Amortized Inference in Probabilistic Reasoning
Recent studies of probabilistic reasoning have postulated general-purpose inference algorithms that can be used to answer arbitrary queries.These algorithms are memoryless, in the sense that each query is processed independently, without reuse of earlier computation.We argue that the brain operates in the setting of amortized inference, where numerous related queries must be answered (e.g., recognizing a scene from multiple viewpoints); in this setting, memoryless algorithms can be computationally wasteful.We propose a simple form of flexible reuse, according to which shared inferences are cached and composed together to answer new queries.We present experimental evidence that humans exploit this form of reuse: the answer to a complex query can be systematically predicted from a person's response to a simpler query if the simpler query was presented first and entails a sub-inference (i.e., a sub-component of the more complex query).People are also faster at answering a complex query when it is preceded by a sub-inference.Our results suggest that the astonishing efficiency of human probabilistic reasoning may be supported by interactions between inference and memory.
Click a node to pin it, click the empty canvas to go back to this paper, or hover to preview. Open a node’s page from its title.