Continual familiarity decoding from recurrent connections in spiking networks.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Memory Systems

Background:

• Familiarity memory allows recognizing familiar inputs without detailed recall, crucial for adaptive behavior.
• Existing models often require extensive training or struggle with temporal generalization.

Purpose of the Study:

• To present a novel spiking neural network (SNN) model for encoding and decoding familiarity.
• To investigate the role of unsupervised spike-timing-dependent plasticity (STDP) in familiarity encoding.
• To compare the SNN model's performance against LSTM on continual familiarity detection tasks.

Main Methods:

• Developed an SNN with lateral connectivity shaped by unsupervised STDP.
• Encoded familiarity through local plasticity events within the network.
• Decoded familiarity using frequency (spike count) and temporal (spike synchrony) characteristics of spike trains.
• Evaluated performance on a continual familiarity detection task, comparing with LSTM.

Main Results:

• The SNN model successfully encodes familiarity via local plasticity.
• Familiarity can be decoded using both spike count and spike synchrony.
• Temporal coding shows superior performance under sparse input conditions, aligning with brain principles.
• The SNN model demonstrates better temporal generalization than LSTM on the continual familiarity detection task.
• Input stimuli are naturally encoded in recurrent connectivity, eliminating the need for a separate training stage.

Conclusions:

• Unsupervised STDP in SNNs provides an effective mechanism for familiarity memory.
• The model's ability to use both frequency and temporal coding enhances its robustness.
• The SNN approach offers a promising, biologically plausible alternative to traditional deep learning models for temporal tasks.