BrainTune

AUTOMATED LEARNING OF DEEP REPRESENTATIONS AND HYPERPARAMETER TUNING FOR BRAINLINKS-BRAINTOOLS DATA


Relevant for Research Area

A - Foundations

B - Core Technologies

C - Applications

PIs

Prof. Frank Hutter

Prof. Tonio Ball

Prof. Wolfram Burgard

Dr. Michael Tangermann

 

former PIs

Prof. Martin Riedmiller


Summary

The constantly growing amount and complexity of data involved in brain studies makes manual parameterization of data processing and modelling algorithms a tedious and time-consuming task. This can preclude several BrainLinks-BrainTools projects from reaching their full potential due to the use of poorly parameterized methods.

This project aims to provide tools for automated hyperparameter tuning to enable continuous exploration of capabilities and improvements within BrainLinks- BrainTools projects. The special focus is on automated deep representation learning aimed at extracting discriminative patterns from brain signals.

Many projects within BrainLinks-BrainTools have to extract patterns from high-dimensional brain signals. Whether the problem is to detect seizures in epilepsy patients or to decode neural activation patterns in BCI, these brain signals are not yet understood well enough to effectively be used as inputs to standard machine learning algorithms. The young field of automated learning of deep representations aims to learn discriminative features directly from such raw data.

The biggest problem with deep learning of representations is that performance is extremely sensitive to appropriate choices of the deep network's structure and to the setting of its hyperparameters: depending on its configuration, a network may perform extremely poorly or define the state of the art for a problem. The same is true for other algorithmic problem solutions to neuroscience problems, whose performance – and thus impact could be substantially improved by using better parameterizations of the method.

This project addresses these challenges which cut across BrainLinks-BrainTools projects whereas it comprises foundational research on automating deep learning and hyperparameter optimization, as well as its applications in the context of BrainLinks-BrainTools.

Hyperparameter Tuning and Deep Learning

The success of machine learning heavily relies on finding the right algorithm and its hyperparameters for the data at hand. Since manual experimentation in this joint space of algorithms and hyperparameter settings is tedious and time-consuming, the machine learning community has recently started to develop automated machine learning (AutoML) systems to remove the need for an expert in the loop.

Automated approaches have been shown to outperform manual tuning by human experts in several domains. Nevertheless, in general they are not yet ready to replace human experts out of the box since humans use intuitive methods to reason across different data sets, determine which parameters actually matter, terminate runs with poor performance patterns early, and parallelize evaluations across a cluster of machines. These fundamental tools are not traditionally available in automated machine learning systems. This project aims to develop such methods and integrate them in automated machine learning systems.

These automated tools are particularly important in the context of deep learning, which is quickly becoming a standard tool to address large data sets and is already being applied in several BrainLinks-BrainTools projects. Deep learning is not only very sensitive to its hyperparameter settings but also to the selection of the right network architecture. Selecting these components appropriately currently requires both expert knowledge in deep learning and a solid understanding of the data being modelled. Since there are only few experts with knowledge in both brain data and deep learning (and the state of the art is quickly advancing), automating parts of the manual experimentation process is very promising to achieve robust and improved performance.

Applications in BrainLinks-BrainTools

We apply our methods in collaboration with three BrainLinks-BrainTools partner projects: Neurobots, FAMOX, and MakeITReaL.

A subproject within Neurobots compares different deep learning pipelines to standard approaches to decode motor imagery brain signals. As there is no common knowledge about the right neural network architecture we guide the design process with our tools.

The project FAMOX applies the SPoC algorithm to study EEG data in order to support rehabilitation of stroke patients. It is expected that tuning of various components of SPoC (frequency bands, thresholds, etc.) will yield improved performance.

In project MakeITReaL, one important problem is to track the motion of mammals (in particular, sheep and humans) based on their skeleton structure. Bone-lengths vary between subjects and need to be optimized in order to yield good tracking results; we automate this task, which would be too tedious to be carried out manually for each subject.