Spiking time-dependent plasticity leads to efficient coding of predictions

Pau Vilimelis Aceituno, Masud Ehsani, Jürgen JostBiol. Cybern 2020 · DOI

TL;DR

This paper explores how Spiking Time-Dependent Plasticity (STDP), a biologically inspired learning rule, can enable neurons to efficiently code and even predict incoming spike trains. By reducing the latency of postsynaptic spikes and concentrating them, STDP improves signal-to-noise ratio, lowers metabolic costs, and lays the groundwork for predictive coding without explicit feedback.

What Problem Does This Paper Address?

Living organisms need to make accurate predictions to survive. While traditional models often rely on supervised or reinforcement learning with explicit feedback, recent findings suggest that sensory neurons can encode predictive information bottom-up, without higher-level feedback. This paper investigates how spiking neural networks, specifically through the mechanism of Spiking Time-Dependent Plasticity (STDP), can achieve this predictive capability. The authors aim to bridge the gap between the known effect of STDP on latency reduction and its potential role in efficient neural coding and prediction generation.

Methods

The authors employ a combination of mathematical analysis and simulations using the Leaky Integrate-and-Fire (LIF) neuron model. They focus on the effects of STDP on postsynaptic spike timing and number. Key aspects of their methodology include:

  • LIF Neuron Model: A standard model for simulating neuronal activity, incorporating a refractory period.
  • STDP Rule: A biologically plausible learning rule where synaptic weights are modified based on the precise timing difference between pre- and postsynaptic spikes. The paper uses a standard STDP kernel with separate parameters for excitatory and inhibitory synapses.
  • Analysis of Latency Reduction: Investigating how repeated exposure to the same spike train leads to earlier postsynaptic spikes.
  • Analysis of Spike Disappearance: Examining how STDP can cause later spikes in a train to disappear if they occur too close to earlier ones.
  • Introduction of Noise: Incorporating Gaussian noise to simulate biological variability and its effect on spike timing and disappearance.
  • Population-Level Analysis: Extending the findings from single neurons to populations to study emergent behaviors like prediction.

Datasets

Not discussed in the paper. The authors use simulated input spike trains rather than real-world datasets.

Key Insights and Ideas

  • STDP as a Prediction Engine: The core idea is that by learning to fire earlier in response to a frequent input pattern, a neuron effectively "predicts" the arrival of that pattern. This prediction is a consequence of the STDP rule reinforcing early-arriving presynaptic spikes.
  • Efficient Coding through Latency Reduction: When neurons fire earlier, their spikes are concentrated in a smaller time window. This makes them easier to decode and reduces the metabolic cost associated with generating many spikes for frequent stimuli.
  • Spike Concentration and Synchronization: STDP leads to a concentration of spikes at the onset of a stimulus, which can increase synchrony within neural populations. This synchronization is crucial for effective communication between neurons.
  • Inhibition's Role: The paper explores how inhibitory plasticity can be integrated with excitatory STDP. They find that under specific parameter regimes, inhibitory plasticity can coexist with and even enhance latency reduction and spike disappearance effects.
  • Bottom-Up Prediction: The findings suggest that predictive capabilities can emerge from local synaptic plasticity rules without the need for explicit top-down feedback mechanisms.

Key Results

  • Latency Reduction: Repeated presentation of a fixed input spike train leads to a significant decrease in the latency of the postsynaptic spike.
  • Spike Disappearance: When postsynaptic spikes occur close in time, STDP can lead to the disappearance of later spikes, especially in the presence of noise.
  • Improved Signal-to-Noise Ratio (SNR): STDP can optimize the SNR of neuronal responses, making the encoded information more robust.
  • Reduced Metabolic Costs: By concentrating spikes and potentially reducing their total number for frequent stimuli, STDP contributes to metabolic efficiency.
  • Emergence of Predictions: The paper demonstrates through simulations that neurons can learn to fire before the actual stimulus onset if the stimulus is preceded by a consistent pattern, effectively exhibiting predictive behavior.

Limitations and Future Work

  • Model Simplifications: The authors acknowledge that their LIF neuron model and STDP rule are simplifications of biological reality. They do not capture the full complexity of ion channel kinetics or all forms of synaptic plasticity.
  • Parameter Dependence: The observed effects are dependent on the specific parameters of the STDP rule and neuron model. Finding the exact parameter regimes for optimal prediction and coding efficiency requires further investigation.
  • Experimental Verification: The paper calls for experimental studies to identify neurons that encode temporally correlated stimuli and to observe latency evolution due to STDP, ensuring it's not confounded by reward systems.
  • Recurrent Connections: While the paper briefly touches upon recurrent connections, a more in-depth analysis is needed to understand their full impact on the emergent behaviors.
  • Generalization to Complex Stimuli: The current study focuses on fixed or random spike trains. Extending the findings to more complex, naturalistic stimuli and sequences is a crucial next step.

Summary updated

BibTeX
@article{aceituno2020spiking,
  author = {Pau Vilimelis Aceituno and Masud Ehsani and Jürgen Jost},
  title = {{Spiking time-dependent plasticity leads to efficient coding of predictions}},
  journal = {Biol. Cybern},
  year = {2020},
  doi = {10.1007/s00422-019-00813-w},
}
Figure