This post introduces a new learning paradigm, Nested Learning (NL), which represents a model as a coherent set of nested, multi-level, and/or parallel optimization problems, each with its own context flow.

This post introduces basic concepts of tokenization, decoding, prompting, tool-augmented agents, RAG, RLHF, VAEs, diffusion models, and LoRA—presented with standard objective functions and probabilistic notation.

This post tidies up the core ideas for how to think about, estimate, measure, and ship performance improvements—mainly for single-binary software (not distributed systems or ML hardware tuning).

This post provides a novel, strongly consistent method for accurately and efficiently estimating the cardinality of complex graph queries.

This post provides novel evaluation strategies for dynamic link prediction, introducing new datasets, robust negative sampling methods, and the EdgeBank baseline for better model assessment.

This post provides a comprehensive overview of RAG and Fine-Tuning methods and introduces Retrieval Augmented Fine-Tuning (RAFT) as a unified approach to enhance language model training.

This post provides a comprehensive overview of RL and graph mining methods and generalize these methods to Graph Reinforcement Learning (GRL) as a unified formulation.

This post introduces Primal Wasserstein Imitation Learning (PWIL) method, based on Wasserstein distance, derives a reward function offline and efficiently replicates expert behavior in MuJoCo continuous control tasks.