Deep Neural Networks (DNNs) are the fundamental component in most artificial intelligence applications. With the increasing number of applications based on artificial intelligence, the performance and energy efficiency of architectures running these algorithms have become crucial, especially for battery-powered platforms. In this work, I propose an energy optimizing memory design framework with a special SRAM/in-memory-computing structure. It also utilizes datapath optimization techniques like quantization and pruning with a fine-level assignment. Compared to other hardware accelerator studies for DNN processing, in this work, I will show that this special memory design, together with the architectural datapath optimization techniques, will have a much better capability of finding the Pareto optimal point in the energy-accuracy trade-off and increase the profitability of the final design.

Biological surfaces are essential for life. Nature employs them as a hub for chemical signalling, active and passive transport of substances, and specialized compartmentation in the building block of life, eukaryotic cell. As such it has been in the main focus of chemist, physicist, and biologist to probe these surfaces with a wide variety of different techniques; from NMR to STORM. This, in fact, is not an easy task since obtaining selective information only from surfaces generally requires to introduce some reporting molecules (chromophore). However, the response of a dye molecule is not always, report on the changes on the biological surface. In this IF proposal, the aim is to enable the probing of biological surfaces, free-floating model membranes, label-free, and non-invasive from the response of interfacial water molecules with second harmonic (SH) response. First, a new high-throughput SH setup will be introduced. The proposed ~3 orders of magnitude enhancement in the SH response will enable to probe at biology relevant (~100 mM) salt concentrations. A big accomplishment for the SH research field. Second, by using this new high throughput setup, two different biological membrane systems and reactions occurring there at will be elucidated, biomimetic lipid droplet organelle, and liposomes, two free-floating membrane model system. Due to its novel aspects, a significant impact is expected from the deliverables of this proposal not only for the scientific field, but also possible intersectoral collaboration prospects as well as opening opportunities for the researcher to achieve a permanent position.

Genomic data carries a lot of sensitive information about its owner such as his predispositions to sensitive diseases, ancestors, physical attributes, and genomic data of his relatives (leading to interdependent privacy risks). Individuals share vast amount of information on the Web, and some of this information can be used to infer their genomic data. Hence, there is a need to clearly understand the privacy risks on genomic data of individuals considering publicly available information on the Web. It is also crucial to protect genomic privacy of individuals without compromising the utilization of genomic data in research and healthcare. The two main objectives of this project are (i) to develop a new unifying framework for quantification of genomic privacy of individuals and (ii) to establish a complete framework for privacy-preserving utilization, sharing, and verification of genomic data under real-life threat models. Graph-based, iterative algorithms previously developed by the applicant to efficiently analyze big data and make inference from it will be the foundation for the new quantification framework. To achieve the holistic genomic privacy objective, cryptographic tools, techniques from information theory, and statistics (differential privacy) will be used. This project will be a significant step towards understanding the privacy risks on genomic data of individuals and protecting the privacy of genomic data. It will also provide a new vision for security and privacy of health-related data in general and will find many implications in other domains such as banking and online social networks. The results of the project will also have an impact on future policies and legislation about protection of health-related data. This EF will have big impact on the future career of the applicant by helping him build new connections, enhance his expertise, increase his visibility in the field of security and privacy, and improve his independent research skills.

Sensor drift is a major problem for inertial sensors and limits their usage in autonomous navigation applications. Inertial sensor data is integrated to find the position and drift leads to error accumulation. A common drift suppression approach is temperature calibration, but ovenized state of the art sensors still exhibit drift. Instead of using temperature as a drift indicator, I have pursued a non-conventional approach and measured on-chip stress that directly correlates with drift. The device interacts with its surroundings through the anchors and on-chip stress accurately estimates drift. I am the leading researcher in the stress compensation field, and I have recently demonstrated that MEMS gyroscope drift could be eliminated with stress compensation. My long-term stability results at 2 days of averaging are unrivaled, but the calibration algorithm is not practical. Different from temperature calibration, stress calibrating a device is difficult. I propose a sensor system that would convert my proof of concept work into a practical 0-drift sensor with self-calibration. The proposed system consists of a circular MEMS sensor with multiple (~100) distributed stress sensors and piezoelectric stress transducers, a machine learning supported analytical calibration model, a custom ASIC for superior noise, and an FPGA for system control and self-calibration. If successful, the proposed approach would improve the MEMS gyroscope stability by >100X to the levels of 10-4 – 10-5°/h, enabling error-free, only gravity-referenced inertial navigation. Unlike GPS or camera, inertial navigation works under all weather, light, and location conditions providing a stable reference to navigation algorithms. With further miniaturization, 0-drift sensors could fit into smartphones, and reliable indoor navigation would become a reality. The compact, low-cost sensor could also disrupt the precision inertial market dominated by bulky and expensive fiber-optic and laser sensors.