Source: Deep Learning on Medium
Applied AI — Navigating Challenges Through Critical Thinking
Perhaps the most (mis)used term today across industries, one would have to agree, is Artificial Intelligence (or Applied Artificial Intelligence, more specifically). Although AI seems to be percolating in almost every space, there are some important question that perhaps not many are asking about their implementation of AI: Is this a responsible use of AI? Do we have the right data (from both quality and quantity perspectives) to begin with for solving a problem? Is the data sufficient? Is my implementation bias-proof? Are the results statistically significant? Or have we just jumped the gun to join the mad race of being associated with the trend?
As awareness builds around AI, there have been developments on the need to use critical thinking for building frameworks for guiding development of AI. One such initiative is “The Montreal Declaration for the Responsible Development of Artificial Intelligence” that provides a set of ethical guidelines for the development of artificial intelligence. In one of my own research to be presented at NeurIPS-2019 in Vancouver and recently published at The Lancet (link), we have used AI to identify young migrants in Central Asia and Africa who might be at risk of being sexually exploited. In this research, we have delineated the process that has much more to do than just AI implementation. For instance, given the problem is imminent and needs solving NOW, we have created a webapp for helping migrants by making them aware about their rights and providing safety tips to follow while migrants are on the move. This app also acts as a source of gathering data from the migrants that could be processed using AI to come up with the patterns and insights for making migration safer. Finally, given there is no readily available dataset of migration, we have hypothesized a solution based on a manully curated dataset. An important aspect of the study is that it does not only rely on some pseduo AI based solution, but explores different dimensions for addressing problems associated with migration.
What is Applied AI?
Applied AI is the application of Artificial Intelligence to enable “human-like” decision-making in myriad fields. It can also be viewed as a metamorphosis of AI from research to practice. The advance of AI has put a great value on applied AI that can quickly convert years of research into software nuggets that can be plugged in real-world systems to make them smarter.
Applied AI is different from Generalized AI in its specificity of application. Generalized AI, as per Artificial General Intelligence Society is “an emerging field aiming at building of ‘thinking machines’; that is, general-purpose systems with intelligence comparable to that of the human mind (and perhaps ultimately well beyond human general intelligence)”. Applied AI, on the other hand, is aimed at solving a problem at hand, making the solution more precise, contextual, intelligent, intuitive, or all of them combined.
What is Critical Thinking (CT)?
According to the Foundation for Critical Thinking, a leading non-profit organization that promotes essential change in education and society through the cultivation of fair-minded critical thinking, “CT is that mode of thinking — about any subject, content, or problem — in which the thinker improves the quality of his or her thinking by skillfully analyzing, assessing, and reconstructing it. CT is self-directed, self-disciplined, self-monitored, and self-corrective thinking. It presupposes assent to rigorous standards of excellence and mindful command of their use. It entails effective communication and problem-solving abilities, as well as a commitment to overcome our native egocentrism and sociocentrism.”
As we go about our lives, we sometimes unwittingly cave into cognitive biases like confirmation bias and selection bias. CT helps cut through the irrelevant and rid these biases. CT promotes an ideology of being aware of the problem by gathering and assessing relevant information, appreciating and formulating the problem statement after understanding its complexity and assumptions, reaching a well-thought conclusion questioning each facet of the solution, and thinking about alternatives with an impartial point of view.
CT is becoming increasingly relevant in the current context. According to a The Wall Street Journal article titled “Bosses Seek ‘Critical Thinking,’ but What Is That?”, frequency of the skill CT in job postings have doubled since 2009. The site, indeed.com, which collates job ads from several sources, found that more than 21,000 health-care and 6,700 management postings contained some reference to the skill.
In his book, “Thinking, Fast and Slow”, Kahneman explains the two systems that drive the way we think. System 1 is fast, intuitive, and emotional; System 2 is slower, more deliberative, and more logical. He argues “The impact of loss aversion and overconfidence on corporate strategies, the difficulties of predicting what will make us happy in the future, the challenges of properly framing risks at work and at home, the profound effect of cognitive biases on everything from playing the stock market to planning the next vacation — each of these can be understood only by knowing how the two systems work together to shape our judgments and decisions.” System 2, here, represents CT and it is not an inhibitor but it rather helps you form an inquisitive outlook, one of an explorer, while understanding a problem before attempting to solve it. It is the System 2 that argues for the need of logic and systematic approach in solving a problem.
How can CT help?
Given the vastness of data and multitudinous researches conducted across domains, it is easy to get lost in the deluge of information that exists. Different disciplines like Statistics can help, but the essential part is to ask the right questions and to be aware of the pitfalls and alternatives, which is where CT might help because it facilitates in identifying a. when what is relevant b. missing information (data points) c. when data is insufficient and when the result is statistically significant d. alternatives for the proposed solution.
Another instance of abundance is while reporting the results of a machine learning experiment, there are many statistics to choose from, namely, True Positive, False Positive, True Negative, False Negative, a combination of these like Precision and Sensitivity / Recall, Specificity, F1 score and what not. However, one cannot substitute the other and in fact, some of them can be very misleading or misrepresent the results. For instance, there can be a case when, even though, accuracy is 99%, still, not even a single instance of one of the two classes of a boolean problem is predicted correct. As weird as it might sound, it is true, and this is not the only example. The list goes on. CT would help you choose the relevant metric(s) for you for your problem. Talking about spurious accuracy metrics, I would recommend this site that lists several weird cases of spurious correlations.
CT models in real-world problem solving
There are many models, in academia and industry alike, that implement Design Thinking, of which CT is an integral part of. For instance, Google came up with its Design Sprints I (shown in the figure below) and II, focussed on Design Thinking, as guiding principles for its product development. Stanford has a school, Hasso Plattner Institute of Design, which is a hub for innovation, collaboration and creativity, dedicated to design thinking and its application in different domains. Many top universities teach courses on Critical Thinking and Design Thinking, some of which are also among the top trending courses on online platforms like Coursera and Udemy.
Although applied AI might have started to emerge as a shining iceberg tip, one must also heed to the fact that over 90% of an iceberg’s volume (and mass) is underwater. There is a lot of uncertainty that has to be tread carefully. In such a situation, abandoning CT and throwing caution to the wind might not be our best option forth.