Personalized movies, TV shows, and music accompaniment suggestions are made with powerful AI algorithms, which increase viewer engagement and retention.
Carlos Pantsios*
Artificial intelligence (AI) is a technology, whose theoretical bases began to be developed in the 1950s (John McCarthy, Dartmouth 1953), and which is emerging as a new transformative force in today's technological world. It has been regarded as "the science and engineering to develop intelligent machines," or "the activity devoted to making machines intelligent."
Such definitions raised questions about what "intelligence" means. The word implies a philosophical-cultural curiosity, with intelligence as a characteristic of capacity that apparently gives human beings a preponderant place over other forms of life.
AI suggests here that it can be simulated by technological means, and that it remains different from the "natural" or, possibly, "real" type of intelligence of human beings.
AI is made up of a complex robotic system that learns and adapts in a unique way, using a large amount of external data for training.
The heart of the AI system is an Artificial Neural Network (ANN), which can be considered a highly specialized artificial "brain", capable of improving decision-making processes, revolutionizing today's industrial world. Its design is based on a very complex architecture, made up of billions of artificial neurons (NA) interconnected with each other, sometimes provided with non-linear transfer functions, and which are grouped into layers that learn and adapt through the external data supplied to the (ANN).
The potential for AI to process large amounts of data, and produce an inside look in just a few seconds, is extraordinary. If this technology is properly controlled, it will be able to deliver the best of expectations over many application sectors in the world of today, and tomorrow. However, this ability can and, in fact, obscures the mechanisms behind the decisions that (AI) delivers on its way out, giving rise to challenges that demand the attention and creativity of human beings to solve the problems that (AI) hides. The rapid adoption of AI-powered solutions in the testing of all types of software, healthcare, and fintech, among others, has raised concerns about transparency, accountability, and trust. The more AI progresses, in terms of its capacity and complexity, the more complicated the problems that lie behind the technological, ethical and social dimensions linked to it become.
The increasing availability of data monitoring and advances in computing mean that AI and machine learning (ML) have become key tools for network operators, for example, to automate complex network management. The solutions offered by (AI) systems have demonstrated supra-human capabilities in solving a wide spectrum of problems in real life. Among them, the approach to very challenging management problems stands out, such as fault detection, identification of fault causes, failure prediction and location of failures, producing the widespread adoption of (AI) in the industry.
However, AI-based solutions have often sparked skepticism among practitioners of this technology, as they are often used as "black boxes". This black-box nature of AI often creates challenges in understanding how decisions are made, leading to skepticism among developers, regulators, and end-users. This lack of interoperability is particularly worrying in applications where the stakes are high, such as AI bugs in testing all kinds of software, causing system failures, or where algorithms for finance can result in unfair practices.
The problem raised occurs in almost every domain, including the media and entertainment industry. Here, AI has a wide spectrum of effects, including marketing techniques, efficiency, personalization, and content creation. With algorithms that evaluate massive amounts of data to create content specifically targeted for audiences, content production has become more dynamic and focused. The entertainment experience now includes voice recognition, chatbots, personal assistants, which allow hands-free operation, and better user assistance.
Personalized movies, TV shows, and music accompaniment suggestions are made with powerful AI algorithms, which increase viewer engagement and retention.
Likewise, in post-production and in the editing of videos made with automation with (AI), productivity and cost reduction have been greatly increased. All this, and much more, makes more relevant the great concern that exists in blindly trusting AI-based systems, as the crucial elements of the 21st century to reform the industry.
It should be noted here that the internal reasoning of the (AI) models is not exposed, preventing the humans who operate them from trusting it and, even more, not understanding how it works. For this reason, AI systems require permanent assistance from trained human beings, who ensure their logical response.
(AI) is transforming today's society from a legal system to a corporate environment. She promises unprecedented efficiency, inner vision, and autonomy. However, the more capable and complex (AI) becomes, the more complicated the answer to the question: is (AI) reliable? If not, how can we ensure that its results can be trusted? AI systems, especially deep learning models, can fail due to several inherent causes:
- AI is based on statistics and not on something deterministic: AI models (such as neural networks) do not understand reality – they identify patterns from the data provided and extrapolate from there. If these patterns are incorrect, incomplete, or biased, the prediction may be wrong.
- Characteristics of the data supplied: If the training data is of poor quality, biased, or incomplete, the quality of the predictions is directly affected.
- Overloaded/Underloaded: An AI model could memorize irrelevant noise (data overload) to fail to capture essential patterns (data underloading), predictions can be inaccurate.
- Exceptional cases: AI models are complicated by rare, novel or out-of-distribution situations, for which they were not trained.
- "Black box" nature: Complex models such as deep neural networks (numerous hidden layers) possess trillions of parameters that are not remotely transparent, intuitively, to humans. The models then work blindly, like black boxes.
The predictive processes performed by AI models are not well known: Especially in deep learning models and complex (ML) models, the following problems arise, which have not been solved so far:
- The way models combine features and parameters to achieve their predictions is distributed across layers of abstract representations.
- Unlike the typically human, rule-based design of a system, there is no explicit logical sequence that can be traced from the input to the output of the (AI) model.
Usually, when model developers (AIs) cannot clearly explain why a model predicted something specific, they respond that statistically the model responds that way with similar inputs.
For some time now, "eXplainable Artificial Intelligence" (XIA) has been created, a pivotal technological solution that fills the gap, offering methods of interpretation of the models (AI) without compromising their operating qualities. The aim is to transform the black box into a "transparent box".
In the testing of software operated by (AI), the (XIA) helps to develop fault debugging tests, understand bias in the models, improve test coverage by providing convincing human explanations.
Through (XIA), practitioners will be able to understand how the (AI) model can arrive at decisions, and even more, know the reasons behind its decisions and, therefore, know if the decisions made are based on correct or wrong reasoning. All of this can generate crucial insight into how the (AI) model can be improved and refined.
By integrating (XIA) techniques – such as SHAP, LIME, and rule-based explanations – organizations of all types will be able to mitigate these risks while maintaining the quality of operation of the (IA).
In conclusion, the challenge here is to develop, implement and adopt (XIA) techniques in (AI) models that seek to fill the gap by providing approximations or interpretations of what happens inside the models, in order to make (AI) safer, fairer and more reliable.
*Text written by Carlos Pantsios Markhauser, PhD, IEEE. he is a Telecommunications Engineer, PhD in telecommunications electronics, Master in Communications from the Simón Bolívar University, with a Specialization in Satellite Telecommunications and Networks from The George Washington University - School of Engineering & Applied Science, Specialization in Digital Telecommunications from the University of Colorado Boulder. He works as a postgraduate professor in the telecommunications schools at the Simón Bolívar University and Andrés Bello Catholic University. In addition to being a professional consultant in TV projects based in Argentina.
