Biologically Plausible Models
Exploring how psychological and neuroscientific knowledge can advance machine vision models presents a promising research direction. Our group is interested in investigating machine models that not only mirror biological systems but also provide tangible advantages for applications in the real world.
Interestingly, it is crucial to recognize that many machine vision models have been influenced by discoveries in neuroscience. For example, the foundational principles of popular machine models such as convolutional neural networks, recurrent neural networks, and reinforcement learning can be traced back to theories and findings within psychology and neuroscience. Given the brain’s evolutionary refinement over millions of years for efficient adaptation to the visual world, it can serve as a valuable reference for understanding and modeling intelligence.
The rapid advancements in deep learning techniques has sparked debate over the significance of the brain in understanding and modeling intelligence. Despite some machine models achieving remarkable performance from purely engineering-driven approaches, we believe that they still lack essential components of human vision, which could be pivotal for enhancing their efficacy and reliability.
Developing machine models that are inspired by biological principles not only has the potential to enhance machine performance but also offers unique perspectives on certain problems. This approach involves integrating biological design principles, such as recurrent processing and predictive coding-based top-down feedback mechanisms, as well as the adoption of ecological training regimes that emulate human developmental progression.
