Determinantal Point Processes for Memory and Structured Inference
- Steven Frankland, Princeton Neuroscience Institute, Princeton University, Princeton, New Jersey, United States
- Jonathan Cohen, Princeton Neuroscience Institute, Princeton University, Princeton, New Jersey, United States
AbstractDeterminantal Point Processes (DPPs) are probabilistic models of repulsion, capturing negative dependencies between states. Here, we show that a DPP in representation-space predicts inferential biases toward mutual exclusivity commonly observed in word learning (mutual exclusivity bias) and reasoning (disjunctive syllogism) tasks. It does so without requiring explicit rule representations, without supervision, and without explicit knowledge transfer. The DPP attempts to maximize the total ”volume” spanned by the set of inferred code-vectors. In a representational system in which combinatorial codes are constructed by re-using components, a DPP will naturally favor the combination of previously un-used components. We suggest that this bias toward the selection of volume-maximizing combinations may exist to promote the efficient retrieval of individuals from memory. In support of this, we show the same algorithm implements efficient ”hashing”, minimizing collisions between key/value pairs without expanding the required storage space. We suggest that the mechanisms that promote efficient memory search may also underlie cognitive biases in structured inference.