Babies possess the ability to predict what will happen in their environment based on intuition about gravity, motion, and mass of objects. This has lead to the idea that the human brain contains a “physics engine”, which is a structure built of nerve cells that can simulate events in the physical world. However, the exact neural architecture that mediates these abilities, and how it operates mechanistically, has remained a mystery. A major aspect of general intelligence could be revealed if animals with experimentally tractable nervous systems possessed such innate physical abilities.
In a recent paper published in eLife, Andrew Bolton and colleagues in the lab of MCB professor Florian Engert examined whether the larval zebrafish, a primitive organism with 1 million times fewer neurons than a human, could show similar prediction abilities. For their study, the group chose to closely examine the zebrafish’s prey capture behavior, during which the fish pursues and consumes fast swimming single-celled microbes. A high-resolution 3D setup and custom computer vision algorithms were used to reconstruct prey capture sequences from the perspective of the fish. For the first time, this allowed the examination of how zebrafish choose their movements according to prey position and velocity. Interestingly, the team discovered that fish make accurate predictions about where prey will be in the future. Moreover, computational modeling of the fish’s energy consumption and speed during prey capture showed that this physical prediction ability is critical to the successful and efficient hunting algorithm that the fish has evolved.
Perhaps the most interesting feature of the zebrafish’s hunting algorithm, however, is that it is somewhat random. When fish are further away from their prey, both in angle and distance, they implement more variable movements; these movements are more precise as fish become closer to their targets. To quantitatively understand this phenomenon, the Engert lab collaborated with Josh Tenenbaum and Vikash Mansingha’s groups at MIT to implement a “probabilistic programing” approach. Using these tools allowed the team to accurately mimic the zebrafish’s stochastic hunting strategy, which remarkably uncovered that the fish’s random behavior is actually beneficial to its hunting efficiency. This surprising result conflicts with our typical notions that randomness in behavior is a nuisance to be overcome.
Bolton and Engert hope to pursue the neural basis of the fish’s stochastic algorithm and show how the fish’s predictive abilities arise in the brain. They hope that the discovery and illustration of predictive abilities in zebrafish will shed light on the field’s long-held interest in how humans have evolved their complex understanding of the physical world.