From steepest descent's Euclidean blindspot, through the Fisher Information Matrix and Riemannian manifolds, to a practical curvature-aware optimizer you can actually run on your laptop. A geometric dissection of why Adam's diagonal approximation lies — and what natural gradient does instead.

Why your neural network wants to memorize noise — and how circles, diamonds, and Bayesian priors force it to generalize instead. A geometric and probabilistic view of regularization from first principles.

From the ruins of sequential recurrence to the elegance of parallelized attention — a step-by-step autopsy of "Attention Is All You Need," covering QKV projections, multi-head parallelism, scaling laws, BERT vs GPT, FlashAttention, RoPE, and the world model hypothesis.

Continuous operator learning, spatial bottlenecks, and predicting shock fronts in fluid dynamics. A 4-dimensional branch network combined with an adaptive trunk learns to approximate solutions of the Burgers PDE across 819K spatiotemporal points.

SNNs are highly energy-efficient, but scaling them remains a challenge. Biologically inspired constraints, mathematical continuous relaxation traps in CVXPY, and Iterative Combined Importance Scoring to conquer the 50% sparsity wall.

Combining text and vision to improve sentiment analysis models — integrating BERT and ViT with CLIP-based alignment, and why Dempster-Shafer evidential fusion outperforms standard concatenation by explicitly modelling uncertainty across modalities.