In the evolving landscape of neural network research, a significant shift is occurring from rate coding to temporal coding of signals. Temporal coding, which encodes information in the precise timing of spikes, offers advantages in processing speed and energy efficiency. This shift has brought synaptic delays into the spotlight, as they are crucial for processing signals with exact spike timings, known as spiking motifs. However, synaptic delays in the brain are bounded, often shorter than the duration of a motif, which poses a challenge for motif recognition methods that rely on heterogeneous delays to synchronize input spikes on a single output neuron.
To tackle this issue, researchers Thomas Kronland-Martinet, Stéphane Viollet, and Laurent U Perrinet have developed a novel method for detecting motifs of arbitrary length using a sequence of output neurons connected to input neurons by bounded synaptic delays. Each output neuron in this network is associated with a sub-motif of bounded duration. A motif is recognized when all sub-motifs are sequentially detected by the output neurons. This approach effectively bypasses the limitation of bounded synaptic delays, enabling the recognition of longer, more complex motifs.
The researchers simulated this network using leaky integrate-and-fire neurons and tested it on the Spiking Heidelberg Digits (SHD) database, which consists of audio data converted to spikes via a cochlear model. They also tested the network on random simultaneous motifs. The results were promising, with the network demonstrating an effective recognition of motifs of arbitrary length extracted from the SHD database. The method achieved a correct detection rate of about 60% in the presence of ten simultaneous motifs from the SHD dataset and up to 80% for five motifs, showcasing the network’s robustness to noise.
Furthermore, the results on random overlapping patterns indicated that the recognition of a single motif overlapping with other motifs is most effective for a large number of input neurons and sparser motifs. This research provides a foundation for more general models for the storage and retrieval of neural information of arbitrary temporal lengths. The implications of this work extend beyond neural network research, potentially influencing fields such as audio processing and pattern recognition, where the precise timing of signals is crucial.
