Human Feedback: a must in Machine Learning?
Today we live in an era where Artificial Intelligence (AI) is becoming more and more present in our daily lives. From voice assistants like Siri, Alexa, and Google, to self-driving cars, trains, and drones, AI is transforming the way we interact with the world around us. But what would happen if machine learning - the fundamental technology it is based on - were left to itself without any human input?
Before answering this question, it is important to understand what machine learning is and how it works. Simply put, machine learning is a set of programming techniques that allows computers to learn through "experience" (to use a human term), rather than being explicitly programmed. This is done through the use of algorithms that analyze large amounts of data and detect patterns, trends, and statistics. As a result, this information can be used to make predictions and automated decisions.
Machine learning is divided into two categories: supervised learning and unsupervised learning. In supervised learning, algorithms receive human input in the form of "labels" or "tags". These are essentially data that describe the content of an image, text, or other type of data. For example, in a photo of a dog, the label or tag could be "dog". These "markers" are inserted by human programming to make it easier for the machine learning algorithm to understand the information contained in a data point. In supervised learning, algorithms analyze the labeled element and then perform a logical inference process on an image or text, assisting the comparison processes of a system's models as they are deemed valid from the beginning. These inferences are used to train a neural network, which is then tested to determine, for example, if it can distinguish a picture of a dog from that of a cat or bird, as in the previous example.
Although the process may seem simple, labeling requires some experience on the programmer's part. Labels must be accurate, clear, and free from any possibility of data misinterpretation. Errors in this phase can affect the results of the machine learning algorithm itself and thus compromise the learning of the neural network. In addition, the labeling process can be laborious and time-consuming, especially when there is a large amount of data to manage. This is why tools are used for classification and indexing of various types of data (images, texts, mathematical formulas, etc.).
In contrast, in unsupervised learning, humans do not assign a meaning to each data point; rather, the algorithm autonomously analyzes information and tries to identify recurring patterns or specific characteristics without external labeling.
In the world of deep learning, which is the most advanced form of machine learning, algorithms are designed to simulate the structure of the human brain. These algorithms utilize various "layers" of algorithms (also called deep layers or hidden layers), consisting of numerous processing unit processes to analyze data. The results obtained through deep learning are built on a foundation of information that would require much time and be too complex for humans to derive, hence the automatic learning machine is used.
But what would happen if machine learning were left to itself without any human input or output? In the absence of sufficient verified data in each new cycle of "autophagy" (i.e. self-generated), the generative model would gradually decrease in quality and precision, a phenomenon known as Model Autophagy Disorder (MAD). As of now, with today's algorithms, computers are not able to determine with certainty whether the result of a processed data point yielded a positive or negative value, partly because deep learning programs are in any case able to analyze enormous amounts of information in very little time, however, without human guidance, they wouldn't know if the data is relevant, reliable, or correct. This lack of reliable verification means that the entire learning process would be inefficient because there would be no certainty about the quality of the acquired or processed information.
Moreover, automatic learning in the absence of human intervention would be extremely vulnerable to error. If a wrong path is taken in data analysis during the learning process, the system may continue to work without realizing the error and perpetuate and amplify it in subsequent elaborations. What AI is currently able to execute in an efficient manner is the interaction between models that would require humans months or years to develop, and human feedback for quick verification. This modus operandi is currently used by AI to combat long-term climate change, help with faster and targeted medical diagnoses, reporting market trends that the human mind would not be able to elaborate quickly.
It should be noted that we are still at the beginning, perhaps even in the pre-history, of AI. Even with GPT4 models (the most advanced technology with which AI currently interacts with humans), there are projects that are far ahead of these systems. The neuromorphic approach, decoding primate neural information, and hopeful creation of a neural model in a 1:1 relationship between man and machine, makes it impossible to predict the future of AI. For now, we are only simulating the human neural process, but what will happen when we have a complete emulation of it? It should be remembered that the nouns "simulation" and "emulation" are not synonyms...