Cutting edge developments in the AI, known as deep learning, are making significant headway toward helping us solve complex problems. Financial Sense Insider spoke with Nikhil Buduma, from Remedy Health and author of Fundamentals of Deep Learning: Designing Next-generation Machine Intelligence Algorithms. He discussed new capabilities in image processing, natural language processing, predicting disease and the implications these developments will have on national security.
See How Deep Learning Is Changing Healthcare for audio.
What’s Deep Learning?
Deep learning is an active area of research within the larger field of artificial intelligence, Buduma explained. It is essentially a methodological approach to creating software that emphasizes the power of artificial intelligence to analyze data and make decisions.
Broadly speaking, Buduma noted, the goal of artificial intelligence is to build computer systems that can reason in a way that meets or exceeds human-level capacity.
It is easiest to divide the history of artificial intelligence research into three segments, Buduma said, consisting of “old-school” AI approaches, classical machine learning approaches—or shallowing learning approaches—and deep learning approaches, which have become more popular over the past decade.
The old-school approach involved researchers hard coding all the rules for software to follow in order to explicitly make decisions or analyze data. This approach was cumbersome and did not produce many useful results, Buduma noted.
Within a decade, the field had advanced to the point that, instead of preprograming software with a hard set of rules, researchers built a framework or a model to describe the world and would then feed this information into a software system. Instead of trying to get the software to process raw data, this approach provided the software system with a base model to test approaches, make mistakes and craft a newer, updated and more accurate and useful internal model to make decisions and process data.
The problem with this approach, Buduma stated, was that it still relies on researchers to extract the most useful or relevant information to craft the model the algorithm uses to test its own internal model.
The next iteration allowed researchers to automate the process of feature selection and extraction so that software can take raw data and figure out what pieces of information are important for its internal model to function accurately under real-world conditions.
“These models often ended up being far more accurate than the handcrafted feature selection approaches,” Buduma said. “Over time, the software becomes better and better at making the right decisions. As computational power started to explode, deep learning began to emerge.”
An Explosion of Practical Applications
Modern computer hardware and the explosion in computing power have allowed researchers to build algorithms that seem to replicate the way in which neurons in human brains talk to each other, Buduma stated.
This framework for building artificial intelligence algorithms is essentially the basis for deep learning advances today, he added. The field has unlocked entirely new capabilities in image processing, including the ability for software to take the raw data of an image and intelligently interpret what it is in such a fashion that we are now seeing serious headway being made toward self-driving cars.
Researchers and developers have also made significant headway in natural language processing, which involves allowing software to understand, interpret and reproduce human speech.
“We're starting to see some pretty incredible headway in this space,” Buduma said. “We are already seeing some pretty serious implications of this from a national security perspective. For example, you can take a video and make it seem like somebody is saying something they never said. … I believe that for us to implement artificial intelligence responsibly and in a way that is truly good for the world, there are a lot of human issues that we have to explore at the same time.”
Cutting Edge Healthcare Applications
The promise of artificial intelligence will unlock new ways of thinking about problems in biology, healthcare and economics, Buduma stated.
This approach underlies his work with Remedy, he noted, where he and other researchers are crafting Sentinel, an AI-powered product designed to aid healthcare professionals in identifying patients who have an undiagnosed chronic disease.
Chronic diseases account for roughly 70% of healthcare spending, Buduma noted, and early detection of these pathologies could substantially reduce costs in the long run.
To accomplish this, Sentinel acts as an intermediary between medical staff who are not trained in diagnosis to be able to address patients and ask the right questions to lead doctors to identify pathologies that they otherwise might have missed.
Sentinel is designed to essentially give less knowledgeable healthcare staff a first-line screening tool to use with patients to figure out who has the greatest healthcare needs, in turn enabling doctors to head off chronic conditions as early as possible.
“Often, physicians just don't have enough time to address all the issues that might be existing and emergent in a particular patient,” Buduma said. “In a lot of ways, this frees up physicians to really grapple with more patients who need their attention, and not necessarily need to spend too much time on patients who are relatively healthy.”