---

Random Signal Analysis ECE629

Spring 2023
Professor Sanghoon Sull
School of Electrical Engineering, Korea University

Reading Assignment

1. Transformer: Attention Is All You Need, NIPS 2017
https://jalammar.github.io/illustrated-transformer/

2. GPT3: Language Models are Few-Shot Learners, NeurIPS 2020
https://jalammar.github.io/illustrated-gpt2/

Week2 (03.15): ChatGPT (Overview of transformer/GPT)

https://openai.com/blog/chatgpt

https://platform.openai.com/docs/model-index-for-researchers

Papers for ChatGPT

L01-1: Training language models to follow instructions with human feedback, NeurIPS 2022 (InstructGPT)

L01-2: Learning to summarize from human feedback, NeurIPS 2020

L01-3: Asynchronous Methods for Deep Reinforcement Learning, ICML 2016

L01-4: Deep Reinforcement Learning from Human Preferences, NIPS 2017

L01-5: Fine-Tuning Language Models from Human Preferences, arXiv 2019

L01-6 (PPO): Proximal Policy Optimization Algorithms, arXiv 2017

References

R. Sutton and A. Barto, Reinforcement Learning: An Introduction, MIT Press, Cambridge, 2018

Trust Region Policy Optimization, ICML 2015

Week3 (03.22): ChatGPT (Policy optimization, TRPO/PPO)

Week4 (03.29): ChatGPT (DRL from HP)

Week5 (04.05): ChatGPT (FT LM from HP)

Week6 (04.12): ChatGPT (InstructGPT)

Week7 (04.19): AlphaCode

L02-1: Competition-level code generation with AlphaCode, Science 2022
L02-2: Supplementary for Competition-level code generation with AlphaCode, Science 2022
L02-3: TEXT GENERATION BY LEARNING FROM DEMONSTRATIONS, ICLR 2021 (GOLD)

Week8 (04.26): Midterm (ChatGPT)

Week9 (05.03): DDGM (L03-1: Tutorial Part 1, L03-2: DDPM)

Papers for DDGM

L03-1: Denoising Diffusion-based Generative Modeling (DDGM): Foundations and Applications, CVPR 2022 Tutorial
L03-2: Denoising-diffusion-probabilistic-models (DDPM), NeurIPS 2020
L03-3: Deep Unsupervised Learning using Nonequilibrium Thermodynamics, ICML 2015

Week10 (05.10): DDGM (L03-2: DDPM)

Useful link:
What are diffusion models? https://lilianweng.github.io/posts/2021-07-11-diffusion-models/

References
Prince, S.J.D., Computer Vision: Models, Learning and Inference, 2012 (5.6 Product of two normals)

Week11 (05.17): DDGM (Score-based: L03-4)

L03-4: Generative Modeling by Estimating Gradients of the Data Distribution, https://yang-song.net/blog/2021/score/
L03-4 ref-1: Sliced Score Matching: A Scalable Approach to Density and Score Estimation, UAI 2019 (Section 2.1)
L03-4 ref-2: Estimation of Non-Normalized Statistical Models by Score Matching, JMLR 2005 (Theorem 1)

Week12 (05.24): DDGM (Score-based: L03-5 1/2)

L03-5: SCORE-BASED GENERATIVE MODELING THROUGH STOCHASTIC DIFFERENTIAL EQUATIONS, ICLR 2021
L03-5 ref-1: A Connection Between ScoreMatching and Denoising Autoencoders, 2011 ((4.3), (4.4))
L03-5 ref-2: Generative Modeling by Estimating Gradients of the Data Distribution, NeurIPS 2019 ((1), (2), (5), (6))

References
Simo Sarkka and Arno Solin. Applied stochastic differential equations, volume 10. Cambridge University Press, 2019 (Def. 3.9 (White noise), Def. 4.1 (Brownian motion), Theorem 5.4 (Fokker-Planck-Komogorov equation), Equations (5.50) and (5.51))

Week13 (05.31): DDGM (Score-based: L03-5 2/2)

Week14 (06.07): DDGM (Accelerated Sampling)

L03-6: DENOISING DIFFUSION IMPLICIT MODELS, ICLR 2021 (DDIM)

Note: Download the lecture note from the course BBS.

Week15 (06.14): Final (DDGM)

Week16 (06.21): DDGM (L03-1: Tutorial Part 2 for Conditional generation and cascaded generation for Imagen and DALL·E 2)

L03-7-0.9: Diffusion models beat GANs on image synthesis, NeurIPS 2021 (Section 4.1-4.2)
L03-7-1:  CLASSIFIER-FREE DIFFUSION GUIDANCE,  NeurIPS Workshop 2021 (Section 3)
L03-7-2: Cascaded Diffusion Models for High Fidelity Image Generation, JMLR 2022 (Section 2-3)