Projects

IKBT: the Autonomous Inverse Kinematics Solver

Symbolic inverse kinematics analysis solves the problem of how to control the robot joints to achieve desired end effector location. Up till now, closed-form symbolic IK analysis required human experts to apply their high-level mathematical reasoning needed. We developed a system, called "IKBT", to automate this process with Behavior Tree, providing better modularity and versatile node types needed for handling high-cognitive problems such as IK analysis. The results showed that IKBT is capable of solving up to 6 degree-of-freedom (DOF) robot and rendering closed-form symbolic solutions without supervision or human input, all in a matter of seconds.

Read More

Predicting Parkinson's Disease from Voice Recordings

It is estimated that over 7 million individuals worldwide are affected by Parkinson’s disease, whose degenerative nature makes early detection and diagnosis vitally important; however, effective diagnosis is presently an expensive and time-consuming process.  Here is presented an attempt to use features extracted from speech recordings to diagnose both the onset and severity of Parkinson’s.  Differentiation between healthy and diseased subjects was highly successful, with a 0/1 loss percentage of 5.2% for the most successful algorithm.  Less successful was regression to determine UPDRS scores: with a best average difference of approximately 7.3 for prediction of motor UPDRS scores and 9.5 for total UPDRS scores, regression was useful for a ballpark estimate but insufficient to serve as a substitute for clinical evaluation.  Ultimately, although the regression techniques applied proved insufficient to predict the severity of Parkinson’s patients’ symptoms separate from clinical evaluation, classification of Parkinson’s patients vs. healthy individuals through speech recordings shows great promise as a rapid, inexpensive method to determine whether the full, clinical diagnosis procedure is warranted.

Read More

Face and Emotion Recognition with Eigenfaces

How Facebook/Amazon Prime Photos/Google+ knows it is you in the photo without you tagging photos one-by-one? Here is how.  In this project, I implemented a set of algorithms to classify faces and emotions.PCA was used to reduce data dimension. Several classification algorithms were implemented and compared, including linear regression with least square, logistic regression, stochastic gradient descent (SGD), RBF kernel with large feature space combined with SGD (state-of-the-art classifier!!), and neural network. All codes were written from scratch, used basic numpy library. <1% 0/1 loss was achieved.

Read More

State of the Art Classification of Hand-written Digits

Identification of hand-written digits (MNIST) using PCA and stochastic gradient descent with RBF kernel (non-trivial feature space: 60k features).  < 2% test set 0/1 loss was achieved.

Note that the large dimension of features means it’s important to construct matrix on the flight instead of explicitly declare the entire mapping at the same time, it may cause numerous issues including insufficient memory and inefficient computing.

Autonomous Pacman Agent

AI agent that plays classic Pac-man game better than me. The automatic decision-making process involves: search algorithms(DFS, BFS, A* search) for navigation, probabilistic reasoning (MDP, HMMs, reinforcement learning, and particle filter) for sensing and dodging ghosts. The platform is developed by Berkeley's AI course.

FIQ – Fluorescent Image Quantifier

If you are a biologist working with microscopy images, FIQ is what you need to save your time and sanity. FIQ the “Fluorescent Image Quantifier” is a software suite to process large amount of fluorescent images in a short time. The first FIQ prototype was born in March 2015, several updates have been made since and currently under beta testing.