Neural Population Coding for Effective Temporal Classification

Zihan Pan, Jibin Wu, Malu Zhang, Haizhou Li, Yansong Chua

TL;DR

This paper demonstrates that the way we encode real-valued data into spikes is just as critical as the learning algorithms used in Spiking Neural Networks (SNNs). By using population coding—distributing a single signal across multiple neurons—temporal information is projected into a high-dimensional spatial domain, making complex auditory patterns linearly separable and significantly easier for simple classifiers to solve.

What Problem Does This Paper Address?

In the world of Spiking Neural Networks (SNNs), most research focuses on the "back-end": the learning rules and architectures used to classify data. However, the "front-end"—how we convert continuous signals (like speech or sound) into discrete spikes—is often treated as an afterthought.

The authors argue that this encoding step is a powerful abstraction that can either simplify or complicate the learning task. Specifically, they tackle the challenge of temporal classification. Real-world signals like speech are "temporally rich," meaning the information is buried in how the signal changes over time. If the encoding scheme is poor, the SNN has to work incredibly hard to find patterns. If the encoding is effective, the classification task might become trivial.

Methods

The researchers conducted a comparative study of five biologically inspired neural encoding schemes:

  1. Latency Code: Information is represented by the timing of the first spike (earlier spikes = higher intensity).
  2. Phase Code: Spikes are timed relative to a background sub-threshold oscillation (SMO).
  3. Population-Latency/Phase Code: These extend the above by using a "population" of neurons with overlapping Gaussian receptive fields to represent a single input value.
  4. Threshold Code: This tracks the trajectory of a signal, firing "onset" neurons when the signal crosses a threshold from below and "offset" neurons when it crosses from above.

To evaluate these, they used two classification back-ends:

  • Support Vector Machine (SVM): Used as a "probe" to test the linear separability of the encoded patterns.
  • Tempotron: A biologically plausible SNN learning rule that learns to fire for specific spike patterns.

They also compared two vectorization techniques to feed spikes into the SVM: one that collapsed temporal information into average firing rates (Spatial only) and one that preserved precise spike timings (Spatio-temporal).

Datasets

The study utilized two standard auditory benchmarks:

  • TIDIGITS: A dataset of isolated spoken digits ('0'-'9' and 'oh') spoken by various people with different accents.
  • RWCP (Real World Computing Partnership): An environmental sound dataset containing samples like bells, phone rings, and other non-speech sounds.

Key Insights and Ideas

  • Time-to-Space Projection: The most significant "aha moment" is that population coding effectively projects temporal information into the spatial domain. By spreading one signal across many neurons, the network creates a high-dimensional spatial map that captures what was previously a purely temporal sequence.
  • Linear Separability is the Goal: The authors found that with population coding, the encoded patterns become linearly separable. This means the "features" are so well-defined by the encoding that a simple linear classifier (like an SVM) can achieve near-perfect accuracy without any complex deep learning.
  • The "Neglected" Front-End: The paper proves that an effective neural encoding scheme can simplify a classification problem to the point where a very simple SNN (like the Tempotron) can succeed where it would otherwise fail using basic latency codes.
  • Efficiency vs. Fidelity: Threshold coding emerged as a highly efficient "sparse" representation. It uses significantly fewer spikes (lower energy) while maintaining high classification accuracy, though it might be less suited for reconstructing the original signal.

Key Results

  • SVM Performance: Using population-based encoding, the SVM achieved 95% accuracy on TIDIGITS and 100% accuracy on RWCP.
  • The Power of Population: Single-neuron latency coding performed poorly when temporal information was collapsed (only ~20-48% accuracy). However, population-latency coding jumped to 92.7% (TIDIGITS) and 100% (RWCP) under the same conditions.
  • SNN Agreement: The Tempotron results mirrored the SVM findings. On the RWCP dataset, population-latency coding reached 99% accuracy, compared to a mere 10% (chance level) for simple latency coding.
  • Spike Sparsity: Threshold coding was the most "biologically efficient," using an average of only 12 spikes per individual neuron compared to 950 for basic latency coding, while still maintaining ~95% accuracy on TIDIGITS.

Limitations and Future Work

The primary limitation acknowledged is that the study focused on isolated patterns (single digits or distinct sounds). In the real world, speech and environmental sounds are continuous. The authors note that extending these findings to continuous signals—where the boundaries between "events" are not clearly defined—remains a significant challenge.

Future work will likely look at how these population codes can be integrated into deeper SNN architectures and how they handle the "segmentation" problem in continuous streams of data.

Summary updated

Abstract

Neural encoding plays an important role in faithfully describing the temporally rich patterns, whose instances include human speech and environmental sounds. To classify such spatio-temporal patterns with the Spiking Neural Networks (SNNs), how these patterns are encoded has a direct impact on the complexity of the task. In this paper, we study several existing temporal and population coding schemes in speech and audio recognition. We show that, with population neural coding, the encoded patterns are linearly separable using the Support Vector Machine (SVM). We note that the population neural coding effectively project the temporal information into the spatial domain, thus improving linear separability of the patterns. We achieve an accuracy of 95% and 100% on TIDIGITS and RWCP datasets respectively with SVM classifier. We further implement the Tempotron as an SNN-based classifier on the same datasets and achieve similar results. The study suggests that an effective neural coding scheme is just as important as the classifier.
BibTeX
@inproceedings{pan2019neural,
  author = {Zihan Pan and Jibin Wu and Malu Zhang and Haizhou Li and Yansong Chua},
  title = {{Neural Population Coding for Effective Temporal Classification}},
  booktitle = {IJCNN},
  year = {2019},
  doi = {10.1109/ijcnn.2019.8851858},
}

Loading related content...

Figure