What is the Most Impressive Task that AI can Accomplish Today?
Artificial Intelligence has already impressive achievements to present, for example, the ability of machines to ‘understand’ natural language, to ‘see’ and detect patterns, and to make sophisticated, real-time decisions.
But what would be the most impressive achievements of Artificial Intelligence so far? How is AI being leveraged across industries, from manufacturing to healthcare? How is AI helping the world become more connected?
Cynthia Rudin, Jordi Guitart, Jyotirmay Gadewadikar, Aruna Kolluru, and George Panou share their insights.
Cynthia Rudin
Professor of Computer Science — Duke University • USA
AI is helping society in numerous ways, though as we know, it is also being weaponized. Let’s consider some of the positive examples first.
IBM was a 2020 Edelman finalist competitor for their work on using Machine Learning to predict computer server failures. These servers form the backbone of the Internet, so it is important that they don’t fail. IBM used natural language processing to analyse trouble tickets and put them into a predictive model for server failures. It was an astonishing data science project that was quite technically difficult. As a second example, a charity called ‘Give Direct’ uses Machine Learning on satellite images to estimate which villages were the poorest. Give Direct can then target those villages to give out money.
My lab is working on automatically generated poetry and music. — Cynthia Rudin
There are a lot of wonderful, innovative, and beautiful examples like the two above. In my own projects, I’ve worked with the power company in New York City to predict power failures (namely fires and explosions on the underground electrical grid) using data from the time of Thomas Edison. I’ve also worked with neurologists to design models that can be used in the intensive care unit to predict whether a patient might have a brain-damaging seizure in the near future. As another example, my algorithms are used by police departments to determine which crimes might be part of a crime series, which is a set of crimes committed by the same people.
As I mentioned, I’m quite concerned lately about the weaponization of AI. AI does help us be more connected since it underpins the recommendation systems used in social media and search engines, but it also allows bad actors to propagate misinformation, leading to genocide, bullying, and violence in the real world. I am quite honestly concerned about what AI is doing to the future of the world and I do not see any clear way to stop it. There are other serious concerns about the interaction of AI with personal privacy. We now have facial recognition systems that are extremely accurate, even across different races. But to build them, a company scraped all our personal photos from the internet! Of course, we did not intend our biometric information to be used for this purpose. Unfortunately, there are almost no regulations on the collection of such data and the use of AI on it, so these problems will continue to occur.
In terms of the characteristic examples of AI, I think the most basic one is automated handwriting recognition. Whenever you put a check into an ATM to be deposited at the bank, or when you send a letter through the mail, an AI system reads your handwriting so that the amount does not need to be typed up manually. In the future (not quite yet) we will hopefully have reliable vision systems for self-driving cars, which could be another characteristic example of AI — after we get it to work! But we should be careful not to launch AI systems too early, in cases when it is dangerous to do so. Another characteristic example of AI is the algorithms that are used for playing games, such as freestyle chess and Jeopardy.
In the future, I hope that AI will help humans with creative tasks too. My lab is working on automatically generated poetry and music. It’s not as good as a human, but we are getting there. I find it easier to edit a computer-generated poem than to write a poem myself!
Cynthia Rudin is a professor at Duke University. Her goal is to design predictive models that are understandable to humans. She is the recipient of the 2022 Squirrel AI Award for Artificial Intelligence for the Benefit of Humanity (the “Nobel Prize of AI”).
Jordi Guitart, PhD
VP of Artificial Intelligence — Aizon • USA
AI is here to stay, and we must agree that the hype is now over according to the steady growth of investment in AI, spanning virtually all industries and sectors along with their digital transformation. The tasks AI can accomplish today are almost infinite as per our human unbounded capacity to think and dream. Still, the financial payoff has not achieved the spectacular promised results due to a mix of failures and successes almost all enterprises and verticals have experienced by adopting Artificial Intelligence.
Regrettably, many companies have discovered that having large amounts of data is not enough to secure successful outcomes from AI. And many more have paid high bills aiming at addressing wrongly stated problem statements, often too broadly defined. Without a delimited and well-defined purpose, no ROI should be possible and AI technology, whether Machine Learning or Deep Learning, is not an exception. But the resilience of first AI adopters will end up rewarding the big majority of them thanks to their unique learnings gathered across their AI transition journey.
