What is Deep Learning?
The most frequent question asked by my students is: Do I need to learn deep learning? Beyond the buzzwords bounced back and forth in blog posts and news articles, deep learning is probably the most revolutionary technology of the last century. Discovered in the 1950s and 60s, and further developed in the 1990s and early 2000s, deep learning remained mostly an academic topic until 2012, when a team led by Geoffrey Hinton won the Imagenet Challenge using an algorithm based on deep learning.
Although 2012 seems relatively recent, interest in deep learning exploded since then, as you can see in the chart below from Google Trends.
Deep Learning Explained
Deep learning is a branch of machine learning that has proven to be formidable in multiple domains and applications ranging from computer vision, natural language processing, speech synthesis, product recommendation, and robotic automation.
Deep learning is based on a technology called an artificial neural network, which is a very configurable mathematical function that can learn complex mappings between pairs of inputs and outputs. For example, a neural network can be trained to recognize objects in images by feeding it with a lot of data pairs (e.g., image and corresponding label). Similarly, it can be trained to translate English to Chinese by feeding it with lots of English sentences as input and the corresponding Chinese sentences as output.
If deep learning is just another machine learning technique, why has it become so popular and in high demand?
The main reason for the success of deep learning is that it works incredibly well with unstructured data, such as images, text, sound, time-series of events and so on. Traditional machine learning is capable of finding relations between pairs of input and output data represented as numbers. For example, a bank that is developing an application to score loans, will use known information about its customers and their past loans to train a machine learning model. In this case, both the input and output data are numbers, and traditional machine learning does a fairly good job.
In contrast, developing a computer vision system to recognize objects requires learning to represent images as numbers, a process called feature extraction. Deep learning is particularly suited to learn the best features to solve a problem, using large amounts of data for training.
Success Stories of Deep Learning in Enterprise
According to Andrew Ng, the current deep learning revolution is akin to the invention of electricity, where a single technology became embedded in thousands of products and businesses, and it completely revolutionized society. In his book AI Superpowers, Kai-fu Lee further expands on this concept, explaining that we live in an era of implementation, where engineering talent, data, and computing power are the main drivers for the application of AI in the enterprise, and not necessarily the availability of a rockstar AI researcher in your team.
This fact, combined with even cheaper computing, large datasets and open source frameworks like Tensorflow and Keras, brought deep learning into the enterprise, unlocking tremendous pools of value. Google, one of the companies at the forefront of the deep learning revolution, is including deep learning features in its products (e.g., autocomplete suggestions for Gmail, image search for photos, language translations, etc.). In addition, Google is using deep learning to optimize its business internally to improve search results, reduce the energy consumption of its data centers, and thousands of other projects. While other technology giants like Amazon, Facebook, Microsoft, Baidu, and Alibaba are similarly embedding deep learning in their products, enterprise players are also starting to find useful applications for deep learning technology.
Change Healthcare, a healthcare technology company with 15,000+ employees, has recently unveiled a Claims Lifecycle AI capable of predicting healthcare service denials, representing $6.2 billion in forecasted allowed amounts and millions in potential administrative savings for the U.S. healthcare system. Companies like Arterys and Enlitic are deploying deep learning models to automate the analysis of medical imaging to detect cancer and heart diseases. Banks and insurance companies apply machine learning and deep learning to solve several issues. For example, Visa uses it to tackle fraud and Capital One uses it to improve customer experience.
How to Evaluate if Your Business Could Benefit from Deep Learning
A few simple questions can help you determine if your business could benefit from developing machine learning and deep learning capabilities.
- Do you have data?
- What kind of data do you have?
Let me expand on each.
1) Deep learning needs data, lots of it. Depending on the application, this may mean tens of thousands, to millions and even billions of records. For example, to train a model that correctly recognizes a thousand different categories of objects in pictures, researchers competing on Imagenet used 1.2 million labeled images. To correctly train a neural network for speech recognition, researchers routinely use hundreds of hours of recorded conversation with corresponding transcription.
This may seem intimidating at first, but you shouldn’t be discouraged if you don’t have large datasets available. For one, your particular problem may be solved with much smaller datasets. There are successful examples of relevant business problems like predicting churn or propensity to buy using only a few thousand data points. Also, if you don’t have data available yet, you may be able to find an open dataset that closely resembles your problem and start prototyping your solution while you collect original data.
2) Although deep learning works with any kind of data, it is particularly suited to work with unstructured data such as images, text and sequences of events. Of these, the most relevant to enterprise is text data. Application of machine learning and deep learning to text data involve sentiment analysis, document classification, topic modeling, automated summarization, and generation of original text from a given prompt and a set of sources. A law firm may use deep learning to sift through thousands of cases to help lawyers find the most relevant ones to use. Similarly, an e-commerce business may use deep learning to recommend relevant products to its users as well as to predict what users may spend or how likely they are to click on an ad.
Deep learning and artificial intelligence are General-Purpose Technologies (GPTs) — a technology that can affect an entire economy and every sector of business and society in the same way electricity, railroads and steam engines did.
Since the impact of deep learning on a problem strongly depends on the amount of data available, early adopters of data collection and deep learning will reap increasing returns on their investment and will advance at increasingly faster rate. This will create a secure defensible competitive barrier against latecomers, because the benefits of AI on products and sales will bring more customers, generating more data to train better models on. This is the reason why all major tech companies are competing to adopt and deploy deep learning as fast as they can, and this is the reason why you should make understanding deep learning a priority for your team. To get trained in deep learning take our video course or our in-person bootcamp.
If you want to catalyze value in your data, visit Catalit‘s website. Catalit is a data science consulting and training company focused on machine learning and deep learning.