What Is Deep Learning?

What is the difference between machine learning and deep learning?

Deep learning is a subset of machine learning. All deep learning is machine learning, but only some machine learning methods and applications are deep learning.

Machine learning has become a very wide field within the even broader discipline of artificial intelligence. In the broadest sense, machine learning revolves around the creation of machine learning models, which represent information learned about a given set of data. Those models can then be applied to additional data to make some kind of observation or prediction about it. For example, one common machine learning model is a simple email spam filter. A machine learning model is trained on thousands or millions of email messages — some known to be legitimate, some known to be spam. As part of this training, the model “learns” and develops a set of rules about what is likely to constitute spam (all caps in the subject line, pharmaceutical keywords in the message body, etc…). When a new email message arrives, the model can then make a determination about whether it is legitimate or not by applying those rules to the question.

Deep learning operates using a considerably different process. A deep learning model can be developed to solve the same kinds of challenges that more traditional machine learning models are used for. For example, when you encounter a spam message, you’re usually well aware of it, even though one spam message may be completely different from another. Over time, and after being exposed to thousands of spam messages, you have developed an internalized, abstract idea about what constitutes a spam email.

Deep learning algorithms are designed to work the same way. It steps each message through layers and layers of processing units that each work toward a calculation of the probability that the message is spam. As the message passes through each layer, the confidence of the calculation improves until, eventually, the model makes a final judgment.

What is the difference between AI and deep learning?

As with machine learning, deep learning is a subset of artificial intelligence (AI). AI is even broader than machine learning, comprising a wide range of technologies designed to mimic human behavior in some way. AI in the broad sense does not have to rely on neural network technology the way that deep learning does. Applications including computer-controlled video game “bots” and resume filtering tools are just a few of the examples of AI that don’t typically have a deep learning connection.

How does deep learning work?

In its simplest explanation, deep learning is something that mimics the way the brain works, albeit on a significantly smaller scale.

Deep learning is always built atop of an artificial neural network. Deep neural networks are designed with artificial neurons, which are processing nodes that perform various computations. These nodes are arranged in layers, including an input layer (far left) and an output layer (far right). The layers in between, known as hidden layers, are where the meat of the processing and “learning” takes place. All deep learning implementations, by definition, have more than one hidden layer, some with hundreds of hidden layers, stacked atop one another.

All of the neurons in one layer are connected to all of the neurons in each layer above and below that layer. When an input is passed from a higher layer to a lower layer, it is assigned a weight that determines the relative importance of the input. While these weights are initially random, they are refined over time as the model learns based on the quality of the output of the model; in this way the model determines what variables are important and which ones are not. As the model is trained on a large amount of data, it learns how to optimally predict an output based on a set of inputs.

One hypothetical, if not greatly simplified example, is forecasting the loan rate that a given mortgage will have. The raw data would be considerable, including the amount of the mortgage, the location of the home, the size of the down payment, the income of the applicant, the net worth of the applicant, their credit score, and potentially dozens of additional variables. All the input data would comprise the inputs along the left side of the image seen above.

The model is trained on historical mortgage data with all of these variables and the corresponding interest rate that was offered to the applicant. The neural network measures the strength of connection from one neuron to another as data is fed into the network. At the first level in our sample network, one neuron would measure the importance of credit score vs. income, another would measure income vs. down payment, and so on. Each of these connections is given a weight based on how the model predicts it will impact the output. The process is repeated at each level of neurons until it reaches the output level. The network then compares its calculated output (the mortgage rate) against the actual mortgage rate suggested by the training data.

If this all sounds complex, it is, and some very advanced math and data science is required to make this workable and efficient. However, the accuracy of a deep learning model is often found to be better than that of a traditional machine learning model, which makes it worthwhile in high-value situations.

Learn more about the Splunk App for Data Science and Deep Learning here.

Here’s a look at a wide range of deep learning applications that are aimed at solving real-world problems with big data.

