AnimalLab
The intention of this web page and accompanying downloads is to implement the concept of "really reproducible research" (Buckheit and Donoho 1995, Schwab et al 2000) for a simulation environment dedicated to the analysis of the behavior, sensory signals, and neural activity of one particular animal, the weakly electric fish. Weakly electric fish are a leading model system for understanding the information processing principles of sensory systems in animals. Some images created with AnimalLab can be viewed on the Uropatagium.
From Open Source to Open Science. We hope that by releasing the code base and data that we've used for our most recent publications, others may benefit in the same way that programmers have benefitted greatly from the open source movement: rather than always reinventing the wheel, people can build on existing infrastructure and develop things that would otherwise be prohibitively time consuming. For example, someone interested in the statistics of natural scenes could use AnimalLab for data on the complete spatiotemporal profile of either the estimated electrosensory signal natural scenes of 116 trials of prey capture data, or the estimated neuronal spiketrain natural scenes of these trials. The effort just to collect the data for this required the better part of a PhD thesis. Similarly, people interested in spike train analysis, detection theory, and non-renewal spiking neuron models could use the synthetic neuronal data in AnimalLab for their studies. Spiketrains of all 14,000 afferents that comprise the active electrosensory input to the blackghost knifefish, over the time course of 116 trials, are included in the afferent module that will be released with its corresponding publication in the winter of 2007. At the behavioral level, someone interested in the kinematics of animal locomotion could make use the motion capture data included with AnimalLab.
Whereas there are mechanisms to foster reproducibility and the cumulative growth of knowledge within molecular neuroscience, such as the sharing of published antibodies, vectors, and transgenic animals, there seems to be no such mechanism in systems-level neuroscience. The mechanism in molecular neuroscience functions quite well normally, in part because it is the policy of key journals in this field that if you fail to share these things from papers they have published, they will no longer publish papers from your lab. No such policy exists in systems-level neuroscience. It would be an enormous boon to our field if more labs made the data and analysis code for their publications available in a usable form, along with mechanisms to quickly find relevant work that has been made open. The
Science Commons initiative of
Creative Commons may be a big help in this regard. AnimalLab is offered under
a Creative Commons License.
Really reproducible research. If you add the
literate programming
movement (Knuth 1984) to the
open source movement, along with a
few concerns over
the reproducibility of results which depend on large complex systems
of software, you arrive at the notion of "really reproducible research."
As David Donoho states on the
WaveLab website,
An article about computational science in a scientific publication is not the scholarship itself, it is merely advertising of the scholarship. The actual scholarship is the complete software development environment and the complete set of instructions which generated the figuresOf course, it has always been a key part of scientific practice that results should be reproducible by others. As published results become increasingly mediated by large complex software systems, this obligatory feature of rigorous scientific practice runs into a variety of difficulties. One is that even if the mathematics behind the computations are made transparent, it will often take an impractical amount of time to encode the math into algorithms to reproduce the results. The code base described here is a good example, having been developed over the course of six years. Further, because so much of the results depend on the settings of various parameters, which the tight space constraints of many journals prohibit the publication of, even if this coding effort is undertaken there is no guarantee that this heroic effort will lead to the reproduction of results. Therefore, providing the source code in a form designed for reproducibility transforms results that are practically irreproducible into results that are easily reproducible. In addition to reproducibility, releasing the code has the further benefit of enabling studies that can utilize this code in some way, as alluded to above. It is a challenge to release code in a readily usable form, but there are now tools available, such as
MATLAB (a multi-platform numerical computation package) and the documentation system
MATLABWEB
(a variant of WEB, the literate programming tool invented
by Donald Knuth), which make it feasible. AnimalLab relies on these
two tools exclusively, except for one case where a costly
commercial package ( RapidForm, by INUS Technology Inc.)
was needed for an analysis in the code for the figures of
"Matched spaces for animal sensation and movement."
The results of this analysis were placed in a MATLAB m-file for generation of
the corresponding figure.
Background. Malcolm MacIver wrote much of the software while he was a doctoral student in
Mark
Nelson's laboratory at the University of Illinois. Mark Nelson has made several contributions
to the code, including the afferent model algorithm.
A significant amount of AnimalLab was developed
while MacIver was a Post Doctoral Fellow with Joel Burdick
at Caltech, including all of the code for analyzing and generating the sensing and motor
volumes of the fish, based on work by MacIver and Burdick on extending the control
theoretic notion of the small-time reachable set to animal movement.
In the short term, AnimalLab will be distributed as an adjunct to publications
that utilize the software, and as a resource for other computational neuroscientists and
neuroethologists. In the long term, it will serve
as a problem solving environment for a monograph on neuromechanics that is
is being written and tested in graduate classes at Northwestern University.
