Cortex-inspired goal-directed recurrent networks for developmental visual attention and recognition with complex backgrounds
Juyang Weng, Matthew Luciw · 2010
It is unknown how the brain self-organizes its internal wiring without a holistically-aware central controller. How does the brain develop internal object representations for a massive number of objects? How do such representations enable tightly intertwined attention and recognition in the presence of complex backgrounds? Most vision systems have not included top-down connectivity or treated bottom-up and top-down separately. Yet almost all excitatory pathways in the visual cortex are bidirectional; evidence suggests the top-down connections are not fundamentally different front bottom-up connections. This dissertation presents and analyzes a hierarchical self-organizing type of network with adaptive excitatory bottom-up and top-down connections. This two-way network takes advantage of grounding—both the sensory end (visual patches) and motor end (action control) are input ports. Internally, local neural learning uses only the co-firing between the pre-synaptic and post-synaptic activities. Such a representation automatically boosts action-relevant components in the sensory inputs (e.g., foreground vs. background) by increasing the chance of only action-related feature detectors to win in competition. After learning, the bidirectional networks showed topographic semantic grouping and modular connectivity. It is shown how and why such modular networks can take advantage of recurrent excitation for recognition. In Where-What Network-3, top-down connections enabled type-based and location-based top-down attention and synchronization of neurons over multiple levels to bind features into holistic representations.