The most impressive task that AI can accomplish is yet to come and it will deal with Human health. - Jordi Guitart, PhD
Currently, Machine Learning and Deep Learning are being intensively used in research with promising results, particularly Neural Networks as they dramatically accelerate computationally intensive processes, e.g., drug discovery[1]. Research benefits from a high degree of liberty that is hardly seen in other areas, thus being the ideal candidate for investing in powerful, though complex, Neural Networks architectures — only limited by budget allocation capacity.
On the other end, there are all those industries that are subject to heavy regulations for which interpretability of Deep Learning models is a problem. Thus, following the example of drug discovery, the adoption of Neural Networks for the subsequent manufacture of these new drugs is a no go right now, according to Good Manufacturing Practices regulation in Pharma and Biotechnology.
Moving away from highly regulated industries, AI has already demonstrated its feasibility to generate potential value from the use of pattern recognition (people’s faces, sign language, etc.) to fully autonomous equipment (self-driven cars, home vacuum cleaners, etc.) thanks to Computer Vision using Convolutional Neural Networks; and from voice-led instructions to activate appliances, select and execute tasks, to text mining or simultaneous translation using Natural Language Processing technology.
No doubt Deep Learning is the field of AI that delivers the closest Human-to-Machine and Machine-to-Machine interactions without simply mimicking human behaviour as Robotics Process Automation does from a collection of pre-set rules.
For sure, we can all look astonished at the performance of a fully autonomous car, but AI still has a lot to do to make the self-driving experience fully secure, and again regulation has a lot to say. Then maybe simple yet irrelevant things like telling Siri that you feel happy, and it selects one of your most-loved songs to play becomes impressive as we are absolutely unconscious of the swift connection we have built over Artificial Intelligence.
Still, to me, the most impressive task that AI can accomplish is yet to come and it will deal with Human health, and more precisely, the diagnosis and prognosis of patients’ life-threatening conditions. The challenge is paramount as there are no two equal persons in the world — yes, even identical twins have genetic differences[2].
Additionally, any intervention, such as any initiated treatment, severely biases patients’ health data. Hence, a predictive model aiming at diagnosing, for instance, the risk of metastases in breast cancer patients, is subject to an excessive number of factors (variables aka features) that dramatically reduces any subset of available patients to an insufficient number of samples that could be decently used for model training.
And here the word ‘decently’ has all connotations: legal, medical, social, cultural, ethical, and economical. How to deal with data privacy and data protection is just an example of issues that easily prompts our minds regarding patients and healthcare institutions, data controllers, and data processors. But these issues are all inherent to AI and they must be addressed in full alignment of all healthcare stakeholders, a common will, and a titanic effort.
Jordi Guitart, PhD has been recently appointed CEO of the Barcelona-based healthtech startup Science4Tech Solutions, coming from Aizon where he served as VP of AI. He is concurrently Adjunct Professor of Strategic Management at ESIC Business & Marketing School.
Jyotirmay Gadewadikar
Chief Engineer (Enterprise AI — Systems and Mission Analysis) — MITRE • USA
We all interact with AI products, for example, through an in-home voice-activated personal assistant or receiving a product recommendation on your favorite e-commerce website. It is abundantly clear that AI is omnipresent, driving innovation and influencing how businesses run and compete. AI has been known for nearly six decades. However, AI was thrust to the forefront of the new industrial revolution only recently. AI is the dominant driving force for modernizing a seemingly never-ending list of industries and functions such as communications, healthcare, media, education, audit, taxation, and operations. In short, the current and future use of AI in innovation can change how people live, work, play, and even think. Below are a few examples of what AI can accomplish today.
Natural Language Understanding and Generation. AI can process natural language, and it can read and comprehend what the intent is in documents, news articles, blogs, books, and emails. Furthermore, it can generate summary information by processing the text and label the text with topics. The technology can detect a question in text and group similar questions together. Note that humans have a lot of variation in speech, as someone can ask the same question differently.
Similarly, AI can use existing unstructured data to find answers to these questions. The ability to process unstructured text and respond to the text has further been helped by the technology to hear using state-of-the-art pattern recognition speech-to-text (STT) techniques. STT application programming interfaces can convert spoken words by humans into text in real-time.
