An Automated and Cost-Effective System for Early Antimicrobial Susceptibility Testing
Antimicrobial resistance (AMR) is estimated to cause up to 10 million deaths annually by the year 2050, with the largest toll in the developing world. We have developed a deep learning-based system for accelerated, cost-effective antimicrobial susceptibility testing (AST). When tested on Staphylococcus aureus from UCLA Health patients, bacterial growth was detected after an average of 5.72 hours, as opposed to 18–24 hours required by the gold standard method, demonstrating the potential to save precious time when treating resistant infections. The poster presentation of this work is here.
Schottky-Diode-Based Wake-Up Receiver for IoT Applications
We present an always-on ultra-low-power (ULP) wake-up receiver (WuRx) with a two-phase architecture resulting in a 12% reduction in the average power consumption compared to conventional single-phase architectures. The proposed system operates at 750MHz and achieves a low wake-up latency of 200µs, a -50dBm sensitivity at a data rate of 200 kbps, and achieves a FOM~8.5pJ/bit.
Deep learning-based virtual staining of unlabeled tissue samples
We discuss a deep learning-based method which we have developed to help pathologists diagnose diseases such as cancer by eliminating the need for staining of biopsy tissue sections. Using this “virtually staining” technique upon tissue that has not been labelled by any chemicals, these diagnoses are made faster and more affordable. The poster presentation of this work is here.
Pathological Crystal Imaging Using Computational Polarized Light Microscopy
We introduced a single-shot computational polarized-light microscopy (SCPLM) method, which is able to reconstruct the transmittance, retardance, and slow-axis orientation maps of pathological crystals with a single image exposure. We demonstrated this method by imaging and reconstructing the birefringence of crystals in synovial fluid, e.g., MSU and CPPD crystals.
Provably Efficient Exploration for RL with Unsupervised Learning
This work introduces a novel and provably efficient method for reinforcement learning based on unsupervised learning. It is published as a spotlight presentation in NeurIPS (2020). The poster presentation of this work is here.
How does an approximate model help in RL?
In this work, we study how many samples it takes to learn a near-optimal policy in reinforcement learning, provided an approximate prior model. Both upper and lower bounds are established using approximation error, learning accuracy, and model parameters. The poster presentation of this work is here.
Intelligent Material Classification with a Silicon-Based Millimeter-Wave Frequency Comb Receiver
In this work, we present a miniaturized sensor for material detection using their broadband millimeter-wave spectral information. This sensor could classify and predict the type and the thickness of different polymer materials with high accuracy. We hope that this work can lead to more interdisciplinary research, where low-cost integrated circuits are assisted by machine learning techniques for high-performance sensing applications. The poster presentation of this work is here.