Ying Zhang

• Addressed length-compressed diffusion language modeling, where prior two-stage pipelines fix an autoencoder before training the generative model, by jointly training a compressor, a Flow Matching model, and a token decoder end to end so the compressed representation adapts to generation.

• Coupled JEPA-style joint-embedding prediction with Flow Matching by using the Flow Matching model itself as the predictor at the final noise step, so one network both predicts and denoises the compressed embedding during generation.

• Benchmarked against seven autoregressive, discrete-diffusion, and flow-based baselines including ELF on LM1B and OpenWebText with five-seed evaluation, achieving the lowest generation perplexity and highest throughput among all diffusion and flow models — 26.20 Gen-PPL on OpenWebText at 2.4× ELF's throughput (10,511 vs. 4,416 tokens/s).

• Formalised the precursor illusion, where pre-onset labels are defined by temporal position alone and point adjustment is applied, causing reactive detectors to appear to deliver advance warning they never had.

• Designed a model-agnostic, placebo-calibrated test for whether pre-onset intervals actually contain detectable evidence, finding that fewer than one in seven benchmark events qualify once the placebo baseline is subtracted.

• Showed that after multiplicity correction none of 35 model–dataset cells across 7 detectors discriminates precursor from matched normal windows, while point adjustment inflates mean detection F1 from 0.186 to 0.549 with scores and thresholds held fixed.

• Derived a closed-form Klein batch normalisation that replaces the iterative Fréchet mean required by prior gyrogroup BN with a single-pass Einstein midpoint, removing the inner optimisation loop at negligible accuracy cost.

• Developed the first systematic set of intrinsic Klein layers, spanning classification, fully connected, and normalisation heads in both geodesic and horosphere form, from a single Klein hyperplane distance. The model had stayed unexplored while the field standardised on the Poincaré ball and Lorentz hyperboloid, despite being the only one of the three admitting both Euclidean geodesics and a closed-form midpoint.

• Benchmarked against Poincaré, Lorentz, and PV counterparts on image classification (CIFAR-10/100, Tiny-ImageNet, ImageNet-1k), graph link prediction, and node classification, where Klein matches or improves on all three while remaining stable under feature clipping, curvature mis-specification, and gradient scaling.

• Built a 3-layer DNN for simulation-based inference integrating normalizing flows, sufficient statistics, and CLT structural regularization, simulated gamma & beta datasets, benchmarked mixture of gaussians vs. normalizing flow.

• Designed conditional normalizing flow architectures for i.i.d. Poisson models, achieved accurate inference of non-analytic posteriors in large-N regimes, proved gaussian-alignment improved stability over standard neural inference.

• Presented project outcomes at the University of Nottingham Undergraduate Research Poster Exhibition (Nov 2025), highlighting multi-distribution benchmarking and novel CLT-based regularization to an audience of faculty and peers.

• Developed GRU-KAN and LSTM-KAN architectures for early loan default detection (3–8 months ahead), achieved +12% AUC over GRU baseline, sustaining >92% accuracy at 3-month and >88% at 8-month horizons in OOT tests.

• Engineered borrower time-series features with out-of-time splits and "blank-interval" simulation to mimic real-world prediction scenarios using Freddie Mac Single-Family Loan-Level data, demonstrated utility for bank risk mitigation.

• Conducted extensive comparative experiments against LSTM, GRU, and Transformer-based models, rigorously validating consistent performance gains across varying feature windows, sample sizes, and early prediction intervals.

• Co-authored paper accepted at Applied Soft Computing (2025), also accepted for presentation at CSCR 2024, advancing interpretability and generalization via knowledge-injected sequential modeling for robust credit risk.

• Developed SPEM with a Sustainability Index (Z) and Economic Productivity Index (EPI) to quantify environmental–economic trade-offs in Juneau's tourism sectors spanning glaciers, whale watching, and rainforests etc.

• Estimated sector revenues, environmental costs, and social burdens using adaptive weights via AHP, CRITIC, and GRA; built dynamic programming models to forecast long-horizon outcomes of taxes, caps, and investment choices.

• Optimized dual objectives with NSGA-II across fifteen consumption paths, deriving differentiated elasticities—Glacier 1.39, Whale 1.98, Rainforest 2.57—to guide targeted tax schedules, visitor caps, and infrastructure investment.

• Extended the framework to Bali and Santa Barbara with income-tiered taxes; simulations projected five-year sustainability gains of +0.17 in Juneau and +0.15 in Bali despite initial visitor declines and revenue transients.

• Developed a Transformer–BiLSTM architecture for post-issuance credit risk profiling, integrating borrower repayment trajectories with macroeconomic signals via sliding-window and interval-masking features methods.

• Engineered hierarchical temporal encoding combining short-term behavioural shifts with long-range loan history, supported by an automated preprocessing pipeline for 20M+ loan records, reduced manual feature engineering time.

• Achieved up to 15% reduction in false negatives over LSTM/GRU baselines and 2% recall gain over Attention-BiLSTM in high-risk borrower detection, validated through multi-year out-of-sample tests and ablation studies.

