Overview
Biological sensory neurons exhibit selective receptive fields structured along meaningful stimulus coordinates, such as frequency, motion direction, or retinotopic position. Previous computational investigations involving simpler neurons across various modalities have indicated that such organization may stem from efficient coding principles and biological constraints related to neuronal activity, connectivity, and wiring. This study aimed to determine if structured receptive fields provide a computational advantage beyond mere resource efficiency, and whether this advantage persists even when individual neurons possess high expressivity.
Research Context
The existence of structured receptive fields in biological sensory neurons is a recognized phenomenon, with examples including organization based on frequency, motion direction, or retinotopic position. Prior computational research has explored the genesis of this structure, linking it to efficient coding and biological limitations on aspects like neuronal activity, connectivity patterns, and wiring. The current research specifically addresses the question of whether these structured receptive fields offer a computational benefit that extends beyond resource efficiency. A key aspect of this inquiry is to ascertain if such an advantage remains relevant even when individual neurons demonstrate a high degree of expressivity or computational complexity.
Approach
The investigation utilized recurrent networks composed of Expressive Leaky Memory neurons. This architectural choice allowed for independent manipulation of two key variables: the computational complexity of individual neurons and the organization of feed-forward receptive fields. The research evaluated performance across two distinct classification tasks: an auditory classification task and an event-based visual classification task. The methodology involved comparing the test accuracy achieved with receptive fields aligned with task-relevant sensory coordinates against that obtained with budget-matched random receptive fields. Further, the study examined the impact of scrambling sensory coordinates and aligning receptive fields with task-irrelevant coordinates. The role of increasing neuronal complexity was also assessed. Finally, the researchers explored the effects of generic synaptic sparsity regularization on input selectivity and performance recovery.
Findings
- Receptive fields aligned with task-relevant sensory coordinates consistently improved test accuracy in both auditory and event-based visual classification tasks. This improvement was observed when compared to budget-matched random receptive fields.
- The computational advantage conferred by structured receptive fields disappeared under two specific conditions: when sensory coordinates were scrambled, or when receptive fields were designed to follow task-irrelevant coordinates. This indicates that the benefit arises from alignment with the task's inherent geometry, rather than solely from restricted connectivity.
- An increase in neuronal complexity, referring to the expressivity of individual neurons, led to a reduction in the performance advantage provided by structured receptive fields.
- Generic synaptic sparsity regularization was found to induce input selectivity and partially recover performance. However, the performance achieved through sparsity regularization alone remained substantially below that of explicitly structured receptive fields. This suggests that sparsity by itself is insufficient to fully recover the computational benefits associated with task-aligned receptive fields.
Why This Matters
The study's findings suggest that appropriate receptive fields can function as a computational prior, offering benefits beyond what is achieved through sparsity alone. The research also highlights that the value or impact of these structured receptive fields is contingent upon the computational expressivity of individual neurons within a system.