Machine learning and statistical analysis of neuronal activity and animal behavior

Understanding the brain requires knowledge of the relationship between neuronal activity and behavior. Many novel insights into the workings of this relationship can be derived from date-driven analysis on large-scale recordings of behaving animals. Machine learning and statistical analysis are increasingly used to process the ever-increasing amount of data in science and industry, due to their ability to automate very labor-intensive tasks, or to recognize patterns in the data that are hard to discover for humans. In this thesis, we develop novel deep learning techniques as well as analysis workflows that outperform current state-of-the-art approaches and enable researchers to process large amounts of neuroscientific data automatically. For more rapid coupling of recorded neuronal activity with specific animal behaviors, we developed a supervised deep learning-based analysis pipeline for classifying behavioral motifs in groups of animals. Our approach, entitled SIPEC, outperforms multiple state-of-the-art methods. By developing novel neural network architectures we improve performance in segmentation, identification, behavioral classification and pose estimation of multiple animals, while reducing numbers of labels needed. We evaluate this approach by analyzing freely behaving mice as well as socially interacting non-human primates. Furthermore, we drive the forefront of unsupervised representation learning, trying to find generative factors that extract meaning from large-scale data through disentanglement. Disentanglement has been shown to improve generalization, interpretability and performance in downstream tasks. Our work with disentanglement thus far has overcome previous bottlenecks in robustness by introducing a training scheme, which iteratively and consistently identifies generative latent factors. Compared to previous approaches, we were able to demonstrate close to 2-fold improvement in disentanglement performance with decreased variance, almost reaching performance levels of semi-supervised approaches. With that, we pave the way to apply this unsupervised representation learning technique in practice, particularly to neuroscientific data. Lastly, we tackle major obstacles in statistical analysis and data management in rodent functional magnetic resonance imaging (fMRI). We developed a principled approach for preprocessing mouse MRI data and registration to a standard geometric space to overcome problems of reproducibility with currently used workflows. Moreover, we introduce 4 metrics for evaluating this process that are also usable for evaluating human fMRI processing. With that, we set a new standard for the processing of preclinical fMRI data. Overall, we introduce multiple lines of work for the supervised, unsupervised and statistical analysis of behavioral and neuronal data. We release our work as open-source frameworks, enabling enhanced analysis over a broad range of data in basic neuroscience research. Finally, We give an outlook as to how these technologies might be combined and applied to spur neuroscientific research in the future