• Healthcare imaging and disease diagnosis: Models are trained on large datasets of x-ray medical imaging and other medical scans to determine whether a tumor or some other pathology may be present. These images can assist human technicians by pointing out areas they might have missed or, increasingly, making diagnoses by themselves.
• Fraud detection: By mining thousands or millions of financial transactions, a deep learning model can discover anomalies that may be indicative of financial crime and fraudulent behavior. This can be useful for retailers, financial institutions and even consumers themselves
• Network intrusion detection: Deep learning systems can be used by intrusion detection systems to learn how to recognize behaviors related to cyberattacks on the enterprise, helping to protect computer systems even before specific vulnerabilities have been identified.
• Chatbot development: Deep learning-based chatbots are considered much richer and more natural at carrying on conversations with humans than those trained using more traditional, simpler ML techniques. These chatbots study real human interactions and rely on reinforcement learning to understand the ins and outs of naturalistic language.
• Real-time translation: New tools driven by deep learning computing power are making it possible to smoothly translate one language to another in real-time, with minimal delay, both via text and synthesized voices.
• Spam detection and elimination: Deep learning is the main catalyst for today’s most successful, high-performance spam filtering technologies.
• Robotic motion: Today’s robots are no longer single-purpose machines like those in the automotive factories of yesteryear. Deep learning can help robots learn multiple skills and pick up new ones as time goes on.
• Facial recognition: Facial and other image recognition technology, which includes feature extraction, is increasingly being used for everything from unlocking smartphone devices to social media applications to security systems for businesses, airports and law enforcement.

What are the benefits of deep learning?

Although deep learning is not appropriate for every data analysis or computing problem, it offers numerous benefits over traditional machine learning, including:

• Machine learning algorithms can often hit compute limits after they’ve been fed a relatively small amount of data. For most deep learning scenarios, more training data will improve the model’s accuracy beyond that of a traditional machine learning model – although there are limits to how far it can go.
• Traditional machine learning sometimes requires a technique called “feature engineering,” wherein a data scientist cleanses, curates, and labels data to be used in the training process. That said, while deep learning is still reliant on feature engineering, you often don't need to do as much exploratory data analysis to understand your features in the way you do with other forms of machine learning.
• Though it can generally work with any type of new data, deep learning is designed with unstructured data in mind, particularly suited to processing unstructured data sources, such as images or text, as compared to other machine learning techniques.

What are some of the challenges of deep learning models?

While immensely beneficial, deep learning is not a panacea for every computing and cloud computing problem. Some of the challenges of deep learning include:

• Slow to learn: Because deep learning requires vast amounts of data in order to train its models, training can take significantly longer than is required by traditional machine learning. The calculations are also more complex and can therefore take longer to train a model.
• Costly: Deep learning requires a large amount of high-power computing resources, often in the form of powerful servers outfitted with costly graphic processing units (GPUs.) While modern enterprises can rely on cloud-based resources to complete the work, expenses can add up quickly.
• Lacks process transparency: The neural network used in the deep learning model is essentially a “black box” that is largely opaque in the way that it makes decisions, which can frustrate users looking for more insight into how a decision is made or if a decision is without inherent biases.
• Overfitting solutions: Overfitting happens when an algorithm is trained on so much data that it begins fitting its model to noise in the training data. Because the model has made accommodations for so many outlier scenarios, tests on training data could possibly become more accurate, while tests on real-world data might be less accurate.
• Lack of qualified staff: If you want to get started with deep learning in your organization, it can be difficult (and expensive) to find trained computer science and data science workers.

What is the future of deep learning?

While deep learning is already the subject of intense investigation, a significant amount of research is giving the field even more room to grow.

For example, new models such as the transformer network architecture are increasing the speed of deep learning processes and training processes by improving the way parallel processing works. Other techniques such as contrastive learning are improving how deep learning technologies determine what is missing — or obscured — in a piece of visual data. (For example, contrastive learning might be used to guess what is obscured behind a person in a photograph.)

Other technologies are being actively developed that will continue to nudge the architectures and processes of today’s neural networks closer to those of the human brain. These techniques will allow deep learning systems to determine not just what a visual object is, but how it would interact with other objects by understanding its composition and other physical characteristics. Similarly, these techniques are improving the way deep learning image analysis works in a three-dimensional space, a critical function for applications such as autonomous driving.

While deep learning may be seen today as a category appropriate for theoretical research and resolving complex but largely esoteric problems, it is set to become increasingly accessible to the masses, applied to an ever-broadening collection of business and societal issues.

Deep learning has a vast number of applications and activations in the business world, many of which we are only just beginning to realize. While the technology is complex and challenging, the transformative capabilities that it enables are impossible to ignore. In the business world, deep learning has applications ranging from product development to manufacturing floor design and from financial decision-making to supply chain management. We’re still in the early days of developing deep learning tools that can use a robotic camera to capture images of new products coming off the production line and predict whether finished product quality is going to decline, but those tools are coming. While all practical applications may not be immediately available, smart enterprises are already investigating the many ways that deep learning can be put to potential use in their organizations.