Every day, we make numerous decisions, whether in our personal lives or while working in businesses and institutions. Often, the consequences of a decision are not obvious. Furthermore, these consequences are stretched over time and, as such, interact with each other, reinforcing, weakening, or correcting one another. To act rationally or intelligently, we must analyze sequences of decisions.
We believe that the development of techniques for effective analysis and decision-making in sequences will lead to the creation of intelligent and autonomous systems. This will translate into many practical solutions, ranging from controlling robots or autonomous vehicles to multi-step decision or deductive procedures, such as mathematical proofs.
This topic is fascinating from a scientific point of view as it lies at the intersection of many machine learning disciplines. To investigate it properly, we need to develop planning techniques, reinforcement learning methods, and skilfully harness the benefits of deep learning and modern language models. We hope that over time, progress in these techniques and their skilful application will lead to the creation of highly versatile solutions that can be taken off the shelf and applied to a wide range of problems.
Can androids have common sense?
Philip K. Dick asks in the title of his cult book, “Do Androids Dream of Electric Sheep?” To begin with, we want to ask a simpler question: Can androids have common sense? Building an artificial equivalent of common sense can have a fundamental impact on the paradigm in which we build automated systems.
Automation and autonomy are major drivers of economic and technological progress. However, in many cases, automation is shallow; for example, robots on industrial assembly lines operate in a tightly defined and controlled environment, unlike the chaotic environment of everyday life (e.g., drones navigating urban spaces). How can we change this? We need a fundamental redefinition of control methods. They cannot be based on manually written scripts, as is the case on production lines, but must flexibly respond to poorly defined changes (or those difficult to manually scale).
Within the work of my team, we are developing methods and models that exhibit the characteristics of functional common sense equivalents. Key features include zero-shot generalization and adaptability without forgetting. Zero-shot generalization means the ability to operate in situations not tested during training, which can be considered an artificial equivalent of common sense. One of the most significant achievements in computer science in recent years is that such generalization can be achieved through careful model scaling and data volume. Adaptability without forgetting, in accordance with the established terminology of continual learning, involves the ability to rapidly acquire new skills (i.e., transferring knowledge from previous tasks) and not forgetting (many model skills will be used sporadically).