I study how predictive information is represented inside neural systems, how it can be translated into faithful evidence, and how people understand and act on that evidence.
Research presentation · Neural representation analysisEach theme addresses a different layer of the same problem: how to build powerful models whose evidence, limitations, and human effects can be examined.
Methods that expose where concepts and predictive signals appear across neural representations, with emphasis on non-interventional analysis and human-readable evidence.
Empirical evaluation of explanation modalities, cognitive efficiency, trust calibration, understanding, and decision support.
Adaptive, physics-guided, and structured models for heterogeneous energy consumers and other real-world time series.
Metrics and evaluation frameworks that connect computational signals with human perception, consistency, and practical quality.
Linear Lens and related work investigate how predictive information emerges across a network while keeping the deployed model unchanged.
Explore Linear LensDefine the scientific and human question before choosing the architecture or explanation method.
Compare against transparent baselines, contemporary methods, and meaningful ablations.
Combine predictive metrics with robustness tests, statistical analysis, visual diagnostics, and user evidence when appropriate.
Package code, metadata, figures, and documented assumptions so results can be inspected and extended.
Extending concept-level and layer-wise analysis into explanations that support understanding without sacrificing faithfulness.
Studying differences across textual, graphical, and interactive explanations using trust, understanding, cognitive load, and decision outcomes.
Combining learnable structure, physical guidance, and robust missing-data methods for transferable forecasting.