A Deep Dive Into Deep Learning

How the human brain works has puzzled researchers for millennia — so actually building a brain from scratch is an unquestionably daunting task. But that's essentially what scientists today are attempting to do within a subset of AI dubbed Deep Learning.

By Stephanie Walden, Contributor

Despite complexities of the human brain, scientists today are ostensibly creating one from scratch with one subset of artificial intelligence called deep learning.

The basic building blocks of deep learning are artificial neural networks—algorithms that replicate the biological structure of the brain through “neurons” that contain discrete layers and connections to one another. Every layer of a neural network focuses on a certain type of task, such as recognizing patterns in digital images. Collectively, these layers form depth—hence the moniker, “deep learning”.

It is this concept of deep learning that is not only likely to create something that looks and acts very much like a human brain, but also plays a big role in the next era of human-machine partnerships.

More Data, More Power

The idea of deep learning isn’t new in the AI space, but in the past, there have been significant challenges to its widespread adoption—namely, insufficiencies in software, computing, and analytical power.

AI—the umbrella field that encompasses both machine and deep learning—has historically struggled with basic things such as recognizing faces or different types of four-legged animals. While a human child can learn to decipher the difference between a dog and a cat after seeing just a couple of examples, deep learning algorithms require an immense amount of training—and data—to be able to make this same distinction. (Though these tasks are intuitive for humans, it took millennia for the brain to evolve to the point where those recognitions were processed automatically.)

In addition to plentiful data, deep learning requires data that is free from anomalies and biases in order to accurately inform a machine’s decision-making process. An algorithm, for instance, may incorporate biases of its human creator. In addition, machines face limitations from incomplete data sets or unrepresentative events (bad weather affecting a self-driving car’s performance, for example).

To combat these challenges, methods for collecting, standardizing, labeling, and cleansing data sets are becoming more prevalent. Regulators are even stepping in. In the European Union (EU), the newly implemented GDPR requires companies using machine learning to provide “meaningful information about the logic” used by algorithms for things like credit decisions.

Non-profits such as AlgorithmWatch are also dedicated to the evaluation and development of responsible, non-biased algorithms, and open-source resources like TensorFlow help pool collective human brainpower to develop more advanced AI and neural network frameworks.

Yet, the availability of clean data is only part of the formula for deep learning success. Once a large data set has been collected, researchers need a staggering amount of processing power to store and parse through that information. Although hardware technology is experiencing an exponential rate of progress, engineers are only now developing the types of chips that can handle a process as intricate and data-intensive as deep learning.

The development of optical and quantum computing chips, which are capable of far greater levels of processing power than traditional central processing units (CPUs) and graphics processing units (GPUs), as well as “neuromorphic” chips, which mimic functions of the human brain, are paving the way for deep learning progress.

Wins—and Changing Industries

Indeed, humans have seen some of deep learning’s gains in the past few years. Last year, AI enthusiasts rejoiced in the wake of the AlphaGo victory, a computer that beat the world’s top human player at the notoriously complex Chinese game, Go.

Using deep learning algorithms, AI has identified discrete objects in YouTube videos without human guidance or labeling. And natural language processing, which enables computer recognition, translation, and mimicry of human speech, is also increasingly shifting from statistical models to neural networks.

Proponents of deep learning hope that the technology will soon allow machines to learn without guidance from explicit algorithms. In other words, they’ll be able to develop an unprecedented level of autonomous thinking.

By 2030, analysts forecast that AI will have a global economic impact of $15.8 trillion. As corporate interest in deep learning morphs from mere curiosity to real investment, there are countless industries that will be impacted by new machines that can “think” like humans.

Healthcare: In the field of healthcare, deep learning may help doctors and patients with diagnostics, and researchers discover new drugs. Already, pharmaceutical companies like Merck & Co. and Johnson & Johnson are using deep learning to identify molecules and compounds that could aid in the development of medication for certain neurological conditions.

Doc.ai, a platform that helps users store and manage health data, uses deep learning computing capabilities for predictive analytics. Users track their medical data—detailed accounts of seizures or asthma attacks, for example—on Doc.ai’s blockchain-enabled platform. The data is provisionally shared with data scientists, who then use it to figure out things like potential triggers for health events. These insights are then condensed into actionable advice for patients.

Transportation: Autonomous transportation technology is another arena seeing deep learning advancements. In 2017, an autonomous truck developed by Chinese company TuSimple completed a 170-mile driverless test journey from San Diego, California to Yuma, Arizona—the training for which was fueled by deep learning algorithms. The company will deploy its fleet of autonomous trucks later this year.

Finance: In the finance and retail sectors, deep learning and natural language processing are fueling communication-related tasks. Customer service chatbots—and even “robo-advisors”—help customers manage investments. These prolific conversationalists can adapt to the users with whom they interact.

A Future of Thinking Machines

Thinking machines that can learn, reason, and adapt to input without any human hand-holding present exciting possibilities for technological progress across industries.

Thanks to deep learning neural networks, we’re on the cusp of realizing adaptive processes that will open up new possibilities from streamlining business operations for things like hiring, tech support, and customer relationship management—to altering the course of human healthcare.

But with AI autonomy comes great responsibility. Machines are, after all, only half of the equation when it comes to human-machine partnership, and it will be up to humans to find effective, responsible ways to put deep learning to work for us, one layer at a time.