The Database for Reaching Experiments and Models

Ben Walker, Konrad P. KördingPLoS 2013 · DOI

Abstract

Reaching is one of the central experimental paradigms in the field of motor control, and many computational models of reaching have been published. While most of these models try to explain subject data (such as movement kinematics, reaching performance, forces, etc.) from only a single experiment, distinct experiments often share experimental conditions and record similar kinematics. This suggests that reaching models could be applied to (and falsified by) multiple experiments. However, using multiple datasets is difficult because experimental data formats vary widely. Standardizing data formats promises to enable scientists to test model predictions against many experiments and to compare experimental results across labs. Here we report on the development of a new resource available to scientists: a database of reaching called the Database for Reaching Experiments And Models (DREAM). DREAM collects both experimental datasets and models and facilitates their comparison by standardizing formats. The DREAM project promises to be useful for experimentalists who want to understand how their data relates to models, for modelers who want to test their theories, and for educators who want to help students better understand reaching experiments, models, and data analysis.

Introduction: The Fragmentation of Motor Control Research

In the study of motor control, "reaching" serves as the quintessential experimental paradigm. However, the field currently suffers from significant fragmentation. Most publications present a single experiment paired with a single model. Because every laboratory typically employs its own proprietary data formats and naming conventions, it is exceedingly difficult for researchers to test a new model against external data.

This "insularity" prevents the rigorous falsification of models and slows scientific progress. If a model only succeeds on the specific dataset for which it was designed, it may not be capturing a fundamental principle of neural control. The Database for Reaching Experiments and Models (DREAM) addresses this by creating a "common language" for reaching data, facilitating comparisons across labs, species (humans and non-human primates), and experimental conditions such as force fields and visual rotations.

The DREAM Framework and Methodology

The authors developed a curated database and a suite of software tools built on the MATLAB platform. The core innovation of DREAM is the transition from inconsistent laboratory-specific formatting to a unified, consistent format that allows for automated analysis and model testing.

The DREAM workflow: converting inconsistent lab data into a standardized format for models and experiments.

The project utilizes a standardized, hierarchical data structure to organize complex experimental information:

  • Standardization: All data is converted to SI units (meters, seconds, Newtons) to eliminate unit-mismatch errors.
  • Hierarchy: Data is organized logically, starting from the Subject level (including metadata like age and gender) down to individual Trials containing time-series kinematics.
  • Extensibility: A Special field allows researchers to include non-standard data, such as auditory cues or neural spikes, without compromising the core format.
The hierarchical data structure of the DREAM database, organizing experiments by publication, task, subject, and trial.

By treating model predictions as "first-class citizens"—storing them in the same format as experimental data—DREAM simplifies the workflow for model-data comparison. A "Run Model" tool takes a standardized dataset as input and produces a standardized prediction, allowing the same visualization and analysis tools to be used for both.

Database Composition and Benchmarking

At its launch, DREAM included 39 experiments sourced from 16 different publications across 15 institutions. These datasets cover diverse conditions, including visuomotor rotations, force-field adaptation, and uncertainty tasks. Beyond simple kinematics, some datasets include multimodal data such as fMRI scans, EMG (muscle activity), and neural spike timestamps.

To demonstrate the utility of the database, the researchers ran several models against the entire repository. For instance, the Berniker_2011_m1 force-field adaptation model performed well on experiments involving physical perturbations, reducing error by approximately 51% compared to a non-learning model. However, the same model generalized poorly to visual rotation tasks, highlighting that models of physical adaptation do not inherently explain visual adaptation.

Discussion and Future Directions

DREAM lowers the barrier to entry for researchers and students alike. Standardized data acts as an educational tool, allowing those from adjacent fields like robotics or machine learning to engage with high-quality motor control data without writing custom parsers for every paper.

While the database is currently rich in experimental data, the authors emphasize the need for more community-submitted computational models. Future iterations of DREAM aim to reduce platform dependency (currently tied to MATLAB) and introduce tools to automatically reproduce specific figures from original publications, further enhancing the transparency and reproducibility of motor control research.

Summary updated

Abstract

Reaching is one of the central experimental paradigms in the field of motor control, and many computational models of reaching have been published. While most of these models try to explain subject data (such as movement kinematics, reaching performance, forces, etc.) from only a single experiment, distinct experiments often share experimental conditions and record similar kinematics. This suggests that reaching models could be applied to (and falsified by) multiple experiments. However, using multiple datasets is difficult because experimental data formats vary widely. Standardizing data formats promises to enable scientists to test model predictions against many experiments and to compare experimental results across labs. Here we report on the development of a new resource available to scientists: a database of reaching called the Database for Reaching Experiments And Models (DREAM). DREAM collects both experimental datasets and models and facilitates their comparison by standardizing formats. The DREAM project promises to be useful for experimentalists who want to understand how their data relates to models, for modelers who want to test their theories, and for educators who want to help students better understand reaching experiments, models, and data analysis.
BibTeX
@inproceedings{walker2013database,
  author = {Ben Walker and Konrad P. Körding},
  title = {{The Database for Reaching Experiments and Models}},
  booktitle = {PLoS},
  year = {2013},
  doi = {10.1371/journal.pone.0078747},
}
Figure