Ultra Low-Power Biomedical and Bio-Inspired Systems
Massachusetts Institute of Technology
Biological systems are incredibly energy efficient and compute with unreliable and noisy components to perform reliable and precise computations. For example, the brain computes with a staggering energy efficiency of approximately 0.2 femtojoules (fJ) per floating-point operation, and the cell, which is even more energy efficient, uses only 20 kT of energy per active biomolecular operation, i.e., only 8 × 10–20 Joules or 20 units of thermal energy (Sarpeshkar, 2010). Similarly impressive numbers for the energy efficiency of the eye, ear, and organs of the body may be found in the literature (Sarpeshkar, 2010). The creation of circuits that are inspired by biology can lead to novel architectures and systems that have applications outside of biology. Such circuits can also be immediately applied to repair biological systems when they do not work (i.e., in neural prosthetics). This talk illustrates how the synergy between biological and electronic circuits has led to ultra-low-power and noise-robust systems for people who are deaf, blind, and paralyzed and to advanced ear-inspired radio receivers. It concludes with a discussion of cytomorphic or cell-inspired systems that establish an important bridge between electronics and chemistry (Sarpeshkar, 2010). Such systems lay a rigorous foundation for an analog-circuits approach to systems and synthetic biology, fields highly important in the future of medicine and biological engineering.
A rigorous comparison of the pros and cons of analog versus digital computation (Sarpeshkar, 1998) reveals that analog computation, which exploits freely available physical basis functions in the underlying technology that are not necessarily logical or linear to compute, is more energy efficient than digital computation at low precision and vice versa (Sarpeshkar, 1998). Biology exploits this insight to compute in a novel and highly energy-efficient collective analog or hybrid fashion that is not purely digital or purely analog but an intimate combination of both (Sarpeshkar, 2010). Engineering systems can take inspiration from biology to also compute in this fashion, and they can improve energy efficiency by delaying digitization after an optimal amount of analog preprocessing
One example of a bio-inspired collective analog system is the radio frequency (RF) cochlea (Mandal et al., 2009), an electronic chip that takes inspiration from the spectrum analysis of the inner ear or cochlea to create an energy-efficient and ultrafast broadband RF spectrum analyzer. This chip exploits the fact that the ear’s spectrum-analysis architecture is the fastest and most hardware efficient known to man—faster than a digital fast Fourier transform or an analog filter bank. It efficiently maps the partial differential equations that describe fluid membrane-hair cell interaction in the biological cochlea at kilohertz audio frequencies to inductorcapacitor-amplifier interaction in the RF cochlea at gigahertz frequencies. The resulting broadband RF cochlea chip operates with 20-fold lower hardware cost than a traditional analog filter bank or with 100-fold lower power than a system that directly digitizes its RF input to perform spectrum analysis. The RF cochlea is useful as a front end in advanced cognitive or software radios of the future (Sarpeshkar, 2010).
The use of analog circuits to perform energy-efficient spectrum analysis is also useful in bionic ear or cochlear implant processors for people who are profoundly deaf. Cochlear-implant processors compress spectral information present in a microphone signal in a nonlinear fashion such that it is suitable for charge-balanced tonotopic current stimulation of a cochlear electrode array implanted near the auditory nerve. For example, a digitally programmable analog cochlear-implant processor described in the literature (Sarpeshkar et al., 2005) lowered power consumption by 20-fold over a conventional design that performs analog-to-digital conversion followed by digital signal processing; it enabled flexible 86-parameter programming in a patient who understood speech with it on her first try (Sarpeshkar, 2006); it was highly robust to several sources of noise including transistor mismatch, 1/f or pink noise, power-supply noise, RF crosstalk, thermal noise, and temperature variations; and it is at or near the energy-efficient optimum even at the end of Moore’s law. Thus, this processor is amenable to fully implanted and low-cost systems of the future: Its 251-ìW power consumption enables it to function on a small 100-mAh battery with 1,000 wireless recharges for 30 years. A more advanced 357-ìW bio-inspired asynchronous interleaved sampling cochlear-implant processor uses a novel bio-inspired method of nerve stimulation similar to that present in the auditory nerve. It enables fine-time encoding of phase information in a signal without requiring a high sampling rate (Sit and Sarpeshkar, 2008). Hence, it enables music information to be encoded in an energy-efficient fashion without requiring a high number of electrodes or requiring high-stimulation power consumption, a bottleneck in the field of cochlear implants. It is also important for improving speech understanding in noise. Similarly, a companding algorithm inspired by tone-to-tone suppression and gain control in the cochlea has led to improved speech performance in noise (Oxenham et al., 2007; Turicchia and Sarpeshkar, 2005).
