Why it's time to learn more about deep learning

[ComputeExperts]

Deep learning is a subset of machine learning, and chances are you've already used it whether you know it or not.

While you most likely have heard the terms deep learning, machine learning, and AI, you might not be as familiar with real-world deep learning solutions. As you might suspect, the terms are interrelated. If it isn't already, deep learning should be on your radar.

Digging into what deep learning means

What is deep learning? It is a subset of machine learning, and both fall under the umbrella of artificial intelligence, which refers to systems built to carry out tasks that normally require human input. Machine learning software teaches a computer how to perform a task, rather than programming it to do so. This type of software gives computers the ability to tackle and solve more complex problems within deep neural networks, which couldn't otherwise be handled. These neural networks pose a series of binary questions and classify all the data based on the answers received, notes Forbes.

How can deep learning help you? One area where deep learning solutions are proving effective is in recognizing and generating images, language, and speech by enabling machines to learn how features in the data combine into increasingly higher-level, abstract forms, according to ZDNet. The technology is changing almost every industry—from healthcare to retail to insurance to automotive, to name a few. One of the significant ways deep learning is making a difference is in making predictions. For example, a team of researchers at New York University is using an open-source neural network to detect certain types of lung cancer—with 97 percent accuracy, according to VentureBeat.

Deep learning solutions are pervasive

You may already use products that utilize this technology. In facial recognition applications that are commonly found on smartphones, for example, software examines how pixels in an image create lines and shapes, how those lines and shapes create facial features, and how the facial features are arranged into a face.

These systems are used by Amazon's Alexa virtual assistant to understand what someone is saying and by Google to translate text from a foreign language. As ZDNet notes, every Google search uses multiple machine-learning systems to understand the context of the language in a query. When the search engine personalizes results, a fishing enthusiast searching for "bass" isn't inundated with results about guitars.

Other areas where deep learning is making a mark include:

• Computer vision for driverless cars, drones, and delivery robots
• Speech and language recognition and synthesis for chatbots and service robots
• X-ray identification of tumors for radiologists
• Analysis of IoT sensor data that enables predictive maintenance on infrastructure

These are just a few of the myriad use cases for deep learning.

First steps before getting in too deep

Organizations of all types are beginning to unlock data insights to achieve better business outcomes and stay competitive. If you have a large amount of mostly unstructured data, consider applying this type of system. The first step is to develop the right data strategy, utilizing either public data sets or building your own data sets; outsourcing the work may prove worthwhile in the latter case.

Then you need to build and train models based on your particular business challenges. It's a good idea to start with a large model to ensure good results, advises Forbes. You can always scale back the model size.

Developing this type of model requires a robust data and analytics system. The data must be clean, available, and reliable, according to O'Reilly. Your organization must also have highly available and reliable data streams that you can integrate with data pipelines wherever the data is stored.

Enterprise data is increasingly spread across a hybrid cloud environment and different storage formats. It's important to ensure there are connections between data residing in public clouds, on-premises data, and data that persists in different kinds of object and file storage. Once in production, deep learning solutions must also be able to iterate rapidly, and you need to ensure the data's quality and governance through visibility, monitoring, and system retraining.

More considerations for deep learning

The nice thing about deep learning algorithms is that they can take unorganized and usually unlabeled data—such as video, images, audio recordings, and text—and organize that data to make useful predictions, notes ZDNet. Because the strength of deep neural networks is in making predictions on mostly unstructured data, these networks can help solve complex problems like image classification. The software can build a hierarchy of features that make up, for example, a dog or cat. Deep learning is also ideal for natural language processing and speech recognition.

Bear in mind that while their strength lies in giving order to large data sets, deep learning solutions also require access to significant computing power. The cost of training is another factor, because it typically requires access to high-powered hardware, generally high-end graphical processing units (GPUs), or GPU arrays. You should also invest in a robust data and analytics foundation.

Deep neural networks are also difficult to train, because of the so-called vanishing gradient problem, observes ZDNet. As more layers are added, it becomes difficult to calculate the adjustments needed at each step of the training process. It also requires a long time to train a neural network to reach an acceptable layer of accuracy.

There are several types of deep neural networks, and one model is not inherently better than the other. It boils down to which is best suited for a particular type of task.

Generative adversarial networks (GANs), for one, are extending what neural networks can do. In a GANs framework, there are two neural networks: one acts as a discriminator, and the other a generator. These two networks battle each other to build the best algorithm for solving a problem. The generator network uses feedback it receives from the discriminator to get better at producing fake data, while the discriminator gets better at spotting it.

The idea is that by pitting the two networks against each other during training, both can achieve better performance.

Developing better business outcomes

There's no doubt that in today's increasingly fast-paced world, you must use your information assets to gain a better understanding of your customers, products, competitors, industry, and employees.

At their best, these systems can remove humans from the machine learning process and lessen the potential for system biases. But you need significant network infrastructure and the right platforms and configurations for real-world deep learning workloads to unlock data insights and achieve better business outcomes.

"Human decision making is increasingly inadequate in a new digital world with an ever-expanding universe of data," notes Gartner. "Data science and especially machine learning excel in solving the kind of highly complex data-rich problems that overwhelm even the smartest person."

Use cases for deep learning are increasing as more and more organizations recognize the benefits of the technology. If you're looking for information as you begin your deep learning journey, this HPE AI and Deep Learning for Dummies guide will help you understand what AI, deep learning, and machine learning can offer you and your organization.

---

 Meet Server Experts blogger Esther Shein. Esther is a longtime freelance tech and business writer and editor, whose work has appeared in several online and print publications. She has also written thought leadership e-guides, customer stories, and marketing materials.

 
Server Experts
Hewlett Packard Enterprise

 @HPE_HPC
 linkedin.com/showcase/hpe-servers-and-systems/