Generating human-like speech is an excellent accomplishment of speech synthesis techniques. Speech synthesis is the artificial creation of human speech where a text to speech (TTS) system converts natural language text into speech. This synthesized speech can be created by concatenating pieces of recorded speech that are stored in a database, and the quality of these speech synthesizers has been becoming more and more similar to the human voice every year.
Text to speech technologies leverage Neural Network models to deliver a human-like, engaging, and personalized user experience. This capability of using existing unstructured databases and grouping questions and answers facilitates responding to those questions being asked in real-time.
So not only is AI understanding, but it also generates natural language and appropriately customizes responses into a specific answer based on the context by identifying the sentiment of the question. These capabilities result in creating an engaging user experience in achieving a great degree of containment in virtual voice assistants with interactions that mimic a human conversational style. The ability to be empathetic is critical. For example, “I would like to reschedule my flight” and “I just broke with my girlfriend, I’m rebooking my flight to leave early” would be different because the contexts in both these situations are different.
Advanced Inference Engines. AI facilitates incorporating rapid and advanced inferences, and inference engines are part of the decision systems that apply logical rules to the knowledge base to deduce new information. For example, inference engines based on Bayesian Belief Networks are used for automated breast cancer detection support tools. AI is a viable option for computer-aided detection by representing the relationships between diagnoses, physical findings, laboratory test results, and imaging study findings. In the example above, the work brings essential roles such as Radiologists, Image Processing Scientists, Database Specialists, and Applied Mathematicians on a common platform[3]. By exploiting conditional independencies entailed by influence chains, it is possible to represent extensive cause-effect relationships using little space. It is often possible to perform probabilistic inference among the features in an acceptable amount of time. The inference engines are critical in recommendation systems, such as suggesting which movie to watch next and deciding which price point will maximize the revenue.
Jyotirmay Gadewadikar received the Scientific Leadership award from the US Department of Homeland Security and is an Artificial Intelligence, Decision Science professional engaged in Strategy, Business, People Development, and Thought Leadership. He was previously Chief Product Officer of Deloitte’s Conversational AI Practice and System Design and Management Fellow at MIT.
Aruna Kolluru
Chief Technologist, AI, APJ — Dell Technologies • Australia
Artificial Intelligence is really transforming the way we live and work. From chatbots to recommendation engines we are surrounded by AI systems. The most impressive task that AI can accomplish is Computer Vision — a branch of Deep Learning which helps machines see and perceive the way humans do.
Since the first breakthrough with AlexNet in 2012, the accuracy of the models increased from 50% to 99%. Computer Vision applications are used in every industry making our world a smarter and safer place to live. Computer Vision is really augmenting humans. At Dell Technologies we have been working with customers across a variety of industries, leveraging AI to innovate and grow.
In healthcare, where 90% of all medical data is images, Computer Vision is playing a pivotal role in developing new-gen healthcare systems for diagnosing 2D and 3D images. AI algorithms are used from diagnosing brain tumours to determining which cancer treatment will work best for a patient, enabling doctors and hospitals to better analyze genetics, lifestyle, and past medical history along with the real-time data coming from wearables to provide personalized care to the patients.
In the automotive sector, autonomous vehicles leverage some of the most sophisticated AI technologies. An autonomous car must deal with complex analysis and decision-making for navigation, automatic braking, collision avoidance, lane change, parking assistance, and many other decisions. Self-driving cars capture data from their environment and feed it back to the AI — a loop called the perception action cycle. Then, AI makes decisions that enable self-driving cars to perform specific actions in the same environment. This is an iterative process and the more cycles the more accurate the AI model. Moreover, accumulated data from such cycles from multiple vehicles help produce intelligence that finally empowers every single vehicle.
Every industry has numerous use cases where AI can be applied. In the Retailer industry, companies are using Computer Vision for checkout, seamless transactions, and theft detection. Cities are using Computer Vision for surveillance to keep the population safe. In agriculture, Computer Vision is used to detect weeds and monitor the crop’s health and growth. Construction and manufacturing industries use AI to run operations more safely and efficiently, for instance, they apply computer vision for defect detection and preventive maintenance. Research also benefits from AI as it enables the optimization of costs and increases the productivity of complex research projects: Considering that the vast majority of research activities are led by data, Machine Learning models can simplify the whole process, allowing researchers to go through the cycles faster and improve their outcomes.
