Summary
- Published 2025-06-15 · slug
boostcourse-math-ocr · category "AI and machine learning studies / Computer vision and OCR" · source article recorded in Korean and English. - A retrospective on taking NAVER Connect Foundation's advanced Boostcourse course, "Deep Learning Models That Understand Mathematical Formulas (OCR)." Reading formula images was the first step for the personal project "Math God," whose goal is for AI to read and solve math problems.
- The article first explains why mathematical OCR differs from ordinary text OCR: two-dimensional spatial structure, symbols whose meanings depend on their positions, and structured output such as LaTeX.
- I understood the course's standard architecture as "eyes" (a CNN encoder), a "brain" (a Transformer decoder), and the "connection" between them (attention).
- Next steps for the project were generating synthetic LaTeX data, designing a domain-specific tokenizer, and building a parser from LaTeX to symbolic representations and expression trees.
Mentioned
- NAVER: NAVER Connect Foundation's Boostcourse, the course provider.
Connections
- Writing
- Concepts: LaTeX and tectonic · OCR · Transformer · i18n and multilingual support · Running a technical blog
Local graph
Notes 1 · Links 0