[ИАД, весна 2026] Введение в машинное обучение. Лекция 5: Обучаемая векторизация данных

Value of the Information & Strength of the Argument

The lecture provides substantial value by connecting classical dimensionality reduction techniques (PCA, SVD) with modern deep learning approaches (autoencoders, transformers). The argumentation is solid, with clear mathematical derivations and intuitive explanations. The instructor effectively builds on previous lectures, creating a coherent narrative. He also engages the audience with questions, encouraging active thinking. The use of concrete examples, such as credit scoring and recommender systems, helps illustrate abstract concepts. The discussion of limitations and extensions (e.g., non-negative matrix factorization) adds depth. Overall, the lecture is informative and well-argued, though it assumes prior knowledge of linear algebra and basic ML concepts.

Scientific Rigor, Source Quality, Title Accuracy

The lecture demonstrates scientific rigor through precise mathematical formulations and references to established methods (PCA, SVD, autoencoders). However, it lacks explicit citations to external sources, relying instead on the instructor’s expertise. The title accurately reflects the content, focusing on learnable vectorization. The lecture is part of a structured course, suggesting a systematic approach. The instructor mentions a reference to Chechotka (likely a misspelling of Cichocki) for non-negative matrix factorization, but no specific URLs are provided. The content is consistent with standard ML literature, but without external verification, the overall reliability is moderate.

The title accurately reflects the content: the lecture focuses on learnable data vectorization, covering PCA, autoencoders, and matrix factorization, culminating in transformers and large language models.

8/10

The lecture is a well-structured academic presentation, covering classical methods (PCA, SVD) and modern extensions (autoencoders, matrix factorization) with mathematical rigor. The instructor demonstrates deep knowledge and provides clear derivations. However, it is a single lecture without external citations or peer-reviewed references, and the content is not independently verified.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 00:00Introduction: recap of three key ideas in ML, including learnable data vectorization.
• 03:00Motivation for dimensionality reduction and PCA as a classical method.
• 10:00Derivation of PCA via low-rank matrix factorization and SVD.
• 20:00Discussion of choosing the number of principal components using the elbow criterion.
• 30:00Connection between PCA and linear autoencoders.
• 40:00General matrix factorization for missing data and non-negative constraints.
• 50:00Recommender systems as an application of matrix factorization.
• 1:00:00Stochastic gradient descent for latent factor models.
• 1:10:00Regularization and non-negative matrix factorization.
• 1:20:00Transition to transformers and large language models.

• Chechotka (likely Cichocki) on non-negative matrix factorization — Mentioned in the context of non-negative matrix factorization literature.

• PCA and SVD are standard methods in machine learning — The lecture's treatment of PCA aligns with standard textbook material.
• Autoencoders as generalization of PCA — The lecture's claim that linear autoencoders generalize PCA is supported by literature.