Aruna Kolluru is passionate about business, technology, and emerging technologies. She helps organizations to visualize the art of possible, architect, and build solutions with emerging technologies. Before joining EMC & Dell Technologies, Aruna held various senior roles for IBM & Oracle. She has over 21 years of experience in the industry on varying technologies including AI, Blockchain, Deep Learning, Computer Vision, IOT Big Data & Analytics, and Enterprise Architecture.
George Panou
Head of Innovation Centre — Eurobank • Greece
Artificial Intelligence has numerous applications in almost every discipline and every industry. Technologies like Artificial Neural Networks (ANNs) have contributed significantly to the advancement of AI and its major adoption by industries worldwide. However, Neural Networks are not new; they were first introduced back in 1944 by Warren McCullough and Walter Pitts, two University of Chicago researchers. The technology took off only around 2012 with the creation of ‘AlexNet’ developed by Alex Krizhevsky, Geoffrey Hinton, and Ilya Sutskever which won the ImageNet Large Scale Visual Recognition Challenge. Actually, before that, the scientific community was ostentatiously ignoring ANNs and the whole concept.
But what led to the wide adoption and exponential development of ANNs and why it is so important to the domain of AI? That is mainly because of the exponential growth of computational power with GPUs, faster computer networks, and more efficient and reusable code libraries. Another reason for the growth of ANNs adoption is the huge amount of data generated by users and electronic devices (IoT, smartphones) giving the possibility to train ANNs to perform various tasks. Cisco reported that in 2021 there are 27.1 Billion devices 3.5 per capita creating 33 Zettabytes, predicted to reach a mind-boggling 175ZB by 2025.
The fully autonomous car is a classic example of advanced AI. - George Panou
ANNs have helped in many different industries. In healthcare, for example, ANNs are used to predict Alzheimer’s or Parkinson’s disease by predicting the disorders with accurate diagnosis at early stages. IBM has used its AI Watson to detect cancer from MRIs and correlate clinical trials and medicines specific for the type of cancer through a huge dataset helping save the lives of patients. Actually, by digitizing the neurons of the brain and then training the ANNs properly with the right data you may end up finding correlations between dependent and independent variables that humans may fail to correlate. In the food industry, ANNs can help identify poisonous ingredients or test the quality of food products through a ‘digital nose’ — even in cases where microbes could develop and go unnoticed by conventional tools and technologies.
In the automotive industry, the fully autonomous car is a classic example of advanced AI. Data are continuously gathered through sensors that capture the behavior of the driver, the road conditions, and the objects in the environment in order to provide an effortless and safer driving experience. Although the technology has advanced, you cannot simply program a car to drive itself from point A to B with no risk as there are too many unpredicted parameters during the ride. You must have an evolving unsupervised deep learning system that can autocorrect itself and adapt to unpredictable conditions. Tesla cars for example, by using ANNs and sensor devices, might be able to ‘see’ an accident before it happens; they even update their firmware on the fly to resolve urgent problems. Of course, there is still work to be done as AI is lacking ethical decision-making. For example, in a critical situation, would the car select to hit the pedestrians or try to avoid them by crashing on a wall and putting the lives of its passengers at risk? Google’s Waymo is also disrupting the automobile industry by creating new ecosystems and platforms for self-driving cars.
George Panou is the Head of the Innovation Centre at Eurobank, leading Digital Innovation and Digital Transformation for more than 20 years working with Financial Institutions-Banks, Large Enterprises, and Public Sector in Greece & EMEA.
[1] “Deep Learning Driven Drug Discovery: Tackling Severe Acute Respiratory Syndrome Coronavirus 2”. Front. Microbiol., 28 October 2021
[2] H. Jonsson et al., “Differences between germline genomes of monozygotic twins”
[3] Gadewadikar et al., “Exploring Bayesian networks for automated breast cancer detection,” IEEE Southeastcon 2009, 2009, pp. 153–157, doi: 10.1109/SECON.2009.5174067.
Excerpt from 60 Leaders on AI (2022) — the book that brings together unique insights on the topic of Artificial Intelligence — 230 pages presenting the latest technological advances along with business, societal, and ethical aspects of AI. Created and distributed on principles of open collaboration and knowledge sharing: Created by many, offered to all; at no cost.
