Recognizing Dead-End Technologies

Natural Selection has an uncanny way of trying and rejecting technologies that do not yield long-term benefits.

As an example, I contend that the fact that the wheel-and-axle does not occur in nature means that the approach is actually a dead-end.  The fact that signals, nutrients and waste cannot pass through a fully rotating joint means that such a thing cannot have long-term utility for an organism.  

When humans discovered and began using the wheel it was a wondrous, labor-saving device.  However, it is likely that over the long term the overall energy costs and environmental pollution related to use of the wheel will make it non-viable.  The inability to perform lubrication and maintenance across a fully-rotating joint means that we over-engineer the wheel to deal far beyond the worst-case usage.  This means that non-degradable components outlast their normal application. The added burden, in the form of manufacturing costs and reduced resource availability, must be shared among all users.

Nature already has a stable of alternative options that do not involve these limitations.  Sliding joints can provide load-bearing movement on two axes, are self-lubricating, self-repairing, capable of dynamic growth, allow unlimited passage of fluids and signals, and enable the static connection of musculature across the joint.

A corollary to this is the observation that nature will never use a rotary fan or impeller.  If nature needs a pump it will either be based on peristalsis (the bowel) or the bellows for positive-displacement applications such as heart or lungs.

The natural choices exhibit incredibly high energy efficiency, are quiet, long-lasting, operate successfully at all scales as the organism grows, and are capable of functioning at speeds from zero up to their fluid-dynamic limit.

Natural components are created and assembled in-situ, on demand, at room temperature from a small menu of raw materials, adapted immediately to form and fit, and are, ultimately, easily degraded back into reusable materials.  Conversely, every ball bearing ever created will end up in a landfill, taking their valuable, highly-refined raw materials with them.

These aspects should all be considered part of the “total cost of ownership” of a new technology.  Nature has already factored in these line-items as part of the optimization between competing species.  A wise designer would do well to emulate these successes.  

In the case of Vision Systems, I contend that a similar dead-end technology is the shutter.  No naturally-occurring vision system incorporates a shutter.  

The invention of representational art constitutes a seminal development in human history.  It allowed communication of new concepts across individuals and generations.  The discovery of photography extended the concept in wondrous new directions by increasing the accuracy and reducing the energy expenditure involved in creating still images.  

I believe that we are currently straining the applications far beyond their breaking-point by trying to maintain this frame-by-frame approach in two areas.  First (as output), the use of movie-like sequences of still images to trick the eyes into perceiving an intended version of the real world.  And second (as input), trying to interpret frame-by-frame input data from video cameras within machine vision systems.

Specifically we should evaluate the naturally-occurring vision systems and find better, faster and more resource-friendly methods of performing input and output of visual data.

No natural neural processing chain is synchronized. Dynamic events (sensations) propagate signals through neural pathways on demand.  Parallel pathways propagate data at some inherent speed, independently, and only as needed.  

Machine vision systems try to emulate this parallelism in various ways, given the limited available hardware in modern processors and communication channels.  One example is image convolution - an attempt to perform the operations of parallel neurons.

Unfortunately, the approach is fatally flawed.  The parallelism is an illusion, even recognizing the limitations of the “serial simulation of parallelism” imposed by the limited hardware.  The entire process waits for the “next image”.  Then, like the starting gun at a horse race, all the “parallel” processing paths take their new data and try to deal with it.  The vast majority of these processes will will have consumed time and energy but will yield no actionable results.

Richard Feynman would refer to this as Cargo Cult Science: all the trappings of science but none of the substance.  Natural vision systems tightly couple their sensors to the associated neural processing, and this processing is is activated only on demand.  Synthetic vision systems, on the other hand, provide merely the illusion of an equivalence.  Ultimately, efforts to improve the illusion without addressing the fundamental flaw are counterproductive and doomed to fail.

All natural systems are inherently fault tolerant.  Imperfections are dealt with at all stages of growth, training, and active use.  This capability minimizes resource utilization and increases longevity.

Synthetic systems are, in general, anything but fault tolerant.  Manufacturing processes strive for perfection and vast resources are devoted to refining these processes using such metrics as “process yield”.  For example, image sensors with a single defective pixel are discarded.

As another approach, memory arrays are typically designed with a limited number of alternative, redundant rows which can be enabled during manufacturing when a bad cell is detected in the primary array.

These are brute-force techniques which do not address the actual underlying problem, do not address defects encountered during the life of the product, and needlessly increase the overall cost of the system.

Natural systems tend to treat all defects as background noise with respect to the desired signal.  The same noise-mitigation strategies that lead to improved processing of real, noisy, data from the environment will transparently deal with the noise introduced by defects within the organism.  

Instead of requiring a perfect array of regularly spaced pixels, natural systems simply learn the relationship among the photoreceptors by observing the sequence of sensations encountered during normal vision.  This on-going, dynamic updating of the understanding of the properties of the retina (for example) means that overall system performance is continually optimized over the life of the organism.

Furthermore, nature inherently allows for both the fovea and the blind spot.  The retina has nothing even remotely similar to a regular array of equal-sized pixels. Therefore, it allows a high-density, high-resolution region to be combined with a wide field of view without developmental or computational penalty.  And the blind spot introduced in the area of the interconnection  between the movable sensor structure and the optic nerve does no material damage to the system performance, despite the resulting hole in the visual field.

The natural, continuous-learning process stands in sharp contrast to the modern synthetic vision techniques which tend to incorporate a “train it once at the factory, then use it forever in the field” paradigm.  

Again, the current direction appears to represent an obvious dead-end.