Publikation
A Controller for Adaptive Neuromorphic Sensing using Active Efficient Coding
Aheli Saha; Didier Stricker; René Schuster
In: Proceedings of the International Workshop on Reliable and Sustainable Neuromorphic Hardware. International Workshop on Reliable and Sustainable Neuromorphic Hardware, May 28-30, York, United Kingdom, 2025.
Zusammenfassung
Analogous to computational systems, biological systems, too, are faced with the constraint of limited resources. It has been studied and observed that to meet these energy constraints, the brain has evolved to not only optimize neural representations but also use motor actions to vary and adapt the statistics of the input signals. The Active Efficient Coding (AEC) theory postulates that the organism's behaviour that shapes these input sensory signals is tuned in conjunction with the neural representations to arrive at an optimal performance. We argue that this concept can be extended to artificial vision systems, specifically event cameras, in our case, to further improve their energy utilization. The conditions for triggering an event are generally pre-determined through calibration procedures and applied once using programmable bias circuits. This indicates that sensor transduction is constant, and there is no dynamic adjustment for varying environmental stimuli or the current task requirements. To overcome this limitation while concurrently promoting enhanced efficiency, we propose jointly optimising the input statistics and performance of a task-specific neural network.