When AI Becomes Intuitive
How faster, deeper, more intelligent AI is changing business
A recent article from The New York Timesreports that for the first time, an AI system was able to pass both an 8thgrade and 12thgrade science test. Aristo, as the system is called, demonstrates the progress made in developing technology that can comprehend language and mimic the logic, reason and decision-making of the human brain.
I sat down with Rashed Haq, Global Head of AI & Data at Publicis Sapient, to discuss the business implications of systems like Aristo and how advancements in AI are primed to disrupt virtually every industry.
Natural language processing expedites information-gathering and analysis
Systems like Aristo make use of natural language processing (NLP), which incorporates principles of linguistics, computer science, information engineering, and artificial intelligence to train computers to process and analyze significant amounts of information. In recent years, major strides have been made to build language models that conduct data mining, but also simulate very rudimentary types of reasoning. This allows these programs to understand the material and present them in a meaningful way to humans.
According to Haq, one area where this technology can have major implications is healthcare, as the amount of medical information doubles every 73 days. Doctors and other health professionals can use this technology to collect, analyze and summarize relevant information. It can also be used to process test results or compare medical records, lab results or clinical trial data. For example, a recent study by Dana-Farber Cancer Institute used deep NLP models to review more than 14,000 medical records. The models provided accurate curation that matched those annotations provided by humans—presumably at a fraction of the time and cost.
Looking to the future, this technology is likely to have the same effect on any type of research, from financial asset management and investment research to product development or food safety.
Generative adversarial networks reduce research and improve design
While AI models typically work to identify patterns, generative adversarial networks (GANs) work in the opposite way, pitting two back-to-back models against one another—hence the term “adversarial.” Using this system, one model creates an output, which can be an image, music, speech or written words, while the other determines if the output of the generator is distinguishable from a known good result (e.g.a photo taken with a camera or a sentence written by a human).As the first model continues to iterate, the discriminator tests validity—with the former looking for ways to outsmart the latter. Eventually, the first model produces an output that the discriminator cannotdistinguish between “real” and manufactured.
Haq shares an example from Insilico Medicine to demonstrate how the GAN model helped radically shorten the research phase of the drug-development process. Using a GAN model, the company was able to test 30,000 different compounds and determine a precise molecule that could be used for animal and human testing—shortening the research timeline from 2-3 years to just 46 days.
Obviously, this is a dramatic reduction in time. However, this GAN also has significant cost implications, as companies in various industries can avoid lengthy research periods and help scientists work more productively while producing better results. For example, Autodesk, a company that manufactures parts for cars and planes, uses this technology to create new design components. Partnering with Airbus, they developed a 3-D cabin partition that was stronger than the original, but half the weight.
Deep reinforcement learning improves sequential decision making
One of the most recent developments in AI is deep reinforcement learning (DRL). This model uses trial and error to fulfill a set objective within a defined environment. Initially, DRL was made popular by gaming applications, such as AlphaGo and AlphaZero, which were developed by Google DeepMind.
However, DRL also has powerful business implications. At its core, this technology can assist organizations in process configuration and sequential decision making. For example, by analyzing and optimizing sequences, DRL can determine optimal traffic configurations, dictating precisely when and where to change signals. This model can also be used to optimize resources at data centers and control network traffic. Looking to an effective use case, Google deployed a program similar to AlphaGo to reduce its data center cooling bill by 40 percent.
From Artificial to Real: How companies can begin their advanced AI journey
An expert with a combined 20 years of experience in the fields of AI and analytics, Haq works with organizations from a variety of industries, including financial services, retail, healthcare, energy and automotive, to help them understand how to leverage this technology. As a first step, he works with clients to brainstorm AI use cases and develop a business case. They then map the use cases with different algorithms and models. Finally, they lay out what additional data is needed to begin the project.
Next Starts Now