Scope. Here we describe the activities and results of a research project whose aim is to advance AI by leveraging cognitive theories of how humans make decisions, with particular focus on the "thinking fast and slow theory" of Daniel Kahneman. The project is led by IBM Research in collaboration with several academic partners.

Thinking fast and slow (or System 1 and System 2). According to this theory of human decision making, humans make decisions by employing two main modalities, called thinking fast and slow. When we think fast, we make decision in an almost subconscious way, reacting to the environment stimuli, and generating decision out of past experience on similar conditions. When we think slow, instead, we carefully reason about the problem to be solved, we devote all our attention to the reasoning process, and usually we are more accurate in the result. We use the fast-thinking modality most of the times, when we tackle problems that are cognitively easy or very familiar to us. We instead use the slow thinking modality when we perceive the problem to be solved as cognitively difficult or not seen in our past experience.

Thinking fast and slow in AI. This research project aims to build AI-supported machines that can 1) make decisions with emergent behaviors similar to the human ones and 2) support human decision making through nudging and explanations. To achieve these goals, the team is designing and building a cognitive architecture (SOFAI, see below) to mimic these two broad modalities in a machine.

Neuro-symbolic AI. Our approach is one way to build neuro-symbolic AI, where the two approaches (neuro and symbolic) are kept separate and combined through the architecture layer. State-of-the-art AI has many successful and useful applications, but it is often also narrow in scope and demonstrates several limitations including the lack of deep understanding of information coming from data, the absence of common-sense reasoning, the difficulty in dealing with causality, and the inability to learn general concepts from few data. To solve a given problem instance, AI systems usually employ either machine learning or a logical reasoning approach. Each of these approaches has its strengths and weaknesses, but it is hoped that their combination that will bring about a new generation of advanced AI. It is indeed now recognized by the whole scientific AI community that end-to-end machine learning (or deep learning) approaches, although very successful in specific scenarios, cannot bring AI to the next level, and that we need to carefully and effectively combine both machine learning and reasoning techniques. The importance of this combination of methods can be seen in the many specific approaches to neuro-symbolic AI, as well as several workshops and other initiatives in this space.