Over the last three years, few topics in enterprise technology (or, for that matter, in the larger culture) have taken on more significance than AI. It would be virtually impossible to avoid bumping into articles about AI, even if you tried. While the range of AI-related topics worth addressing is vast, let us narrow our scope to the interests of our enterprise technology audience, and to developments over the two years or so since we devoted an entire issue of The Doppler to AI.
First, a very short reminder about what we mean when we talk about AI in the context of business. Virtually all AI that is useful in business is “narrow”: AI trained to perform a specifically targeted task, such as translating between two languages, or detecting a particular type of cancer in X-rays. In fact, the more narrowly targeted the focus of an AI model, the more accurate it tends to be. Narrow AI contrasts with artificial general intelligence (AGI), whose goal is to create a general, human-like intelligence that is broadly applicable to a virtually unlimited range of cognitive tasks. While AGI is a subject of deep interest and vast potential, it is nearly irrelevant for business in the near- to mid-term, due to its relative immaturity.
Let us further point out that the terms “AI” and “machine learning (ML)” are defined differently by different people. Due in part to this confusion, these terms are often used interchangeably in business contexts. Rather than dive into this debate, in this article we will refer to AI and ML synonymously and collectively, while recognizing there are divergent opinions.
With that out of the way, what has changed in the world of business AI in the last couple of years, and what lessons should we be learning as a result?
By any measure, AI technologies have gotten considerably better performing and more accurate in the last few years. Because of AI’s direct value to business, a large number of researchers at leading universities and commercial firms are constantly innovating and iterating, developing new algorithms, frameworks, techniques and toolsets. Because increased accuracy and performance can translate into significant prestige and profits, these researchers are provided with plenty of resources. The net result is a remarkable rate of progress, as benchmarks for every conceivable way of measuring capabilities is continually bested. One of the best-known examples of this is the ImageNet Competition, in which a massive standardized set of images have been used over approximately a seven-year period as the basis for measuring the accuracy of AI-based image classifications. During that time, error rates dropped from approximately 28 percent to less than 3 percent, surpassing human performance. As rates for all manner of models improve, this translates into more and more business offerings which can be practically applied to solve problems, enhance experiences or create products.
Takeaway: Do not assume that because AI did not offer sufficient robustness and accuracy for an application two or three years ago, that is still the case. It is well worth your while to stay familiar with the current AI performance levels in various areas. You may find that a previously impractical concept is now perfectly feasible.
AI as a Service (AIaaS)
Historically, developing AI solutions has been a carefully crafted custom process, requiring significant investment in both highly skilled data scientists and specialized computing environments to make the most of their talents. For many instances, this is still the best path, but not in all cases. Over the last few years, predeveloped models for various business-friendly purposes, such as image recognition, speech recognition, language translation and text transcription, have been brought to the market by the major cloud service providers (CSPs). While Google was perhaps the first to offer a suite of these pre-engineered services, AWS and Microsoft Azure now also provide similar offerings. Some examples across vendors are the Google Translate API, AWS Rekognition and Azure Bot Service. These are offered as APIs that can be easily called from within any modern business application. To be clear, they do not provide a complete application to any organization wishing to build an AI solution. But if the AI needs of a solution fall into the well-defined capabilities of these standardized CSP offerings, highly effective AI-driven applications can be developed quickly and efficiently without having to create a full AI capability within your enterprise.
Takeaway: Be sure to examine your organization’s plans for AI projects, to check whether some could gain significant speed and efficiency in their development through AIaaS offerings.
Ethical AI: Bias and Explainability
Perhaps the biggest change related to the business use of AI over the last few years is the growth of awareness and concerns about the ethics of AI usage. AI-driven decisions affecting the consumers of a business’s products have the power to materially impact those consumers’ lives: What rates will they pay for insurance coverage? Who gets a mortgage, and who does not? Who gets hired, and who gets passed over?
The first ethics concern has to do with the bias that is implicit in the data used to train and develop AI models. For example, when an AI model is trained on visual data that under-represents women or people of color, the resultant model will be less accurate in recognizing members of those under-represented groups. When trained on data capturing previous hiring decisions, past biases can be learned and built into a model’s decision-making. It is important to note that these concerns are not just theoretical. Specific examples of such problems have been documented in AI tools, and in projects from companies such as IBM, Microsoft and Amazon, with serious potential consequences.
The good news is that researchers are now making significant progress in addressing these problems, through a combination of after-the-fact auditing of results to identify models’ resultant bias, as well as specific data handling and modeling techniques designed to minimize bias.
Historically, various data science algorithms used to make decisions were often directly transparent. This means that for any individual decision made using the algorithm, the reasons for that decision (approve/do not approve, put in this category vs. that category, etc.) could be directly confirmed by tracing the decision path. With some new classes of AI, such as deep learning, this is no longer the case. A given model can be highly accurate, yet still opaque as to confirming why any particular decision was made. These models are, in effect, black box decision-making machines. For many uses of deep learning, this lack of explainability is not really important, but for other use cases, it may be significant. Data protection laws in various jurisdictions are starting to require the ability to document the provenance of these decisions, with Europe’s GDPR a prime example. Aside from regulation, as the use of deep learning models expands, it is clear that more and more decisions made for consumers might be subject to scrutiny for liability reasons.
Because of this, the industry has developed the concept of explainable AI (XAI). This term is used to refer to the various approaches being developed to address the lack of explainability in AI-driven decisions. No currently known magic bullet exists to solve this problem, but there are many different approaches being pursued. However, there is an important consensus that as of today, explainability methods typically somewhat reduce the accuracy of the models, in favor of supporting a greater level of explainability. It is also necessary to understand that in these approaches, the explainability achieved is not absolute, but rather should be thought of as an increase in “confidence level.” As a result, it makes sense to selectively apply XAI approaches with the balance of accuracy vs. explainability that is appropriate to each use case.
Takeaway: Ethics surrounding the use of AI will only become more important to business over time. Any enterprise utilizing AI will need methods, processes and possibly roles, dedicated to proactively assessing whether there are ethical implications for AI-related projects, and addressing them accordingly.
A fairly recent development attracting attention in the field of AI is called “adversarial AI.” It turns out that in many cases, a model that is very good at visual or audio recognition can be fooled by what humans would consider extremely subtle alterations of the image or audio signal being analyzed. While these changes are not perceptible to a human, they can cause a model to decide that, for example, an image of a turtle is instead an image of a gun. Alterations intentionally introduced through sophisticated techniques to trick a perceptual AI are called adversarial AI, which, understandably, has generated significant concern. One can imagine numerous scenarios where tricking a model into producing specific errors could be disastrous, such as in autonomous vehicles, weapons detection or medical diagnostics.
A couple of points to note about the development of adversarial AI. First, it is generally agreed that as of this point in time, no usages of these techniques for malicious purposes have been discovered in the wild. All currently known examples of adversarial AI have been created by researchers in an effort to understand, anticipate and prepare defenses against future attacks. Secondly, researchers are simultaneously working on defenses against adversarial AI, including using another layer of AI specifically to detect adversarial AI!
Takeaway: Adversarial AI may not yet be an active concern for your company, but it is certain to be a growing focus over time. Keep watch on this topic and include it as an item to consider when evaluating risks associated with potential AI-related projects.
AI is a vast movement in industry and in our culture, so the topics discussed above are certainly not exhaustive, nor is the level of detail. But we do feel that an organization would be wise to ensure it is familiar with these developments, as it considers AI-based offerings and features. The benefits of AI are simply too great to ignore, plus, if you do not take advantage of those benefits, your competitors certainly will!