Recent work has reported an ultra-energy-efficient adiabatic energy-recycling neural stimulator that can lower power dissipation of nerve stimulation in implants for people who are deaf, blind, or paralyzed, or in other neural, cardiac, or muscle-stimulation applications by a factor of at least 2 to 3 (Arfin and Sarpeshkar, 2011). Such work can be combined with state-of-the-art micropower neural amplifiers that operate near the fundamental limits of physics (Wattanapanitch et al., 2007), with (1) 1 nJ/bit near-field RF telemetry systems that enable transcutaneous wireless bidirectional data transmission in implants (Mandal and Sarpeshkar, 2008), (2) energy-efficient wireless recharging circuits that operate at the limits of physics set by coil quality factors (Baker and Sarpeshkar, 2007), (3) novel highly area and power efficient battery-recharging circuits (Do Valle et al., 2011), (4) highly energy-efficient bio-inspired processors for neural decoding (Rapoport and Sarpeshkar, 2010), (5) highly energy-efficient imagers and novel cochlear-implant-inspired image-processing algorithms (Turicchia et al., 2008), (6) bio-inspired analog vocal tracts for speech and hearing prostheses that perform well in noise (Wee et al., 2011), and (7) blocking-capacitor-free highly miniature precision neural stimulation (Sit and Sarpeshkar, 2007). The integration of several such ultra-low-power and bio-inspired innovations can enable ultralow-power, low-cost, highly miniature, and fully implanted neural prosthetics for people who are deaf, blind, or paralyzed or for people who have other conditions (Sarpeshkar, 2010). Examples of complete working systems that successfully stop a bird from singing via wireless neural stimulation (Arfin et al., 2009), perform wireless recording from a monkey (Wattanapanitch and Sarpeshkar, 2011), and summarize system aspects of design (Sarpeshkar et al., 2008) are given in the literature. Sarpeshkar (2010) discusses the practical engineering constraints needed in such devices.
As these examples illustrate, analog and bio-inspired circuits have enabled and are continuing to enable noise-robust, highly miniature, and ultra-low-power operation in neural prosthetics, a necessity to reduce advanced research to practical clinical applications (Sarpeshkar, 2010). In fact, the deep and astounding mathematical similarities between a form of electronics termed subthreshold electronics and chemistry (Sarpeshkar, 2010) suggest that the impact of electronics on the future of medicine may not be confined to neural, cardiac (Turicchia et al., 2010),or muscular prosthetics but in fact could be much broader: It could encompass a whole new way of thinking about biological circuits–simulating them, designing them, and fixing them.
The average 10-ìm cell is a marvel of nanotechnology, performing 1 × 107 energy-consuming biochemical operations per second in its stochastic, nonlinear, feedback 30,000-node gene-protein and protein-protein network with just 1 pW of power (Sarpeshkar, 2010). Efficient precise computation with noisy components is achieved via clever nonlinear, feedback, and hybrid analog-digital strategies in cells in biology (Hahnloser et al., 2000) as it is in the most advanced ultra-lowpower analog electronic systems of today. Circuits in cell biology and circuits in electronics may be viewed as being highly similar, with biology using molecules, ions, proteins, and DNA rather than electrons and transistors. The striking mathematical similarities between chemical reaction dynamics and electronic current flow in the subthreshold regime of transistor operation, including the Boltzmann stochastics of current flow (Sarpeshkar, 2010), imply that one can mimic and model large-scale chemical-processing systems in biological and artificial networks very efficiently on an electronic chip at time scales that could potentially be a million times faster. This key idea has been built on to show how to create current-mode subthreshold transistor circuits for modeling arbitrary chemical reactions in protein-protein (Mandal and Sarpeshkar, 2009a) and gene-protein networks (Mandal and Sarpeshkar, 2009b; Sarpeshkar, 2010).
The latter work shows that we can potentially attempt to simulate cells, organs, and tissues with ultrafast highly parallel analog and hybrid analog-digital circuits, including molecular stochastics and cell-to-cell variability on largescale “supercomputing” electronic chips. Such molecular dynamics simulations are extremely computationally intensive, especially when the effects of noise, nonlinearity, network-feedback effects, and cell-to-cell variability are included.Stochastics and cell-to-cell variabililty are highly important factors for predicting a cell’s response to drug treatment (e.g., the response of tumor cells to chemotherapy treatments). In turn, analog circuit-design techniques can also be mapped to design and create synthetic-biology circuits that have been shown to be in accord with biological data (Danial et al., 2011). Thus, they can affect the treatment of gene therapies in diseases like cancer and diabetes, or affect the understanding of how such circuits malfunction, thus leading to better drug therapies.
The deep links between energy and information allow one to articulate information-based principles for ultra-low-power design that apply to biology or to electronics (Sarpeshkar, 2010). Engineering can aid biology through analysis, instrumentation, and repair (medicine). Biology can aid engineering through bioinspired design. The positive-feedback loop created by this two-way interaction can amplify and speed progress in both disciplines and shed insight into both (Sarpeshkar, 2010).