Artificial Intelligence seems to inspire a wide range of thoughts and deliberations these days.
It may be partly the result of its continual success in fields like computer vision, images recognition, language processing, and other ostensibly creative activities. But among the factors contributing to the present hype are also the concerns about its possible impact on our future—which some foresee devastating enough as to pose an existential threat to humanity—and also the unsettling realization that technology is getting closer to replicating the very qualities that define who we are—namely intelligence, thinking, and perhaps consciousness. Although, it has to be said, we still know very little about these processes.
What we know is that AI is an ever-evolving discipline encompassing many different techniques; and because it is always expanding into new domains, describing it without getting lost in a convoluted web of nuances is rather difficult. So, without the presumption of offering exact or exhaustive definitions, I will simply mention a few characteristics, which can be useful to understand the actual potentials of this intriguing technology.
Of all the programming techniques that go under the name Artificial Intelligence, so-called Machine Learning algorithms are among the most advanced and sophisticated. They can classify images, predict trends, identify patterns, detect anomalies, and even imitate styles of painting and literature.
Their deployment often involves a training phase during which the program processes large amounts of data samples in order to infer relationships, or structures, that are mapped through a mathematical function, or model. Trained models are then used as the basis for generating outputs or statistical perditions about new data—typically by approximating some desired target, represented by the function.
In fact, machine learning includes several types of algorithms, each with different, partially overlapping, applications—and names such as supervised, unsupervised, deep learning, and reinforcement learning.
In the case of supervised learning, the training data normally consists of inputs that have been manually labeled—so to provide examples of correct input-output associations. When deployed to classify objects, for instance, the program implementing the algorithm tries to “learn” some general rule that can map inputs to their target value. Then it references the rule in order to assign each new data input to one of a given set of categories of objects.
For problems that involve unspecified or unlabeled inputs, unsupervised learning algorithms are best suited for discovering features and patterns in the data. For example, they can identify the key features of images and use them to group the inputs based on similarities.
Particularly effective for features extraction and for modeling complex data structures are deep learning algorithms. Their multi-layer architecture—made of hierarchical levels of representation called neural networks—allows the software to perform the series of subsequent computations through which more abstract attributes are “learned” by building on the inputs acquired from lower levels of abstraction.
Reinforcement learning algorithms, on the other hand, are designed to improve their performance in relation to executing certain actions in a given environment. As they interact with the environment—represented or simulated by a model—they acquire new information and respond with one of a given list of actions. At the same time, they are provided with reward inputs meant to incentivize those responses that can lead to an optimal performance in achieving a certain goal.
In any case, after they are deployed, programs are usually capable of performing their tasks autonomously; and the accuracy of their predictions tend to improve over time—as they learn from the data.
But even though they may not require specific feedbacks or human interventions, there are significant limitations to their self-sufficiency and learning potentials. To illustrate these limitations I will try to examine the relationships among four dimensions—the context, the problem, the solution, and the data.
The first key aspect of the problem-solution correlation is that the algorithm and the model are chosen based on our interpretation of the problem, and its structure. Then, we must also consider that the effectiveness of the algorithm—the sequence of steps completing the tasks which fulfill the solution—always depends on how well we have comprehended the problem.
Personal assistants, or virtual assistants—now commonly used by many computer systems to interact and communicate with us—are a good example of how unnecessary complexity can be introduced when the intrinsic nature and core priorities of the problem are not fully grasped.
As they try to help us with the scheduling of our meetings, for instance, these programs focus primarily on the conversations related to matching availabilities on our calendars. In doing so, they typically learn from our customary interactions and linguistic habits in order to improve their capacities for processing the natural language required to manage the task. However, the need for interpreting the many redundancies and ambiguities rooted in human language could be reduced, or even eliminated, if the salient purpose of the task were recognized as needing to have conversations converging toward a finalized meeting-event as quickly and efficiently as possible—realization which would lead to solutions that focus predominantly on structuring messages in such a way to narrow the range of possible options at each interaction—starting with invitations that state our own availability.
But it is even more important to appreciate that scheduling is a lesser aspect than ascertaining the value of the meeting. This should be fairly obvious if we consider how many meetings are cancelled or rescheduled after they have been meticulously scheduled. And yet, to establish how their purpose fits in our life we must constantly reassess their relevance, and the related situation, by answering two questions: “do I really need this meeting?” and if so: “is this the best time to meet?”
Which brings us to the next relationship, the one between solution and context. While AI solutions are fundamentally static—in the sense that the algorithm and the model remain essentially the same throughout their deployment—the situations in which they are implemented continue to evolve. So, even when we have picked the best possible algorithm, its ability to adapt is only within the scope of the task. And, eventually, when external conditions have changed enough to cause a shift in priorities, any algorithm loses its effectiveness.
Finally, we can look into the relationship between the problem and thedata. As mentioned, models are usually deployed after they have been successfully trained. However, the rules they have learned become obsolete when they deal with inputs that are overly different from the training data.
In general, the software turns ineffective when the data processed is no longer relevant or significant with respect to the task-problem; and also when the data changes too frequently—which makes it impractical to re-train the model.
In any case, developers are required to identify all the elements of information that can potentially affect the program’s estimations, and must be mindful of the fact that ML programs only learn the correlations that can be found in the given dataset. For instance, when the data contains only features that have no meaningful cause-effect relationship with the categories they are tasked to identify, they will learn to discriminate based on immaterial elements, and therefore reach wrong conclusions. Let’s say, they are required to categorize people as trustworthy or not, and the data consists only of images of their faces. In this case the algorithm would most certainly characterize a person as honest or dishonest based on some traits of its face—e.g. the color of the eyes.
Essentially, when the information provided is meaningless in the context of the task, or important information is omitted, ML learns the bias contained in the dataset.
Anyway, to understand how the above limitations may affect the potentials of Artificial Intelligence in the real world, we should consider that the learning skills of sophisticated organism like us are the result of millions of years of evolution which have enabled them to adapt and survive in highly unpredictable environments.
Natural learning developed as part of the continuous exploration that is necessary to uncover unknown facts, and to recognize what’s relevant and important in different situations. Rather than on intelligence it seems to rely on constant alertness and active attention. Always concentrating on immediate priorities; and at the same time ready to shift to new actions, and engage in the interactions required to solicit or acquire new sets of information.
Furthermore, to simply consider the present limitations of AI inherent to its lack of general intelligence is discounting the fact that, like all biological skills, intelligence could just be another form of specialization. And, after all, there may be no such thing as general intelligence. Instead, it is the unconstrained sense of awareness characteristic of all sentient beings which appears to be beyond the reach of current technologies. And, incidentally, as our environments have become exceedingly complex and dynamic, it also seems to have outpaced our own capacities.
Indeed, AI closely reflects our own limitations. As we have mentioned, it represents knowledge through models and functions which, like our words and ideas, are specific and intrinsically static. And very human-like is also its inability to explain its decisions and predictions.
So, to hope that highly sophisticated technologies like Artificial Intelligence will ever solve our fundamental problems seems downright ingenuous.
Nevertheless, if we could somehow find a way to slow things down and simplify our inner lives, we may still have a chance to recover and trust our innate sense of awareness. Then, as a result of being able to perceive what’s most relevant and important, our relationship with technology will be fully understood, and technology itself will have its place in enabling solutions that address our immediate problems and keep us focused on the unresolved ones.
Sandro Levati
Wideprocess © 2018