• Built a forecasting pipeline using Prophet with exogenous price, holiday, and seasonality signals, delivering category-level demand predictions with R² up to 0.55 and informing cross-category substitution features from ARIMA.

• Optimized a seven-day dynamic pricing and replenishment plan via Simulated Annealing under shelf-life, loss-rate, and demand constraints, minimizing expected spoilage and stockouts in backtests while preserving service levels.

• Formulated SKU-level ordering and pricing as a Mixed-Integer Quadratic Program solved with a Genetic Algorithm, maximizing gross margin subject to demand coverage, storage limits, and budget caps across heterogeneous freshness.

• Defined forward data standards for storage capacity, spoilage rates, transit loss, and policy shocks, enabling robust scenario testing, reproducible pipelines, and cleaner ablation studies for governance audits and future model upgrades.

• Built a score-capture model combining Markov state transitions with pre-match odds to estimate point-level win probabilities, enabling granular, time-aligned momentum tracking and counterfactual evaluation across matches.

• Quantified momentum using leverage and win-probability features with EMA smoothing; validated stationarity via ADF tests; simulated match flows using MCMC; K-S tests rejected random-momentum null with p < 0.0001.

• Developed a multivariable LSTM using serve quality, faults, clutch points, athletic proxies, score differential, and shot placement, achieving 76.3% cross-validated accuracy on momentum swing classification and early-warning.

• Visualized the 2023 Wimbledon Gentlemen's Singles Final to flag turning points with leverage > 0.45; generalized methods to men's doubles and proposed transfer learning paths for table tennis and badminton analytics.

• Conducted an in-depth analysis of the catalytic reactions of Cotton Stalk (CS) pyrolysis products with those of Cellulose (CE) and Lignin (LG), leveraging statistical analysis to establish a functional equation among mixture ratios.

• Applied logistic regression and Grey Prediction Algorithms to accurately predict the optimal catalyst mixture ratio, enhancing the efficiency of pyrolysis processes, improving the recycling selection of catalyst mixture ratios.

• Demonstrated expertise in data-driven decision-making and algorithmic designs, proposed a new standard for catalytic reactions in pyrolysis research, contributed to the uprising field of sustainable energy and material recycling.

• Awarded outstanding prize for leading test loss performance in predicting catalytic reactions in desulfurization ash.

• Developed a neural network model focused on Orchard Monitoring, architecture a customized large scale Residual Neural Networks (RNN) to predict falling & ripeness of products, identifying and outlining the falling area of fruits.

• Practiced regularization techniques including drop-out layers, L2-regularization, stochastic gradient descents with learning rate schedular, ensemble methods over multiple neural networks, achieving best test loss & accuracy.

• Authored a comprehensive paper reviewing the challenges and current researches in the Orchard Monitor field.

• Awarded first prize for exceptional research and leadership in competition, demonstrating the capability of applying cutting-edge techniques in solving problems, contributed to evolving field of Intelligent Orchard Monitor Systems.

• Led analysis of 390 mother–infant pairs; engineered clean datasets and encodings, then ran normality, Spearman, ANOVA, and Chi-square tests to surface maternal indicators linked to infant behavior and sleep outcomes.

• Trained Random Forest classifiers with under- and over-sampling to correct class imbalance, raising minority-class accuracy by 28.3% and stabilizing fold variance through calibration, feature selection, and repeated cross-validation.

• Reduced multicollinearity with PCA and built multinomial logistic regression achieving 70.3% behavioral prediction accuracy; separately scored sleep quality via RSR plus logistic regression reaching 75.6% across four categories.

• Integrated predictive outputs into a cost optimizer using exhaustive search to recommend minimal-cost psychological treatment plans; delivered sensitivity analyses and deployment guidance for maternal-child health program decisions.

• Researched ML algorithms for UAVs modeling, optimizing real-time flight adjustments, achieving a higher stability and efficiency, employed PyTorch Neural Networks to predict a-priori unknown aerodynamic behaviors.

• Engineered sophisticated reinforcement learning algorithms, enabling UAVs to adapt to their environment, minimizing manual calibration while incorporating OOD Detection to recognize unknown territory signals.

• Summarized state-of-the-art machine learning algorithms for genetic simulation of a-prior unknown UAVs parameters.

• Designed, developed, and tested a rolling inventory system for Five Plus using IBM Planning Analytics, integrating master business data for automated calculations; improved product turnover and reduced inventory losses by 10%.

• Built an OTB ordering tag system leveraging multidimensional database analytics to optimize annual order quantities based on sales targets, strategy, and historical performance; replaced manual Excel workflows, cut calculation time.

• Programming: Python (PyTorch, NumPy, SciPy, scikit-learn, geoopt), R, MATLAB, LaTeX, Git

• Research Computing: Slurm job arrays and multi-GPU training on Ada (Nottingham) and CINECA Leonardo (EuroHPC), plus mixed-precision and numerical-stability debugging

• Methods: Hyperbolic and Riemannian deep learning, normalizing flows, diffusion and score-based models, simulation-based inference, Bayesian inference, time-series anomaly detection