Affine Covariant Image Features and Description: Images as documents via bag of visual words.Distributed Map-Reduce with Centralized Control: K-means++; accelerated K-means by triangular inequality with Hellinger distance for visual vocabulary quantization.Deep Boltzmann Machines for Image Sentiment Annotation: Experimental sentiment analysis for visual objects.Joint Indexing of Linguistic and Visual Vocabulary by Inverted Files: Threaded inverter with periodic flush; Packed-integer codec for visual words storage; TF-IDF based similarity and L1- norm scoring.Range Query by In-Memory KD Tree Redis Cache: Supporting browser Javascript for visual object cropping without vision library; Injection into iOS Web View for mobile clients.NGINX as Image File Server with Cascade Sub-requests to Redis and Lucene Upstreams: Decouple Web service into Redis range server and Jetty Lucene server for scalability and availability.

Probabilistic latent variable models as non-negative matrix factorization.

Analytics: latent semantic analysis, collaborative filtering methods, naïve Bayes classifier with kernel density estimation, Poisson-Gamma mixture.

Hierarchical Beta-Binomial model for online product shopping conversion prediction. Product conversion ranking by shrinkage estimates. Categorical conversion estimation by logit-regression. Aggregated seller conversion as co-variate for prediction. Bootstrap methods for product shopping ROI optimization strategy comparison. Bias correction for pseudo-random allocation. Bootstrap p-value computation.

Technology development in the area of machine learning for digital image understanding, visual object detection and recognition, as well as image indexing. Reduced 1-class SVM as object screening by Walsh-Hadamard projection; Reduced 2-class cascade SVM as object detector. Object recognition by pair-wise coupling SVM classifiers and multi-category SVM. Multi-category SVM training by homogeneous self-dual IPM and large-scale sparse LU decomposition.Image pre-processing: variability reduction methods for object detection and recognition. Retinex by two bilateral filters with accelerated slice interpolation; Fast convolution by integer-packing; Multi-slice processing by histogram segmentation; Histogram mode identification by nonparametric isotonic testing.

Algorithms development in machine learning and pattern recognition, statistical analysis and modeling for gene expression data.Supervised learning methods for biomarker discovery: multi-category SVM (MSVM) training, data adaptive tuning. Recursive feature elimination (RFE) and feature ranking method for biomarker discovery. K-nearest neighbor (kNN) classification methods for expression profile classification and missing data imputation. Microarray data exploration and visualization: multidimensional scaling (MDS) method and principal component analysis (PCA) method for expression profile visualization and dimensionality reduction. Gap statistics for K-means clustering, bagging and bootstrapping K-means clustering algorithm for cluster confidence measure.Expression and differential expression detection method comparison: methods to estimate false positive, true positive and total positive. ROC comparison for MBEI, Rosetta error model, RMA (robust multiarray) method, GeneSpring error and P-fold algorithm. False Positive Rate (FDR) and Multiple Test Correction: FDR estimation method by q-value, Benjamini-Hochberg correction for p-values, and permutation t-test for FDR estimation.

Software development in the area of Bayesian data analysis via Markov chain Monte Carlo, survival analysis and nonparametric maximum likelihood methods, and financial time series modeling via GARCH models.C++ libraries for Markov chain Monte Carlo methods and Bayesian data analysis for S-PLUS: 1) Bayes hierarchical models, 2) generic samplers with Metropolis-Hastings, Gibbs, and random direction interior point methods, 3) generalized linear mixed models.Java libraries for nonparametric MLE-based survival analysis software: 1) Cox models with right, doubly and interval censored data, 2) Nonparametric survival distribution estimation with doubly and interval censored data, and truncated data, 3) pseudo-chunking algorithms for fitting Cox model with large data sets. C libraries for radial-spherical Monte Carlo methods for high dimensional integration : random rotations on sphere and adaptive Gaussian quadrature on radial dimension. Linear mixed effect model fitting with a large number of random effects.C libraries for diagnostic methods for B-spline hazard regression with censored data: Cox-Snell residuals with hazard regression models. C libraries for change detection methods for time series models: level-shift, slope change and their combination. FORTRAN libraries of GARCH models for financial time series modeling and analysis: 1) Adapting FORTRAN core for GARCH models for S-PLUS interface, 2) S-PLUS graphical applications with GARCH modeling, 3) Help utility and documentation for S-PLUS GARCH module.