Neuron type-specific translatomes in dorsal hippocampus during early memory consolidation.

Rethinking machine cognition - beyond data-driven learning: Human cognition excels at new tasks with minimal prior experience through reasoning, by dedicating cognitive resources, such as extended thinking time, to planning and strategising. This is enabled by dual-process human cognition: System I enables fast, intuitive decisions based on prior experience, while System II facilitates slow, deliberate reasoning, especially in unfamiliar scenarios. In the context of machines, reasoning or System II thinking is analogous to the ability to consume computational resources, such as inference-time compute, to improve performance, as opposed to using data for improvement, as noted in recent literature. Conventional deep learning systems lack this capacity to perform reasoning or System II thinking, relying heavily on extensive human-labelled data or repeated task exposure to perform well in familiar scenarios. While such data-driven experience yields strong performance within the scope of training data, these systems falter when labelled data is insufficient to represent new tasks. Here, we propose an algorithm which enables System II thinking and equips machines with the ability to improve performance with increasing thinking time (inference-time computation), rather than resorting to additional labelled data. Two systems of human cognition and their transition in psychology: Theories from cognitive psychology have observed that fast and intuitive System I thinking allows humans to make confident decisions for familiar tasks, despite unclear immediate justifications. In contrast, System II thinking consumes cognitive resources at the time of the decision-making. For instance, a beginner chess player can reason about potential moves, plan and develop complex strategies using only the game rules, despite limited gameplay experience. One key cognitive resource consumed in such processes is thinking time, and its mechanisms include comparison of hypothetical scenarios, considering long-term outcomes in greater detail, and breaking problems into simpler parts before solving them. Attention is another critical cognitive resource, which can be consumed through mechanisms such as filtering out irrelevant information (e.g., ignoring background speech while reading). Consuming these resources allows improvements in decision quality, which is an ability considered unique to humans. With experience, humans can transition tasks from slow, deliberate System II to much faster System I thinking e.g., the ability of experienced chess players to execute strategies more quickly and instinctively as their experience grows. The interplay between fast and slow thinking allows versatility in human cognition, which is capable of handling both familiar and unfamiliar scenarios.