Naturally Supervised Learning, new arXiv paper

Here's a paper of mine recently accepted to the arXiv, cross-listed under the cs.HC and q-bio.NC categories:

Naturally Supervised Learning in Manipulable Technologies

This is more work from the physical intelligence project, and it features three experiments I did when the MIND Lab was still up and running:

MIND Lab Website

Here is the abstract and key points:

Objective
It will be argued that haptic and proprioceptive sensory inputs serve a supervisory function in movement production related to the control of virtual environments and human-machine interfaces. To accomplish this, an approach new to human factors called neuromechanics will be used. This involves the introduction of novel techniques and analyses which demonstrate the multifaceted and regulatory role of adaptation in interactions between humans and motion and touch-based (e.g. manipulable) devices and interfaces.

Background
Neuromechanics is an approach that unifies the role of physiological function, motor performance, and environmental effects in determining human performance. In this paper, a neuromechanical perspective will be used to explain the supervisory role of environmental variation on human performance.

Method
Three experiments are presented using two different types of virtual environment that allowed for selective perturbation. Electromyography (EMG) and information related to kinematics were collected. Measures related to human performance dynamics were used to model the results.

Results and Conclusions
Results presented here provide a window into neuromechanical performance under a range of technologically-mediated conditions. Both descriptive and specialized analyses were conducted: peak amplitude analysis, loop trace analysis, and the analysis of unmatched muscle power. These analyses demonstrated that there are myriad consequences to force-related perturbations related to dynamic physiological regulation.

Applications
The findings presented here could be applied to the dynamical control of touch-based and movement-sensitive human-machine systems. In particular, the design of systems such as human-robotic systems, touch screen devices, and rehabilitative technologies could benefit from this research.

Key Points
* emerging manipulable technologies (e.g. touch- and motion-based interfaces) will ultimately feature a number of non-uniform forces and sequences of stimuli that can be simulated using virtual environments with physical intermediaries.

* the dynamics and complex relationships between simulation, the physical world, and human physiology can be better understood through the lens of neuromechanics, an approach that unifies biomechanics, neuroscience, embodied perspectives, and systems engineering.

* it was found that selective perturbation, in relation to a staggered training protocol can uncover various differences in performance, which remain to be formally classified but are suggestive of underlying cognitive and morphological regulatory mechanisms.

* these findings can be integrated with existing mobile and virtual technologies to provide a versatile, programmable tool for rehabilitative and non-medical applications.

Be sure to take a look. As always, comments are appreciated.

Maker Faire, Detroit

Here's an interesting upcoming event (happening later this Summer). Unfortunately, I will not be presenting anything due to lack of spare time:

Maker Faire, Detroit 2011

Looks good, and if Adam Savage can fool around inside of a Faraday cage for the cause, it's gotta be good.

Adam Savage, Tomfoolerier*

* scene from SF Maker Faire, May 2011

Relativistic Virtual Worlds, new arXiv paper

Recently, I submitted a paper to arXiv, cross listed between the cs.HC, cs.CG, and q-bio.NC categories. A link to the paper and abstract are presented below:

Link to paper

In this paper, I will attempt to establish a framework for representation in virtual worlds that may allow for input data from many different scales and virtual physics to be merged. For example, a typical virtual environment must effectively handle user input, sensor data, and virtual world physics all in real- time. Merging all of these data into a single interactive system requires that we adapt approaches from topological methods such as n-dimensional relativistic representation. A number of hypothetical examples will be provided throughout the paper to clarify technical challenges that need to be overcome to realize this vision.
The long-term goal of this work is that truly invariant representations will ultimately result from establishing formal, inclusive relationships between these different domains. Using this framework, incomplete information in one or more domains can be compensated for by parallelism and mappings within the virtual world representation. To introduce this approach, I will review recent developments in embodiment, virtual world technology, and neuroscience relevant to the control of virtual worlds. The next step will be to borrow ideas from fields such as brain science, applied mathematics, and cosmology to give proper perspective to this approach. A simple demonstration will then be given using an intuitive example of physical relativism. Finally, future directions for the application of this method will be considered.


This is part of a body of work in an area I am calling Physical Intelligence (the interaction between the human cognitive/biomechanical substrate and manipulable virtual worlds), and is definitely a work in progress. Please have a look. As always, comments are welcome.

Machinery of Biocomplexity, new arXiv paper

Yesterday, I submitted a paper to the arXiv entitled The "Machinery" of Biocomplexity: understanding non-optimal architectures in biological systems, cross-listed between the nlin.AO, physics.bio-ph, and q-bio.QM categories. Here is a link to the paper, and abstract, and a summary figure:

The "Machinery" of Biocomplexity: understanding non-optimal architectures in biological systems

One popular assumption regarding biological systems is that traits have evolved to be optimized with respect to function. This is a standard goal in evolutionary computation, and while not always embraced in the biological sciences, is an underlying assumption of what happens when fitness is maximized. The implication of this is that a signaling pathway or phylogeny should show evidence of minimizing the number of steps required to produce a biochemical product or phenotypic adaptation. In this paper, it will be shown that a principle of "maximum intermediate steps" may also characterize complex biological systems, especially those in which extreme historical contingency or a combination of mutation and recombination are key features. The contribution to existing literature is two-fold: demonstrating both the potential for non-optimality in engineered systems with "lifelike" attributes, and the underpinnings of non-optimality in naturalistic contexts.

This will be demonstrated by using the Rube Goldberg Machine (RGM) analogy. Mechanical RGMs will be introduced, and their relationship to conceptual biological RGMs explained. Exemplars of these biological RGMs and their evolution (e.g. introduction of mutations and recombination-like inversions) will be demonstrated using block diagrams. The conceptual biological RGM will then be mapped to an artificial vascular system, which can be modeled using microfluidic-like structures. Theoretical expectations will be presented, particularly regarding whether or not maximum intermediate steps equates to the rescue or reuse of traits compromised by previous mutations or inversions. Considerations for future work and applications will then be discussed.

Example of the "cooption" and "inversion" scenarios described further in the paper.

This paper introduces a new metaphor for understanding non-optimality in the evolution of complex traits. This metaphor is inspired by the mechanical Rube Goldberg machine  (or mechanical RGM as I characterize it in the paper).

In mapping this metaphor to biological systems, I introduce the concept of "maximum intermediate steps" in the function of a trait (traits that are not optimized with respect to evolution, structure, or function), which can be characterized by biological versions of the RGM. I have written about this idea in one of the first posts in this blog, but this is a more formal computational look at how this might work in nature.

As always, I would appreciate feedback.

Nathan Sawaya, LEGO Artistry

I have been getting into the work of Nathan Sawaya, who makes sculptures out of LEGO bricks


Nathan Sawaya at work

Check out pictures of his work here.