hide
Free keywords:
-
Abstract:
Humans seamlessly group perceptual sequences in units of chunks, parsed and memorized as separate entities. Chunking is a computational principle essential for memory compression, structural decomposition, and predictive processing. How can this ability be accomplished in a neural system? On an algorithmic level, computational models such as the Hierarchical Chunking Model (HCM) propose grouping frequently occurring proximal observational units as chunks. Chunks, once learned, are stored as separate entities in memory, ready for reuse and recombination. In doing so, the HCM learns an interpretable and hierarchical representation that resembles human chunk learning, without the need for gradient based training. In this work, we propose a biologically plausible and highly efficient implementation of the HCM with spiking neurons: the neuromorphic HCM (nHCM). When parsing through perceptual sequences, the nHCM uses sparsely connected spiking neurons to construct hierarchical chunk representations. Simulation on a standard computer showed remarkable improvement of nHCM in speed, power consumption, and memory usage compared to its original counterpart. Taking it one step further, we validate the model on mixed-signal neuromorphic hardware DYNAP-SE 2, which uses analog spiking neurons in an event-driven way to imitate biological computation. The transistors in the neural cores are run in sub-threshold, reducing energy requirements by more than one thousand times. We verified this implementation’s robust computing properties, overcoming the analog circuits’ heterogeneity, variability, and low precision. This work demonstrates cognitively plausible sequence learning in energy-efficient dedicated neural computing electronic processing systems.
