The earliest machines capable of being programmed were mechanical. Then analog programming provided a new level of flexibility and ability to handle more complex instruction sequences at lower cost. Digital processing has further extended this flexibility and capability at ever-lower costs over the decades. However, more efficient energy dissipation, something that "just happened for free" with each generation in digital processing technology nodes, is an increasingly important performance measurement that is more difficult to maintain with each new generation of processors.

Implanted medical equipment is one application area that demands extreme energy efficiency in order to minimize how often batteries that will reside inside a patient must be replaced. Remote and distributed industrial sense and control modules represent another application area where energy efficiency is a key consideration, because gaining access to the physical module, and its power source, is a difficult and expensive endeavor.

For these types of extreme energy efficient designs, some designers and researchers are exploring how to use more analog processing of signals before they are converted into the digital domain. This generally results in a loss of flexibility compared with a pure digital computation model, but if the problem space is constrained enough, it can result in huge gains in energy efficiency.

Another set of researchers is working on yet another type of processing to handle computational problems that are extremely complex, or currently impossible, when managed in just the digital domain. These systems often not only employ a mixture of digital and analog processing, but also are trying to exploit characteristics of physical or mechanical structures to accomplish passive processing or structural processing.

The current Prying Eyes article presents such a structural processing project for sensing sound. In this case, the team is building an artificial cochlea via a MEMS (microelectromechanical-systems) approach. The cochlea is part of the inner ear behind the eardrum. The artificial cochlea passively separates the incoming (complex) sound wave into its component waves, and the passive structure of the cochlea enables the brain to receive the sound without having to perform the heavy computations of an FFT.

Like the prism, the artificial cochlea needs active energy to detect the separated signal, but the magnitude of energy necessary to detect the components of the signal may be much smaller than the magnitude to computationally and digitally perform the same type of signal separation.

In contrast, traditional microphones could be described as imitating the eardrum, where all of the component waves of a sound input are simultaneously recorded at a single point. To identify the component waveforms from a complex sound wave, the application often relies on some type of FFT on the input before it can do much with the input signal.

I believe the use of structural processing elements integrated with electronic sensing and control systems is an emerging trend, especially for dealing with complex real-world inputs.

Consider the "artificial snot" approach being explored by the teams at The University of Warwick and Leicester University. They have found a way to replicate how the natural nose's mucus enhances our sense of smell. The nose's mucus and "artificial snot" both dissolve scents and separate out different odor molecules so that they arrive at the receptors at different speeds and times. The mucus is effectively performing a passive FFT of smells through a chromatography approach. The result is an electronic nose that is better at sensing smells that gave it trouble without the mucus layer.

UC Berkeley biologist Robert Full's work on feet offers another example of passive structural processing at work. His two talks on TED, "Secrets of movement" and "How engineers learn from evolution" expose insights that are relevant to a structural processing approach of building and managing complex sensing and control systems. The robotic examples in his talks look very similar to the Rhex and RiSE robots at Boston Dynamics. Unfortunately, the last time I approached Boston Dynamics, they did not want to discuss their technology. I will keep trying to gain their cooperation for a future article.

Also, I am working with the single-pixel camera team at Rice University to better understand what they are doing. I think it is a related structural-processing approach, but I have not been able to confirm that yet. If these different groups are open to discussing their work, I hope to follow up with a comprehensive feature article about these projects and structural processing.

If you are aware of any other examples of research, industrial, or commercial use of structural processing, please let me know either through an email to me or by posting in the comments below. I outlined in a past story that embedded designs are incorporating more sensors than ever before and using them together in a correlated fashion to establish a more accurate understanding of the real world around them. As the economics allow including many more sensors in a single design, I expect these types of systems will proliferate and open doors to yet only imagined capabilities.