An Exploration of Mental Health in US Tech Workers
  • We found many references to a study that reported that ~50% of tech workers suffer from mental health disorders. Our team of data scientists set out to find out if this alarming statistic is backed by evidence and what we can do to improve the state of mental health in the tech workplace and beyond.
  • From our research, mental health issues are still not openly discussed, despite negative impact on work productivity. The good news is that in US tech workplaces, more support for mental health is being provided.

    • A Survey of Computer Vision in Sports
  • Reconstructed end-to-end 3D soccer scenes from soccer videos by combining instance segmentation, human pose estimation, human mesh recovery, sports field localization and camera calibration.
  • Identified the video analysis architecture with the highest mean average precision (mAP) among SlowFast, Inceptionv3, and I3D to determine the labels of actions and activities observed in baseball videos.

    • Dynamics Curriculum Learning for Deep Reinforcement Learning Agents
  • Existing curriculum learning algorithms for RL provide an agent with a sequence of tasks characterized by different initial state distributions or goals. Here, we present another approach to design different tasks -- by varying state transition dynamics, ranging from “easy” environments to “hard” environments that are closer to the real world.
  • Our experiment results show that the utility of curriculum can facilitate learning across three reinforcement learning tasks.

    • Predictions on California Road Accidents Severity
  • We obtained 3 million US accidents records and predicted the severity of car accidents in California state using machine learning methods such as k-nearest neighbors (kNN), decision tree, and support vector machines (SVM). The accuracy of our severity prediction can be 92.7%.
  • Various hyperparameters and machine learning tricks, as well as model differences were further analyzed in depth.

    • Image Classification on 20 Categories Pictures
  • We performed data cleaning, exploratory data analysis, feature extraction and selection, as well as different regression models to build an image classifier on 20 caregories pictures.
  • With logistic regression, k-nearest neighbors (kNN), decision tree, random forests, and support vector machines (SVM), we achieved 43% accuracy. With the convolutional neural network implemented, the accuracy can be even 53%.

    • Joystick with 6-DOF Motion Tracking Function
  • We designed and built a joystick that can be easily wear, and could track the arm motion of the user, as well as display it through the built model. The accuracy, size, weight, and working time could meet the engineering specifications, and the expenditure is much lower than the current products in the market.
  • The gamers can feel that controlling the arm of a character in a game is just like controlling their own arms. With this joystick, actions like waving, hugging or shooting will no longer be pre-programmed and they will be completely carried out by the players.

    • A Two-Legged MEMS Walking Microrobot
  • We designed, simulated, fabricated, and tested a two-legged walking microrobot with electrostatic gap closing actuator based angled-arm inchworm motor using standard two-mask SOI process.
  • This microrobot has in total 2 individually addressable legs with 12 pin-joints and driven by 4 inchworm motors. The robot mass is 135 mg, and dimension is 6.8 mm long, 6.8 mm wide, and 7.5 mm tall. This walking microrobot can achieve 1 mm/s walking speed theoretically, with 500Hz applied electrical signals from external sources via wire connections.

    • Scaling of Vibration Energy Harvester
  • We simulated the scaling benefits of a MEMS electrostatic energy harvester with nonlinear springs.
  • By scaling down the device, the output power per mass can be significantly increased; the needed bias voltage and load resistance can also be reduced.

    • 10nm Bulk and SOI FinFET Transistors Design
  • The 10nm technology node bulk and SOI NMOS FinFET transistors were designed and simulated to meet low power requirements based on specifications from 2015 ITRS using Sentaurus three dimensional TCAD simulator.
  • The tradeoffs between bulk and SOI FinFETs, including electrical and thermal characteristics, were compared and explained.

    • LCD Driver Amplifier Design for Smartwatch Display
  • A LCD driver for a 38-mm smartwatch display was designed to drive all pixels sequentially in one period of the refresh rate.
  • The designed two-stage folded cascode amplifier has high open loop gain, high bandwidth, large phase margin, as well as low settling time and low power consumption when put in feedback loop.