional vectors – so it's easier for other computer vision tasks to use. The key is to train the model so similar images are converted to similar vectors.
How deep learning works
Deep learning changes how you think about representing the problems that you’re solving with analytics. It moves from telling the computer how to solve a problem to training the computer to solve the problem itself.
Feature representation
Deep learning is a paradigm shift in model building that moves from feature engineering to feature representation.
Deep learning layers
Instead of using known variables to predict unknowns, deep learning uses looks at layers of the data to recognise latent features of the data.
Deep learning results
The promise of deep learning is that it can lead to predictive systems that generalise well, adapt well, continuously improve as new data arrives. You no longer fit a model. Instead, you train the task.
A traditional approach to analytics is to use the data at hand to engineer features to derive new variables, then select an analytic model and finally estimate the parameters (or the unknowns) of that model. These techniques can yield predictive systems that do not generalise well because completeness and correctness depend on the quality of the model and its features. For example, if you develop a fraud model with feature engineering, you start with a set of variables, and you most likely derive a model from those variables using data transformations. You may end up with 30,000 variables that your model depends on, then you have to shape the model, figure out which variables are meaningful, which ones are not, and so on. Adding more data requires you to do it all over again.
The new approach with deep learning is to replace the formulation and specification of the model with hierarchical characterisations (or layers) that learn to recognise latent features of the data from the regularities in the layers.
Advanced Analytics from SAS
Deep learning is just one technique in the data scientist's toolkit. Learn about other advanced analytics techniques, including forecasting, text analytics and optimisation.
Recommended Reading
-
Article
How to drill a better hole with analyticsFrom drilling holes to preventing health care fraud, learn about some of the new technologies SAS has patented with IoT and machine learning technologies.
-
Article
Machine learning for beginners and beyondWhether you’re an experienced data scientist or a machine learning beginner, you’ll appreciate these 10 tips for getting started with machine learning.
-
Article
Australian Superannuation at RiskWith over $2tn of assets, Australians’ retirement savings are under attack from Fraudsters and Hackers, even more so as Superannuation providers move into Digital with customer portals and faster processing. However, it is not all doom and gloom as behavioural signals of fraud are typically there well before the actual fraud occurs.