[ИАД, осень 2025] Методы глубокого обучения. Занятие 10: Autoregression, VAE, GAN

Value of the Information & Strength of the Argument

The lecture provides a solid conceptual foundation for generative models, clearly explaining the differences between model families and their underlying principles. The argumentation is coherent, building from basic probability to specific architectures. The instructor uses intuitive examples and connects to real-world applications, enhancing understanding. However, the depth of mathematical derivations is moderate, suitable for an introductory graduate course, and some advanced topics are only briefly touched upon.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is high: the content aligns with established literature (e.g., Kingma & Welling, Goodfellow et al.). The lecture does not cite specific papers explicitly, but the concepts are standard. The title accurately reflects the content. No comments were provided, so no public trends are analyzed.

The title accurately reflects the content: a lecture on deep learning methods focusing on autoregressive models, VAEs, and GANs, with practical seminars.

8/10

The lecture is given by a practitioner from the Kandinsky team (Sber AI), covering established generative model families (autoregressive, VAE, GAN) with mathematical formulations and practical seminar coding. The content aligns with standard ML curricula and known literature, though it is a single lecturer's perspective without external citations in the video itself.

• 00:00Introduction
• 08:30What are Generative Models?
• 17:11Autoregressive Models
• 32:33Autoencoders
• 45:37VAEs
• 1:30:20GANs
• 2:07:29Coffee break
• 2:17:44Seminar: PixelRNN on binarized MNIST
• 2:42:20Seminar: VAE on MNIST
• 3:11:54Seminar: GAN on MNIST
• 3:33:30Extra: Modern Architectures

• Pixel Recurrent Neural Networks — Mentioned as the origin of PixelRNN and PixelCNN models.
• Auto-Encoding Variational Bayes — Foundational paper for VAEs.
• Generative Adversarial Networks — Original GAN paper.
• Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks — DCGAN paper, mentioned as first convolutional GAN.
• Denoising Diffusion Probabilistic Models — DDPM paper, mentioned as first diffusion model.

• Deep Learning (Goodfellow et al.) — Standard textbook covering generative models.
• Pattern Recognition and Machine Learning (Bishop) — Classic reference for probabilistic models.