June 13, 2013

Human Augmentation Short Course -- part I

I have been posting many odds and ends I have in my files on the topic on the science of human performance augmentation to my micro-blog, Tumbld Thoughts. After a few posts in this area, I decided to try a Tumblr-based short course composed of "flash" lectures (an innovation I am experimenting with). 

Below is a series of posts that constitute the first part of this short course (mostly introductory concepts). These short lessons can also be found (in their original context) under the #human-augmentation tag on Tumblr. Comments would be appreciated.

I. Basics of Performance Mitigation

In a previous post, I discussed an area of science and engineering called Augmented Cognition. But how does one “augment” cognition? In the modern version, a mitigation strategy is used to enforce optimal performance in a manner similar to supervised learning.

* this requires a well-characterized response function. One example is theYerkes-Dodson law (inverted curve) that characterizes arousal.

* mitigation proceeds using a first-order linear control strategy with feedback. This is similar to the pole-balancing task widely used in robotics.

* to make these judgments, measurements of physiological state are used. In these cases, the measurements determine when there is “too much” or “too little” of a physiological state. For example, “too much” activity in a certain part of the brain leads to an unacceptable amount of arousal, which can then be mitigated using a visual stimulus on a heads-up display.

* but what about more complex responses and the role of adaptation? This will be covered in the next post.

* all images taken from the following lecture: Alicea, B.   Behavioral Engineering and Brain Science in Virtual Reality. Figshare, doi:10.6084/ m9.figshare.155710 (2013).

II. Performance Mitigation vs. Optimization

In my last post on human augmentation, I discussed what a mitigation strategy is. Now I would like to discuss the role of natural variation in human augmentation and mitigation. While some people might view the products of natural selection to be “optimal”, natural variation actually works against engineering optimization in a number of key ways.

* to fully account for natural variation, we must open up the black box of physiological regulation. To do this, we must understand the process of humans interacting with technology as a homeostatic or allostatic process [1].

* physiological regulation as a result of technological interaction occurs at multiple biological scales. These include the molecular bases of learning and memory, tissue-specific gene expression, cognitive memory consolidation, and populations in their environment.

* the nature and potential outcomes of this process can be captured using a fitness landscape or related type of n-dimensional phase space [2]. This allows us to understand the adaptability of specific genotypes or populations of individuals [3].

* the use of fitness landscapes allows us to characterize allostatic regulation as a hill-climbing (or quasi-optimal) process. However, in doing so, we must account for certain regularities of training such as the power law of practice.

III. Role of Natural Variation in Performance Augmentation

In the last two #human-augmentation posts, I discussed the role of mitigation strategies and natural variation in human augmentation. In this post, we will explore an experimental paradigm for training in novel environments that produces a chaotic output via stochastic resonance [4]. As a chaotic output, it can be controlled by various types of feedback [5].

We can use a motion-controlled sports videogame to simulate various human movement regimes (e.g. swinging, reaching). A weighted instrument (e.g. misshapen baseball bat with a forcing chamber [6]) was used to introduce chaotic motion during performance. By switching between this distortion and normal gameplay, we have created an environmental switch that can induce natural variation and the biological substrates that underlie performance. 

This type of environmental switch is found in nature as brain-related preconditioning [7] in humans, or as a means to speed up adaptation in a given population [8]. By presenting each type of movement regime in different sequences, we can augment performance under normal circumstances or control chaotic fluctuations [9].

IV. Review of Performance Mitigation Architectures

Two posts ago in the human augmentation thread, we were introduced to the role of allostasis and first-order linear control in correcting (e.g. mitigating) sub-optimal behaviors related to human performance. In this post, we will explore this theme further using architectures that adaptively control (e.g. augment) optimal levels of cognition and human performance.

The first architecture is the simple feedback with band-pass filter. This is often used to mitigate performance profiles that conform to the inverted U (e.g. arousal). The bandpass filter implements a simple rule used as feedback that reinforces parameter values within a certain range. This first-order linear control manages a unimodal response function as a signal-to-noise problem without excessive computational overhead.
But what about cases where our measure exhibits a greater number of measures? The second architecture demonstrates the simple feedback motif as a parallel array (in this case, two arrays) that contribute to a global assessment of performance (long rectangle). In this case, mitigation is treated as an optimization problem rather than a signal-to-noise problem. This allows us to search for optimal mitigation configurations on a n-dimensional landscape rather than extracting one-dimensional signal from noise.

In cases where the contributions of physiological variance are great, from example in systems which are not well-understood, we can use something called I call an allostatic control architecture. This type of model accounts for a dynamic physiological background as it interacts with performance embedded in its environment. To enforce this type of control, environmental switching [10] can be used. In this case, there is no feedback, but there is a linear filter that enforces a threshold on the response to both sets of environmental conditions. Levels of performance that are robust in both environments are selected for using the filter, and treats mitigation as a problem of stability during adaptation. 

[1] For the concept of allostatic regulation, please see: Schulkin, J.   Rethinking Homeostasis: allostatic regulation in physiology and pathophysiology. MIT Press (2003).

For the concept of technology being an environmental challenge, please see: Alicea, B.    Performance Augmentation in Hybrid Systems: techniques and experiment. arXiv Repository, arXiv:0810.4629 [q-bio.NC, q-bio.QM] (2008).

[2] For more information on the geometry of fitness landscapes, please see: Gavrilets, S.   Fitness Landscapes and the Origin of Species. Monographs in Population Biology, 41. Princeton University Press (2004).

[3] Populations can consist of special needs populations, different ethnic groups, and even people of differing body shape and athletic ability. They carry unique molecular, physical, and cognitive features to specific types of interaction.

For more information, please see: 
Alicea, B.   Natural Variation and Neuromechanical Systems. Cogprints, 6698 (2009).

Alicea, B.   The adaptability of physiological systems optimizes performance: new directions in augmentation. arXiv Repository, arXiv: 0810.4884 [cs.HC, cs.NE] (2008).

[4] Alicea,  B.    Stochastic  Resonance  (SR)  can  drive  adaptive  physiological  processes.  Nature  Preceedings, npre.2009.3301.1 [http://precedings.nature.com/documents/3301/version/1] (2009). 

[5] Hunt and Johnson   Keeping Chaos at Bay. IEEE Spectrum, 30(11), 32-36 (1993).

[6] A forcing chamber is a container filled with a liquid or other material (with a specific gravity) to create a distorted radius of gyration during a swing, a reach, or a stroke.

[7] Gidday, J.M   Cerebral preconditioning and ischemic tolerance. Nature Reviews Neuroscience, 7, 437-448 (2006).

[8] Kashtan, N., Noor, E., Alon, U.   Varying environments can speed up evolution. PNAS USA, 104(34), 13711-13716 (2007).

[9] Ott, E., Grebogi, C., and Yorke, J.A.   Controlling Chaos. Physical Review Letters, 1196-1199 (1990).

Ott, E.   Controlling Chaos. Scholarpedia, 1(8), 1699 (2006).

[10] Shifting between environments continually or in a patterned way. See the last #human-augmentation post for more.

No comments:

Post a Comment