Reliable AI,
from theory
to practice.

The challenge. Preserve fine structural detail while keeping a vision model compact enough for high-throughput inspection.

The approach. A teacher–student pipeline that uses boundary salience, structural complexity, and local feature variation to guide learning.

The result. A 7 percentage-point mIoU improvement to 80% in internal validation of the prototype.

About the cover. The cover illustrates image-to-mask segmentation; it is not a KLA inspection sample or model output.

The challenge. Build more reliable multi-label regression models from noisy, high-dimensional physical measurements.

The approach. Synthesize spectral data with conditional diffusion and evaluate both in-distribution and out-of-distribution behavior.

The context. Industry research at KLA. The implementation and internal evaluation data are not shared here.

About the cover. The signals illustrate conditional generation. They are not industrial measurements or model evaluation results.

The question. How can we understand and improve the stability of learned models for circuit design?

The approach. Analyze input and output graph-based manifolds to identify sensitive circuit elements under changes in graph structure and node features.

The outcome. First-author research at DAC 2025, selected as a Best Paper Nominee.

About the cover. The circuit is schematic. The DES bars replot Table I of CirSTAG: mean relative change in GNN-predicted arrival time, with 10% of nodes perturbed and pin capacitance scaled by 10. Stable: 0.0068%; sensitive: 0.5211%.

The question. How can we assess model and sample sensitivity without relying on access to a model’s internal gradients?

The approach. Use a spectral graph-based measure to study local input–output behavior and inform robustness evaluation and adversarial training.

About the cover. The digit tiles come from the original SPADE project figure. The connecting graph illustrates the method; it does not assign new numerical scores or show clean/adversarial image pairs.

The question. Which text samples are most sensitive to perturbations?

The approach. Compare input and output relationships using Distance Mapping Distortion, then use sample-level stability to guide robustness evaluation and training.

The status. Research preprint; first author.

About the cover. A conceptual view of input–output relationship distortion, not actual model responses or measured stability scores.

The question. Which graph nodes become unstable when connections or features change?

The approach. Build distance-preserving graph-based manifolds with spectral embeddings and probabilistic graphical models, then analyze distortion between model input and output.

The status. Research preprint; first author. Applications include GNN stability, recommendation models, and robustness research.

About the cover. A conceptual graph perturbation and mapping. Highlighted connections illustrate sensitivity; they are not a measured benchmark result.

The question. Can a model learn the relevant physics while spending less time on redundant training samples?

The approach. Graph-based importance sampling for PINNs, studied on computational fluid dynamics workloads.

My contribution. Co-author on the DAC 2024 publication.

About the cover. Sampling locations are illustrative. The pressure-error panel is adapted from Anticev, Aghdaei, Cheng & Feng, SGM-PINN, Figure 4; its field colors and color scale are preserved. This adapted cover is licensed CC BY-NC-SA 4.0.

The work. Developed ML methods for semiconductor metrology and advanced packaging at KLA.

The presentation. Machine-Learning Solution for Cu Dishing Metrology in Advanced Packaging, SPIE Advanced Lithography + Patterning 2026.

The context. Industry research; proprietary implementation and evaluation data are not shared here.

About the cover. A schematic cross-section and measurement workflow. Geometry is conceptual, with no measured dimensions or proprietary sample data.

The work. As a scientific advisor and AI-lab lead, I worked on 3D avatar assistants combining visual, speech, emotion, and language components.

The demonstration. Co-developed RITA, a real-time interactive talking-avatar framework demonstrated at CVPR 2024.

About the cover. The cover uses actual output frames at 0.0, 1.0, and 3.0 seconds from the project demo. These are not input portraits.

The challenge. Improve classification while reducing inference cost in an industrial inspection pipeline.

The approach. Curate the training data, prune the network, and strengthen training against perturbations.

The result. A 42% relative accuracy improvement over the prior model and a 5.8× inference speedup in internal evaluation.

About the cover. The negative and positive examples come from the original LEMA project figure. Boxes and arrows are source annotations; these are not before-and-after model outputs.

The challenge. Detect fine hair-like contaminants against complex material textures.

The approach. Robust training and augmentation, evaluated on a small internal dataset and severity-5 image corruptions.

The context. Scientific advisory work on industrial computer vision at LEMA.

About the cover. The overview and magnified detail use the same original inspection image. The red box is a source annotation, not a new prediction.

The question. How should models be compared when predictions arrive at different times and evidence is incomplete?

The approach. Split data by label-resolution date, combine specialist models with contextual log-opinion pooling, calibrate outputs, and compare matched forecasts with Bradley–Terry/Elo rankings.

The scope. Independent research using daily equity data, reproducible artifacts, and offline evaluation.

About the cover. A conceptual forecasting workflow. The drawn observations and forecast are illustrative, not measured predictions, performance results, or investment returns.