Research notes and engineering essays
Ideas on AI, mathematics,
and software systems.
Practical write-ups, experiments, and technical notes.
No results found on this page.
Positive–Unlabeled Learning: A Short Survey
A phone number with several confirmed spam reports is a useful positive example. A number with no reports is harder to interpret: it might be benign, or it might be spam that nobody has reported yet. Treating every unreported number as negative turns missing information into a potentially incorrect…
Continue readingKiến trúc nền tảng MLOps nội bộ với ClearML và Label Studio
Một nền tảng AI nội bộ cần quản lý nhiều hơn các tác vụ huấn luyện. Dữ liệu thô phải được gán nhãn và quản lý phiên bản; mỗi thử nghiệm phải gắn với đúng mã nguồn, môi trường chạy và bộ dữ liệu; mô hình sau huấn luyện phải có thể truy ngược nguồn gốc. Khi hệ thống bị cô lập Internet, toàn bộ contain…
Continue readingDAVE, MonkeyOCRv2, and DocPO: A Survey of Document VLM Training
Executive summary DAVE, MonkeyOCRv2, and DocPO address different stages of a document vision language model VLM lifecycle rather than competing as three interchangeable training recipes. Method Primary stage Main contribution Best use case DAVE Vision encoder pretraining and multi decoder alignment…
Continue readingUncertainty in Deep Learning: Deep Ensembles and MC Dropout
Why uncertainty matters A neural network can be confidently wrong. A softmax score of 0.99 only says that one logit is much larger than the others; it does not prove that the input resembles the training data or that the prediction is reliable. This distinction matters whenever a prediction drives a…
Continue readingNavier–Stokes Equations: From Conservation Laws to Fluid Flow
Why these equations matter The Navier–Stokes equations describe how the velocity, pressure, density, and temperature of a fluid evolve. They sit behind weather prediction, aircraft design, blood flow simulation, ocean circulation, combustion, and the movement of water through a pipe. Their ingredien…
Continue readingData Studio Architecture and Data Storage Specification
Production AI data platforms need more than a file upload endpoint. This specification defines a complete, reproducible path from Hugging Face compatible repositories to immutable revisions, transactional metadata, content addressed storage, columnar indexes, and revision scoped serving. Live demo T…
Continue readingQ-Learning from Scratch: Solving a Grid World with Python
Introduction Most machine learning models learn from labeled examples. Reinforcement learning is different: an agent interacts with an environment , observes the consequences of its actions, and learns which decisions produce the largest long term reward. In this tutorial, we will implement Q learni…
Continue readingBayesian Inference: Priors, Likelihoods, and Decisions
Why Bayesian inference matters Deterministic pipelines often fall apart when the data distribution shifts or the amount of evidence changes. Bayesian inference keeps a full probability distribution over uncertain quantities, so you can update beliefs as new observations arrive and keep downstream de…
Continue readingBuild a ChatGPT-like chatbot for free with Ollama and Open WebUI
Introduction Has the power of ChatGPT led you to explore large language models LLMs and want to build a ChatGPT like chatbot app? Do you want to create a chatbot with your own personal touch? Do you want to deploy a chatbot tool for your team at work to support daily tasks? This post shows you how t…
Continue readingLinear Regression: Foundations, Estimation, and Diagnostics
Linear Regression: Foundations, Estimation, and Diagnostics Linear regression is a fundamental method for modeling the relationship between a continuous response variable and one or more explanatory variables. It is widely used for prediction, estimation, hypothesis testing, and the analysis of rela…
Continue readingHàm sigmoid dưới góc nhìn xác suất
Trong bài toán phân loại nhị phân, mô hình thường tạo ra một điểm số thực $z in mathbb R $. Tuy nhiên, một số thực bất kỳ chưa thể được diễn giải trực tiếp như xác suất. Hàm sigmoid giải quyết vấn đề này bằng cách ánh xạ $z$ vào khoảng $ 0, 1 $: $$ sigma z = frac 1 1 + e^ z . $$ Nếu đặt $$ P y=1 mid…
Continue readingGiới thiệu về Variational Autoencoder
Introduction Xin chào mọi người, trong bài viết ngày hôm này mình sẽ cùng mọi người tìm hiểu về Variational Autoencoder VAE , một loại generative model trong deep learning. Trong vài năm gần đây, các mô hình generative đang thu hút được sự chú ý của các nhà nghiên cứu và đạt được một số kết quả đáng…
Continue reading