Details of each research project can be found in pdf slides attached to each description or can be found under the publications page.

With advances in fabrication technology, an ever increasing number of sensors can now be integrated within a single package. Examples of such sensors include microelectrode arrays and microphone arrays. In a classical data acquisition paradigm each sensor channel is treated independently with regard to analog-to-digital conversion and any intelligent processing is delegated to a backend digital signal processor. However, in micro/nano-scale high-density sensor arrays, the detrimental effect of finite resolution on signal processing cannot be ignored, especially if the objective is to mine hidden/weak signatures in high-dimensional analog data. Our aim is to explore techniques for integrating machine learning algorithms directly with analog-to-digital conversion such that resolution can be dynamically allocated across sensor channels to achieve optimal system performance. 

Design of High-dimensional Oversampling Data Converters with On-chip learning:  Theory, Algorithms and Hardware Realization (Ph.D Thesis of Amit Gore) 

PDF slides (High-dimensional Delta-Sigma Converters)

PDF slides (Auditory surveillance)

Keywords: Multiple-input Multiple-output Analog-to-digital Converters,
Acoustic Sensor Arrays, Sigma-delta learning, High-dimensional quantizers, Analog-to-information converters. 

Students:  Amin Fazel

Sponsor: National Institute of Health, Johns Hopkins Applied Physics Laboratory.

---

Energy Scalable Decoders for Low-density Parity Check Codes

In analog and digital VLSI, design methods typically follow a top-down approach where proven algorithms are mapped onto optimized computing hardware. In this research, we are investigating a bottom-up approach where computing paradigms are inherent in hardware can be used for designing specialized algorithms. This is particularly relevant for analog VLSI (aVLSI) where the computational hardware is always approximate and is heavily dependent on the device physics. Our objective is to use universal conservation principles (charge, mass or current) to approximate well known functions and then suitably modify the algorithms to achieve the desired or even superior performance. We are using a novel approximation technique called "margin propagation" to design  energy scalable LDPC decoders that can be used for a wide variety of applications ranging from sensor networks to wireless base-stations. In this regard, this study is unique because it investigates the three way trade-off between energy efficiency, signal-to-noise ratio and bit-error-rate performance, where as traditional approaches only focus on SNR and BER trade-offs. We are also translating the findings of the theoretical study into practical implementations on silicon.  

PDF slides I (Margin propagation analog decoders)

PDF slides II (Nano-powered integrated circuits)

Keywords: Margin propagation, low-density parity check codes, analog computing, low-power communications, density evolution, factor graph 

Students:  Ming Gu

---

Forward Error Correcting Biosensors

Environmental variability and stochastic interaction between bio-molecules are major causes that affect the reliability of existing and emerging biosensors. In this project our objective is to develop hybrid bio-CMOS techniques, that combines sensitivity of biological sensors with reliable computation on silicon to improve the performance of existing biosensor. As a model, we have chosen a disposable polyaniline based immunosensor that is relatively easy to fabricate and test. Our approach is to use source-channel coding principles to embed forward error-correction on the biosensor and decode the output produced by the sensor using a reliable silicon based processor.   

PDF slides

Students:  Yang Liu,  Ming Gu

---

Energy harvesting circuits and systems

In this project we are investigating circuit and sensor topologies that can harvest the power directly from the signal being sensed. As a result such circuits and systems can be used for long-term surveillance and monitoring where batteries are impossible to deploy and remote power delivery methods are impractical. In particular, we are developing floating-gate circuits that can be used for long-term health monitoring of biomechanical implants (for example, hip or knee implants) with operational life greater than 10 years.   

PDF slides  (Self-powered floating gate sensor)

PDF slides (PLLA-Carbon Nanotubes strain sensor)

Keywords: Self-powering, floating-gates, energy harvesting, nanowatt, structural health monitoring, implantable sensors. 

Students:  Chenling Huang,  Yang Liu

---

Speech Biometrics

We are investigating real-time learning and signal processing algorithms that can be used for localization, separating and identifying acoustic targets of interest in the environment. This effort is divided into three parts: (a) microphone array and source localization; (b) analog-to-feature extraction for direct acquisition of speech features without any intermediate power-consuming analog-to-digital conversion; (c) noise-robust speech feature extraction; and (d) robust detection algorithms that can identify specific targets in the acoustic field.    

PDF slides

Keywords: Speaker verification, speaker identification, microphone arrays, source localization, source separation, support vector machines, classification, analog-to-feature conversion, sigma-delta learning, kernel features. 

Students:  Amin Fazel,  Ravi Shaga