Supervised Neural Network Structure Recovery

Ildefons Magrans de Abril, Ann Nowé

Abstract

Inferring the underlying topology of neural networks from time-series fluorescence signals is a fundamental challenge in connectomics. Traditional methods often rely on single, unsupervised indicators that struggle with the inherent noise and synchronous bursting characteristic of biological data. This paper presents a supervised ensemble approach that treats network recovery as a regression problem. By engineering a high-dimensional feature set from hundreds of connectivity indicators—each tuned to specific noise and synchronization thresholds—and training an ensemble of Random Forests and Gradient Boosting Machines, we achieved a significant performance improvement. Our method secured a podium finish in the ECML 2014 Neural Connectomics Challenge, reaching an Area Under the ROC Curve (AUC) of approximately 0.94, compared to the 0.88 baseline of Generalized Transfer Entropy.

Introduction

Understanding the "wiring diagram" of the brain is essential for deciphering neural function. While calcium imaging provides high-throughput observations of thousands of neurons, the physical synaptic connections remain hidden. Researchers must infer these connections from fluorescence signals, a task complicated by low temporal resolution (often 20ms frames vs. sub-millisecond firing), background noise, and synchronous bursting events that mask individual connectivity patterns.

Historically, state-of-the-art methods utilized unsupervised metrics like Generalized Transfer Entropy (GTE). However, these "one-size-fits-all" indicators are often too rigid for the messy reality of biological data. We propose a paradigm shift: instead of seeking a single optimal metric, we leverage a supervised machine learning framework to synthesize information from a diverse array of connectivity features.

Methodology

The core of our approach is a multi-stage pipeline designed to transform raw time-series data into robust connectivity predictions.

The multi-stage pipeline for supervised neural network structure recovery, showing neuron-wise and network-wise processing steps.

Feature Engineering Pipeline

  1. Spike Inference: Raw fluorescence signals are processed through a sequential Monte Carlo framework to estimate spike probabilities.
  2. Noise and Burst Removal: We apply varying thresholds for Noise Level (NL) and Synchronization Rate (SR). The SR filter is particularly critical; it identifies and discards time steps where too many neurons fire at once, focusing the analysis on "quiet" periods where true point-to-point connectivity information is most visible.
  3. Connectivity Scoring: We calculate simple correlations between spike trains across a massive grid of 252 different parameter combinations. The computational efficiency of correlation allows for this high-dimensional feature expansion.
  4. Network Deconvolution: This step distinguishes direct connections from indirect paths (e.g., A correlating with C only because both are connected to B).
  5. Z-Normalization: To ensure the model generalizes across different networks, we apply Z-normalization to the deconvolved matrices, standardizing the feature distributions.

Supervised Ensemble

The resulting 252 connectivity indicators serve as input features for a supervised model. We trained an ensemble of Random Forests (RF) and Gradient Boosting Machines (GBM) to predict the likelihood of a connection between any given pair of neurons.

Results

The proposed method was evaluated using datasets from the ChaLearn 2014 Neural Connectomics Discovery Challenge, involving networks of 1,000 neurons.

Performance comparison (AUC) across different training networks and model configurations, demonstrating the superiority of the ensemble approach over Generalized Transfer Entropy (GTE).

As shown in the results table, our supervised approach consistently outperformed the GTE baseline (0.88) across all test networks, achieving AUC scores above 0.94. The ensemble models (ALL) provided the most stable and highest performance, demonstrating that combining multiple weak learners effectively mitigates the noise inherent in individual network topologies. Our final submission achieved the 3rd best score out of 144 competing teams.

Discussion

The success of this approach highlights the "Power of Diversity" in feature engineering. By generating a high-dimensional view of the network through various filters, the machine learning model can navigate the data's complexity more effectively than any single human-tuned metric. While correlation is simpler than GTE and lacks directionality, its stability and speed allowed for the creation of an ensemble that ultimately yielded superior predictive power.

However, the method's primary limitation is its computational cost, requiring significant time for correlation and deconvolution steps on large networks. Future work will focus on optimizing these steps for scalability and incorporating asymmetric indicators to better capture the directionality of neural connections.

Conclusion

By reframing neural structure recovery as a supervised learning task, we have demonstrated a robust path forward for connectomics. The combination of domain-specific filtering and powerful ensemble learning provides a significant leap in accuracy over traditional unsupervised methods.

Summary updated

Abstract

This paper presents our solution to the European Conference of Machine Learning Neural Connectomics Discovery Challenge. The challenge goal was to improve the performance of existing methods for recovering the neural network structure given the time series of neural activities. We propose to approximate a function able to combine several connectivity indicators between neuron pairs where each indicator is the result of running a feature engineering pipeline optimized for a particular noise level and firing synchronization rate among neurons. We proved the suitability of our solution by improving the state of the art prediction performance more than 6% and by obtaining the third best score on the test dataset out of 144 teams.
BibTeX
@inproceedings{abril2014supervised,
  author = {Ildefons Magrans de Abril and Ann Nowé},
  title = {{Supervised Neural Network Structure Recovery}},
  booktitle = {Neural Connectomics},
  year = {2014},
}
Figure