[ИАД, весна 2026] Введение в машинное обучение. Лекция 10: Эволюционные методы машинного обучения

Value of the Information & Strength of the Argument

The lecture provides substantial value by connecting historical developments (Ivakhnenko, Vapnik) to modern concepts like regularization and generalization bounds. The argumentation is solid, with clear logical progression from internal/external criteria to regularization as a proxy for external criteria, supported by theoretical justifications (VC dimension, AIC). The instructor effectively explains why internal criteria cannot be used for model selection and why external criteria are necessary. The discussion of multi-criteria optimization and feature selection is well-structured, with practical insights. However, some claims (e.g., ‘first deep neural network’) are presented without direct evidence, and the lecture assumes prior knowledge, making it less accessible to beginners.

Scientific Rigor, Source Quality, Title Accuracy

The lecture demonstrates high scientific rigor, referencing foundational works in statistical learning theory (Vapnik-Chervonenkis) and information criteria (Akaike). The instructor provides historical context, which adds depth. However, specific sources are not cited in the video or description, limiting verifiability. The title accurately reflects the content, focusing on evolutionary methods. The lecture is well-structured and technically accurate, though some topics (e.g., VC dimension) are covered briefly without full derivations.

The title accurately reflects the content: a lecture on evolutionary methods in machine learning, covering model selection, feature selection, and genetic programming.

8/10

The lecture is grounded in established machine learning theory, referencing foundational works (Vapnik-Chervonenkis, Akaike, Ivakhnenko) and presenting formal definitions and criteria. The instructor demonstrates deep expertise and provides historical context, though some claims (e.g., 'first deep neural network') are presented without direct citations in the video.

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

• 02:00Introduction to the lecture, overview of ML schools, and mention of Ivakhnenko's contribution.
• 10:00Explanation of internal vs. external criteria and the importance of external criteria for model selection.
• 20:00Discussion of regularization as a form of external criterion, linking to generalization bounds.
• 30:00Introduction to Vapnik-Chervonenkis theory and VC dimension.
• 40:00Multi-criteria optimization for model selection, inspired by Ivakhnenko.
• 50:00Feature selection as a discrete optimization problem, NP-hardness, and heuristic approaches.
• 1:00:00Comparison with logical methods and discussion of filter, wrapper, and embedded methods.
• 1:10:00Symbolic regression and genetic programming as advanced evolutionary methods.

• Pedro Domingos' classification of ML schools — Referenced in the introduction as the basis for the lecture structure.
• Ivakhnenko's Group Method of Data Handling (GMDH) — Discussed as a pioneering self-organizing model approach.
• Vapnik-Chervonenkis theory — Mentioned as foundational for statistical learning theory and VC dimension.
• Akaike Information Criterion (AIC) — Referenced as an example of L0 regularization with explicit penalty coefficient.

• Vapnik-Chervonenkis theory — Supports the discussion of VC dimension and generalization bounds.
• Akaike information criterion — Supports the mention of AIC as a model selection criterion.