The Two-Stream Hypothesis as a Foundation for Human-Like Memory and Action
摘要
In the two-stream hypothesis, the mammalian brain’s visual (and possibly other sensory) information diverges into a ventral “what” pathway for object identity and a dorsal “where” pathway for spatial planning. Although this anatomical split is well documented, its computational value and its role in the emergence of human-like intelligence remain debated. We combine a neurobiological review with targeted simulations to evaluate when a split model (i.e., “what” and “where” facts are stored separately) outperforms a unified model (i.e., facts stored together). Simulations (n = 5,000 objects) show that a split model answers integrated queries (object identity plus location) ≈ 17% slower than a unified model. While absolute timings vary by hardware, the qualitative trends remain robust. Parameter sweeps reveal a crossover: the split model becomes more efficient when integrated queries fall below ≈15% of the total (> 85% ask only “what” or “where”), when object memory exceeds cache efficiency, or when there is more parallelism. We argue that this operating regime matches real-world conditions in which rapid visuomotor loops and abstract conceptual learning proceed simultaneously, and that the forced “binding step” between streams may foster compositional thought. The results suggest that representational segregation is not an evolutionary relic but a computationally advantageous principle worth implementing in brain-inspired cognitive architectures (BICAs).