Optimal localist and distributed coding of spatiotemporal spike patterns through STDP and coincidence detection

Timothée Masquelier, Saeed Reza Kheradpisheh

Abstract

Repeating spatiotemporal spike patterns exist and carry information. Here we investigated how a single spiking neuron can optimally respond to one given pattern (localist coding), or to either one of several patterns (distributed coding, i.e. the neuron's response is ambiguous but the identity of the pattern could be inferred from the response of multiple neurons), but not to random inputs. To do so, we extended a theory developed in a previous paper [Masquelier, 2017], which was limited to localist coding. More specifically, we computed analytically the signal-to-noise ratio (SNR) of a multi-pattern-detector neuron, using a threshold-free leaky integrate-and-fire (LIF) neuron model with non-plastic unitary synapses and homogeneous Poisson inputs. Surprisingly, when increasing the number of patterns, the SNR decreases slowly, and remains acceptable for several tens of independent patterns. In addition, we investigated whether spike-timing-dependent plasticity (STDP) could enable a neuron to reach the theoretical optimal SNR. To this aim, we simulated a LIF equipped with STDP, and repeatedly exposed it to multiple input spike patterns, embedded in equally dense Poisson spike trains. The LIF progressively became selective to every repeating pattern with no supervision, and stopped discharging during the Poisson spike trains. Furthermore, using certain STDP parameters, the resulting pattern detectors were optimal. Tens of independent patterns could be learned by a single neuron using a low adaptive threshold, in contrast with previous studies, in which higher thresholds led to localist coding only. Taken together these results suggest that coincidence detection and STDP are powerful mechanisms, fully compatible with distributed coding. Yet we acknowledge that our theory is limited to single neurons, and thus also applies to feed-forward networks, but not to recurrent ones.

Introduction

In neuroscience, a long-standing debate exists between localist coding (the "Grandmother Cell" hypothesis, where one neuron responds to exactly one concept) and distributed coding (where neurons respond to multiple stimuli, and identity is inferred from the population). While previous research showed that a single neuron could learn to detect a single repeating spike pattern hidden in noise, it remained unclear if a single neuron could robustly learn multiple independent patterns. If a neuron can learn many, it becomes a building block for distributed coding. This study addresses whether a neuron can achieve this without a "teacher" (unsupervised) and still maintain a high signal-to-noise ratio (SNR) to avoid false alarms.

Methods and Datasets

The authors used a two-pronged approach. First, they developed an analytical theory extending a previous mathematical framework for a threshold-free Leaky Integrate-and-Fire (LIF) neuron. They derived a formula for the SNR of a "multi-pattern detector" to determine the theoretical limit of how many patterns a neuron can distinguish from background Poisson noise.

Second, they conducted numerical simulations of a LIF neuron equipped with Spike-Timing-Dependent Plasticity (STDP). This included Long-Term Potentiation (LTP) when a presynaptic spike preceded a postsynaptic one, and homeostatic Long-Term Depression (LTD) to prevent weight saturation. Crucially, they implemented an adaptive threshold that increases after a spike and decays over time, preventing the neuron from firing multiple times for the same pattern and encouraging it to respond to other patterns.

Spatiotemporal spike patterns (top) embedded in noise and the resulting membrane potential (bottom). Colored boxes indicate repeating patterns.

The study used synthetic datasets consisting of background Poisson spike trains (e.g., 3.2Hz) and "frozen noise" patterns (100ms duration) embedded repeatedly. To simulate biological reality, the timing of spikes within the repeating patterns was "jittered" at each presentation.

Key Results

The results demonstrate that the neuron acts as a highly efficient coincidence detector. By picking up on spikes that consistently arrive together within a specific time window, it can distinguish patterns from random noise.

The SNR remains high across a range of parameters (left) and follows a predictable distribution (right), indicating near-optimal learning.

Key findings include:

  • Surprisingly High Capacity: The analytical SNR decreases very slowly as patterns are added. A single neuron could robustly learn up to 40 independent patterns. For up to 20 patterns, the success rate for "optimal" learning was 100%.
  • The Threshold is the Key: Previous studies used high thresholds, forcing localist coding. By using a low adaptive threshold, the neuron is allowed to fire for multiple different patterns, enabling distributed coding.
  • STDP as an Optimizer: Simulations showed that unsupervised STDP naturally pushes the neuron toward the synaptic weight distribution that maximizes the theoretical SNR.

Discussion and Limitations

This work suggests that coincidence detection and STDP are powerful mechanisms fully compatible with distributed coding. However, the theory is currently limited to single neurons and feed-forward architectures. Future work is needed to explore these dynamics in recurrent networks where neurons excite and inhibit each other. Additionally, while the paper shows a neuron can respond to multiple patterns, it does not model how a downstream "reader" would disambiguate those responses, which would require a population-level analysis.

Summary updated

Abstract

Repeating spatiotemporal spike patterns exist and carry information. Here we investigated how a single spiking neuron can optimally respond to one given pattern (localist coding), or to either one of several patterns (distributed coding, i.e. the neuron's response is ambiguous but the identity of the pattern could be inferred from the response of multiple neurons), but not to random inputs. To do so, we extended a theory developed in a previous paper [Masquelier, 2017], which was limited to localist coding. More specifically, we computed analytically the signal-to-noise ratio (SNR) of a multi-pattern-detector neuron, using a threshold-free leaky integrate-and-fire (LIF) neuron model with non-plastic unitary synapses and homogeneous Poisson inputs. Surprisingly, when increasing the number of patterns, the SNR decreases slowly, and remains acceptable for several tens of independent patterns. In addition, we investigated whether spike-timing-dependent plasticity (STDP) could enable a neuron to reach the theoretical optimal SNR. To this aim, we simulated a LIF equipped with STDP, and repeatedly exposed it to multiple input spike patterns, embedded in equally dense Poisson spike trains. The LIF progressively became selective to every repeating pattern with no supervision, and stopped discharging during the Poisson spike trains. Furthermore, using certain STDP parameters, the resulting pattern detectors were optimal. Tens of independent patterns could be learned by a single neuron using a low adaptive threshold, in contrast with previous studies, in which higher thresholds led to localist coding only. Taken together these results suggest that coincidence detection and STDP are powerful mechanisms, fully compatible with distributed coding. Yet we acknowledge that our theory is limited to single neurons, and thus also applies to feed-forward networks, but not to recurrent ones.
BibTeX
@inproceedings{masquelier2018optimal,
  author = {Timothée Masquelier and Saeed Reza Kheradpisheh},
  title = {{Optimal localist and distributed coding of spatiotemporal spike patterns through STDP and coincidence detection}},
  booktitle = {Frontiers Comput. Neurosci},
  year = {2018},
  url = {https://openreview.net/forum?id=DAfUDqCX1Nc},
}
Figure