“We’re already seeing AI’s potential to basically revolutionize the world,” said Prof. Fredric M. Ham, interviewed in IMR’s May/June 2018 issue (“Artiﬁcial Intelligence—Where We Are,” pp.8-14). Prof. Ham is Emeritus Professor at Florida Institute of Technology, Co-Founder and Chief Innovations Ofﬁcer, Proteus AI Labs, Past President and Fellow of the International Neural Network Society, a Life Fellow of the Institute of Electrical and Electronics Engineers, and a Fellow of the International Society for Optics and Photonics.
Prof. Ham’s 2018 interview highlighted a broad array of AI applications, real and potential. His article below tightens the focus to four emerging areas in which companies are actually implementing AI applications and not simply conceptualizing their market prospects. These areas include: workplace stafﬁng, robo-advisors for asset allocation, contract governance and legal discovery, and healthcare.
Although AI virtual assistants, self-driving cars, intelligent targeted marketing, ambient computing, and chat bots are very interesting, they are more developed than these four, and as such mentioned only brieﬂy.
Some people, especially those who are not technically savvy, may have a misguided conception of AI. They seem to believe everything will be run and controlled by robots and intelligent machines such that humans have no input to the functions performed, or even worse no purpose in society. This is not, nor should it be, the case, as many of these systems are, ﬁrst of all, machines—they are only human-like.
Prof. Bruno Olshausen (University of California, Berkeley) elaborated on this in a letter to The Wall Street Journal. He wrote,“The real problem is the technologists who are trying to sell us a vision premised on grandiose, unsubstantiated claims about the capabilities of modern AI systems.”
So according to Prof. Olshausen, many technologists are to blame for all the hype that can potentially lead a novice, or possibly anyone, to draw incorrect conclusions. Prof. Olshausen has also said that we’re nowhere near understanding how humans possess common-sense knowledge (an important element missing in AI systems) about the world that we’ve gained through evolution, let alone how to incorporate this into an AI system.
Real and Great AI Strides
There have been, however, great strides in AI that are real and have made many people’s lives all the better for it. For example, AI Virtual Assistants, such as, Amazon’s Echo-Alexa and Google Home, have given the family access to information fast and accurately. For example, “Alexa, what is the Dow Jones at right now?” or “Alexa, what will the weather be at 2 pm today?” These devices can collect the personal information about what your interests are and, for example, offer targeted suggestions for related purchases (for the Echo, obviously for products offered on Amazon).
However, a single device does not have to operate autonomously when interacting with us; what about a single device interacting with other devices? The term IoT (Internet of Things) refers to the interconnected array of computing devices, objects and machines (Amazon Echo, “smart-home” devices, wearable devices, and more). These devices have distinctive identiﬁers capable of transferring data over a network.
This transference of information is without human-to-computer or human- to-human interaction, and when it goes beyond simply collecting and transferring information to actually using the information in an intelligent manner, this cognitive computing is referred to as ambient computing. So, in a nutshell, ambient computing consists of many of our devices basically “talking” to one another in a manner that they become extensions of one another, rather than discrete autonomous devices.
An area of rapid AI advancement is emerging in the workplace for personnel stafﬁng, speciﬁcally in the interview process for strategic hiring. Employee referrals have always been encouraged in the corporate culture, with the person making the referral rewarded with a ﬁnancial bonus many times for a successful hire. However, a fundamental problem with this is the high risk of bias.
Biases by recruiters (internal and external to the company) and hiring managers can take on the form of a candidate’s “likability” and general personality traits. Although these are important, they have been found in many cases to take center stage to what is truly important: the candidate’s qualiﬁcations and skill sets for a particular position.
AI can take the bias out of the hiring process and offer an objective evaluation of a candidate’s potential for a successful career that meets the speciﬁc needs and requirements of a deﬁned employment position. AI can analyze huge amounts of data and make decisions based on this analysis that can expeditiously offer hiring recommendations for the strongest candidate from search pools for deﬁned positions.
There is a related application for parents looking for better information on babysitters for their children. The online service Predictim (www. predictim.com) offers additional information beyond the standard criminal background checks, parent reviews, and face-to-face interviews with the prospective sitter. Using AI, Predictim can assess personality traits based on the person’s posts on Twitter, Instagram and Facebook, for example. A speciﬁc risk rating can reveal negative personality traits such as harassment, bullying, disrespectfulness, and in general a possible “bad attitude.”