The code (all written in MATLAB) implements a set of empirically tested computational models of the body, sensory receptors, and sensory neurons of the black ghost knifefish (Apternotus albifrons). This is augmented by data for the motion of the fish and the prey during prey capture behavior, and data for the electric field generated by the fish. Specifically, the data, algorithms, and empirically tested models included in AnimalLab are
- a model of electrosensory image formation from Brian Rasnow (1996) (in the stimulus module)
- a high resolution fish body model from MacIver and Nelson (2000) (in the body module)
- a complete population of tuberous electroreceptors (13,857) on the body surface at densities measured by Carr et al (1982), mapped to the surface of the body model as detailed in MacIver (2001) and Nelson et al (2002) (in the sensor module)
- 3D motion capture data of the fish surface and prey location in 116 prey capture trials, measured using the custom motion capture system detailed in MacIver and Nelson (1999), with results in MacIver et al (2001) (in the motion module)
- a 3D electric field data set for A. albifrons, collected by Chris Assad, Brian Rasnow, and Phil Stoddard according to the methods of Rasnow and Bower (1996) (in the stimulus module)
- an algorithm to reconstruct the sensory signal input based on these trajectories, as detailed in MacIver et al (2004, submitted) but similar to the process outlined in Nelson and MacIver (1999) (in the stimulus module)
- an algorithm to reconstruct the full neural input to the animal's brain along the active electrosensory modality (13,857 afferent fibers) during these prey capture trajectories, based on the estimated signal input and a non-renewal model of the afferent spiking dynamics (Brandman and Nelson, 2002) (in the afferent module)
- a motion capture data browser/viewer, which allows the 116 3D prey capture trajectories to be played back at user-specified view angles and playback rates (in the etc/trial_browser directory)
- algorithms to generate the prey sensing volumes and immediate movement volumes of A. albifrons (in the papers/opt1/figs directory, see instructions for generating the figures for "Matched spaces for animal sensation and movement" below)
AnimalLab 1.1 is comprised of 570 program and data files (3.6 GB). A smaller version (100 MB) that requires about 38 hours of computation on a 2.8 GHz Pentium 4 is available, as is an option to have a DVD mailed to you for a nominal fee to cover the cost of a blank DVD and postage: follow the link below. MATLAB R13 (6.5) is needed to run the code. The code has been tested under Linux, Mac OSX, and Windows (Win XP pro and Win 2K).
DOWNLOAD
AnimalLab
Reproducing figures from "Spatial congruence of sensation and action: a general design principle for active sensing?" (submitted). This work uses version 2.0 of AnimalLab, and Sun Microsystem's Grid Engine for distributed computing. AnimalLab 2.0, along with instructions for regenerating all of the figures for the paper will be placed here upon publication.
References
Brandman, R., and Nelson, M. E. (2002). A simple model of long-term spike train regularization. Neural Computation, 14(7), 1575-1597.
Buckheit, J. B. and D. L. Donoho (1995).
WaveLab and Reproducible Research.
In: Wavelets and Statistics, ed. A. Antoniadis. New York, Springer-Verlag: 53-81.
Carr, C. E., Maler, L., and Sas, E. (1982). Peripheral organization and central projections of the electrosensory nerves in Gymnotiform fish. Journal of Comparative Neurology, 211(2), 139-153.
Knuth, D. E. (1984). Literate programming. The Computer Journal 27(2): 97-111.
MacIver, M. A. and M. E. Nelson (2000). Body modeling and model-based tracking for neuroethology. Journal of Neuroscience Methods 95(2): 133-143.
MacIver, M. A. (2001). The computational neuroethology of weakly electric fish: Body modeling, motion analysis, and sensory signal estimation. PhD Thesis. University of Illinois at Urbana-Champaign.
MacIver, M. A., N. M. Sharabash, and M. E. Nelson (2001). Prey-capture behavior in gymnotid electric fish: Motion analysis and effects of water conductivity. Journal of Experimental Biology 204(3): 543-557.
MacIver, M. A., M. E. Nelson, J. W. Burdick (2005). Matched spaces for animal sensation and movement. Submitted.
Nelson, M. E., M. A. MacIver, and S. S. Coombs (2002). Modeling electrosensory and mechanosensory images during the predatory behavior of weakly electric fish. Brain, Behavior, and Evolution 59(4):199-210 .
Nelson, M. E. and M. A. MacIver (1999). Prey capture in the weakly electric fish Apteronotus albifrons: Sensory acquisition strategies and electrosensory consequences. Journal of Experimental Biology 202(10): 1195-1203.
Rasnow, B. (1996). The effects of simple objects on the electric field of Apteronotus. Journal of Comparative Physiology A, 178(3), 397-411.
Rasnow, B., and Bower, J. M. (1996). The electric organ discharges of the gymnotiform fishes: I. Apteronotus leptorhynchus. Journal of Comparative Physiology A, 178(3), 383-396.
Schwab, M., M. Karrenbach, and J. Claerbout (2000). Making scientific computations reproducible. Computing in Science and Engineering 2(6): 61-67.
