Inferring Functional Neural Connectivity with Deep Residual Convolutional Networks

Abstract

Measuring synaptic connectivity in large neuronal populations remains a major goal of modern neuroscience. While this connectivity is traditionally revealed by anatomical methods such as electron microscopy, an efficient alternative is to computationally infer functional connectivity from recordings of neural activity. However, these statistical techniques still require further refinement before they can be reliably applied to real data. Here, we report significant improvements to a deep learning method for functional connectomics, as assayed on synthetic ChaLearn Connectomics data. The method, which integrates recent advances in convolutional neural network architecture and model-free partial correlation coefficients, outperforms published methods on competition data and can achieve over 90% precision at 1% recall on validation datasets. This suggests that future application of the model to in vivo whole-brain imaging data in larval zebrafish could reliably recover on the order of 10 6 synaptic connections with a 10% false discovery rate. The model also generalizes to networks with different underlying connection probabilities and should scale well when parallelized across multiple GPUs. The method offers real potential as a statistical complement to existing experiments and circuit hypotheses in neuroscience.

A Review of Network Inference Techniques for Neural Activation Time Series

“…Our experiments showed that unsupervised learning provides a more efficient solution. However, we know that RCNN has the strongest solution with networks of 1000 cells [6] and as we showed, it has a very competitive performance even with networks of 100 cells. Thus, the use of supervised methods can be justified, if certain requirements regarding the computational resources and the volume of data are satisfied.…”

“…However, there is not extensive literature based on neural networks for network inference and analysis, with small exceptions, such as a convolutional neural network that identifies contrastive underlying weighted network structures in fMRI data of different subject groups [17]. In this work, we are going to examine a residual convolutional neural network, which is based on the method that achieved fourth place in the competition [28] and has been extended to a state of the art solution [6]. This method is accompanied with its own preprocessing, as the model works directly with the fluorescence signals.…”

“…To ensure a balance between positive and negative samples in the training dataset, subsets of activity from pairs of connected neurons are sampled more times than the ones with no connection. The architecture of the residual neural network employed [6] can be seen in figure 1 The training of the model was performed with Adam optimization, on 100 epochs, a batch size of 100, early stopping in 20 epochs, a dropout probability of 0.2 for convolutional layers and 0.5 for dense layers. We base our implementation on the python libraries TFconnect 3 and TFomics 4 .…”