The primary function of a robo-advisor is “asset allocation,” which is basically determining how much of your money should be in stocks, bonds and cash. This is typically based on your age and a deﬁned risk proﬁle, similarly to what a human ﬁnancial advisor does. But a robo-advisor offers a more in-depth breakdown of the allocations, for example: equity types—international versus domestic and the size; and different bond types—corporate, municipal and varying durations.
Just like a human ﬁnancial advisor, a robo-advisor can assist in rebalancing your asset allocation based on events that occur in the market and offer tax-loss harvesting suggestions. The major difference is that a robo-advisor is less expensive than a human ﬁnancial advisor.
What does AI have to do with robo-advisors? A robo-advisor (sometimes called automated investing or online investment advisor) uses advanced software speciﬁcally designed to build and manage a client’s investment portfolio. AI is only now starting to be used to power robo- advisors, which are essentially high-level active management systems.
Feeding the AI powered robo-advisor “big data” can produce a different allocation strategy that would presumably perform better than the classic robo-advisors.
There are companies (Responsive, http://www.responsive.ai), for example) now engaging in the business of AI-powered robo-advisors. This emerging AI area seems poised to disrupt the world of ﬁnancial investing.
Contract Governance and Legal Discovery
Like many other businesses, the legal world is drowning in mounds upon mounds of documents and volumes of data. By 2020 it is estimated there will be 44 zettabytes of data in the world that must be carefully analyzed, categorized, scrutinized, and evaluated. Speciﬁcally, contracts for important clients must be precisely crafted to minimize the client’s business risk.
AI can allay concerns over possible inappropriate handling of the masses of documents and data, and assist in appropriate contract life-cycle management, thereby easing the workload of the legal and contract professionals. IBM’s Watson Compare & Comply is one such solution to this problem.
Watson is trained on contract semantics, statements of work, procurement agreements, technical documents, etc. As such, it can understand the semantics of contracts, interpret PDFs, break down details of tables, perform general contract analysis, and make natural language queries (NLQ), thus mitigating risk and reducing cost and complexity.
Legal discovery is a related area. In the course of preparing for litigation, a law ﬁrm can spend an overwhelming amount of time collecting and poring over extreme quantities of documents, including, contracts, email messages, photographs, bank statements, product design documents, marketing material, patents, Facebook pages, Twitter accounts, etc. Even with a highly skilled team of legal professionals, this task is daunting and some key elements pertaining to the legal case could be overlooked. AI can provide a systematic means of clustering documents into relevant groups and assist the legal team in extracting content that is relevant to the ﬁrm’s speciﬁc case.
AI is currently being developed for medical imaging diagnostics, drug discovery, legacy of health records, and general medical diagnoses, to name a few. Here only medical imaging diagnostics will be addressed. This should be considered an emerging ﬁeld and by no means a mature one.
The need for AI in radiology is ever more increasing with the radiologist-to- scanned images ratio declining rapidly. Moreover, in an attempt to ﬁne- tune medical images for better image quality, the resolution of today’s scanners continues to improve, resulting in greater volumes of data. This will result in radiology transitioning from qualitative interpretations of scans to a quantitative discipline (that is, radiomics: deriving clinically pertinent information from vast amounts of data).
The main reason for applying AI to radiology is to speed up radiological workﬂows, minimize decision errors, and assist rather than replace radiologists. The current emphasis is on cancer diagnosis using AI where the output is yes (malignant)/no (benign).
Another application is the segmenting or marking of deﬁned areas in images again for cancer diagnosis and analysis, and also for planning treatments and monitoring conditions.
Deep learning neural networks are probably the most successful recent AI approach for determining image features by training them on vast amounts of image data. As these deep learning methods require a large amount of labeled or marked images (that is, knowing the content), the challenge becomes the accumulation of these labeled images.
Another deep learning challenge in digital pathology is the sheer dimensionality associated with high-resolution images (some can exceed 50,000 by 50,000 pixels, whereas deep neural networks can typically accommodate images no larger than 300 by 300 pixels). This should be an area of concentration in order to further develop the applicability of deep learning neural networks to medical image diagnosis.
The application of AI has exploded in the past few years, and I anticipate this continuing for many more years as there are many areas that can beneﬁt from AI. However, there should be careful scrutiny of such applications in order to ensure that the results are in fact true and meaningful and not marketing hype to merely beneﬁt corporate PR in the marketplace.
A possibly prudent venture would be to create an oversight organization consisting of AI experts who can evaluate these technologies and provide consumers with a deﬁned rating of the efﬁcacy of these new AI technologies.