[ИАД, осень 2026] Методы глубокого обучения. Занятие 9: Autoregression, VAE, GAN

Value of the Information & Strength of the Argument

The lecture provides a solid, well-structured introduction to generative models, balancing mathematical rigor with intuitive explanations. The speaker clearly distinguishes between discriminative and generative approaches, and walks through the chain rule for autoregressive models, the reconstruction loss for VAEs, and the adversarial objective for GANs. The argumentation is coherent, building from basic definitions to concrete examples and code-free illustrations. The inclusion of a historical evolution and current state-of-the-art (Flux, flow matching) adds practical relevance. The main strength is the lecturer’s expertise, which lends credibility to claims about model quality and practical challenges. However, some sections rely on assertions about recent advances without detailed citations, and the comparison between model types could have been more quantitatively supported.

Scientific Rigor, Source Quality, Title Accuracy

The lecture is scientifically rigorous in its presentation of foundational concepts, using correct mathematical formulations (e.g., probability chain rule, negative log-likelihood). The speaker does not explicitly cite papers or provide external references during the lecture, but the content aligns with established literature (e.g., PixelRNN, VAE, GAN). The title precisely matches the content, focusing on autoregression, VAE, and GAN, with no notable discrepancies. The absence of formal citations is a minor weakness, as interested viewers cannot directly access the referenced works. No user comments were provided, so no public reception analysis is possible.

The title accurately reflects the content: the lecture focuses on autoregressive models, VAEs, and GANs, as stated.

8/10

The lecture is delivered by an expert researcher in generative models, directly involved in state-of-the-art model development at Sber's Kandinsky lab. The technical content is accurate and well-structured, covering key classes of generative models with mathematical formalism. However, the lack of formal citations or references in the lecture limits absolute verification.

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

• 00:00Introduction and wait for participants; technical checks.
• 05:00Start of lecture; speaker introduces self and topic: generative models.
• 10:00Quiz: identifying real vs. generated faces; all shown faces are AI-generated.
• 25:00Historical evolution of generative models: from autoencoders to GANs and diffusion.
• 40:00Definition of generative models vs. discriminative models; three main goals.
• 1:15:00Taxonomy of generative models: explicit vs. implicit density; start of autoregressive models.
• 1:40:00Detailed explanation of autoregressive models: chain rule, PixelRNN/PixelCNN, training with BCE loss.
• 2:10:00Discussion of VAE: encoder-decoder structure, latent space, reconstruction loss.
• 2:40:00Introduction to GANs: generator vs. discriminator, training dynamics, challenges.
• 3:15:00Modern architectures: flow matching, Flux, VL3, and reflection on current trends.

• Pixel Recurrent Neural Networks — Mentioned in relation to autoregressive generative models (PixelRNN).
• Conditional Image Generation with PixelCNN Decoders — Mentioned as an autoregressive model variant (PixelCNN).
• Auto-Encoding Variational Bayes — Foundational paper for variational autoencoders (VAEs), referenced during VAE discussion.
• Generative Adversarial Nets — Original GAN paper, mentioned in the context of GAN models.
• Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks — Referenced as the DCGAN architecture, a milestone in GAN development.