Hyacehila's Blog

Are deep learning methods always effective for time-series forecasting? Based on Christoph Bergmeir’s NeurIPS 2024 talk, this post discusses model limits and evaluation traps.

PCR and PLSR are common dimensionality-reduction regression methods under multicollinearity. This post derives their logic and explains how PLSR brings in response relevance.

When data looks too perfect, it may have drifted away from reality. Starting from Benford’s Law, this post reveals statistical fingerprints hidden in fabricated data.

Every fall, Matt Might explains what a Ph.D. means to new doctoral students. This post uses that classic illustration to describe knowledge accumulation through doctoral research.

A summary of empirical scaling laws from Kaplan to Chinchilla, with a brief discussion of tree models and scaling laws.

Hypothesis testing is powerful but assumption-bound; statistical significance does not make a conclusion true, and anyone without statistical literacy can accidentally commit statistical malpractice.

Why are boosted trees still such strong baselines for structured tabular data? This post compares XGBoost, LightGBM, and CatBoost by technical focus and best-fit use cases.

A technical overview of the path from language-only models to vision-language models, covering CLIP, VLMs, diffusion guidance, and native multimodal tokenization.

A deep dive from linear-preserving methods such as PCA and MDS to nonlinear manifold methods such as t-SNE and UMAP, with their mathematical intuition and technical tradeoffs.

Are calculations precise while charts are rough? Francis Anscombe’s quartet challenges that belief and shows why exploratory data analysis matters in statistical inference.