Paper Notes Archive

Reading notes I wrote in 2022 while working through papers on sequence models, transformers, generative models, and holographic displays. Each note follows the same shape: summary, strengths, critique, and the resources I found useful.

These are kept online because people still link to them. Newer notes live in Knowledge Share.

  1. November 2022 Cascaded Diffusion for Super-resolution Image Generation Cascaded Diffusion Models for High Fidelity Image Generation
  2. November 2022 Review of the Guidance in Diffusion Models I plan to use Diffusion Models Beat GANs on Image Synthesis NIPS2021 paper as a foundation and review several diffusion guidance methods in this article…
  3. November 2022 VQ-VAE VQ-VAE is aiming to provide a discrete latent representation space rather than continuous VAE ones so that it could better model discrete modalities such as…
  4. November 2022 VAE: Variational Auto Encoder The motivation of VAE came from improving the representation (latent space, z) quality of the traditional autoencoder, so that we could take random samples…
  5. November 2022 StyleGAN StyleGAN: A Style-Based Generator Architecture for Generative Adversarial Networks
  6. November 2022 GAN: Generative Adversarial Nets Goodfellow, Ian and Pouget-Abadie, Jean and Mirza, Mehdi and Xu, Bing and Warde-Farley, David and Ozair, Sherjil and Courville, Aaron and Bengio, Yoshua
  7. November 2022 GLIDE: Text Guided Photorealistic Image Generation and Editing using Diffusion GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models
  8. November 2022 DDPM: Diffusion Model for Image Generation Come down to the implementation level, there are several modules to implement diffusion model: the noise scheduler, which sequentially adds up noise; the…
  9. November 2022 Hogel-free Holography Praneeth Chakravarthula, Ethan Tseng, Henry Fuchs, Felix Heide
  10. November 2022 DALL-E Ramesh, Aditya and Pavlov, Mikhail and Goh, Gabriel and Gray, and others PMLR 2021
  11. October 2022 Efficient Geometry-aware 3D Generative Adversarial Networks I review this paper mainly for curiosity of the term ‘geometry-aware’ and ‘3D’. We always see works from 2D face generation and the physical correctness the…
  12. October 2022 Holographic Glasses for Virtual Reality Jonghyun Kim, Manu Gopakumar, Suyeon Choi, Yifan Peng, Ward Lopes, Gordon Wetzstein
  13. October 2022 Reformer: The Efficient Transformer The original Transformer self-attention takes NND parameter space in Q K dot product operation. A sequence of 64k will thus result in 16GB of memory. However…
  14. October 2022 CLIP The intuition behind such success is that the task for model now is not blindly find the discrimination line for image data distribution and assign clusters…
  15. October 2022 ImageGPT Recall that for Transformer structure, the parameter space of the self-attention layer is $$O(dn^2)$$ where n is the sequence token number. Even for a 300300…
  16. October 2022 GPT-2 The revision of the Transformer decoder to fit for unsupervised training is quite straightforward. First, since there is no Transform encoder part in the…
  17. October 2022 BERT Language Model Fine-tunning demands the BERT to model the context information (both left-to-right and right-to-left) of each word to better capture the semantics. By that…
  18. October 2022 Self-attention with Relative Position Representations Absolute position encodes the relative position information based on the fact that for any fixed position offset k, $$sin((pos+k)/c)$$ can be acquired by the…
  19. October 2022 Attention Is All You Need For encoder stack, the intuition behind self-attention comes from the coreference phenomenon in language. When considering the semantics of a word in the…
  20. October 2022 Pixel Recurrent Neural Network The LSTM contains two components: state-to-state and input-to-state. For input-to-state component, the Row LSTM will take the whole input map and apply k1…
  21. September 2022 Sequence to Sequence Learning with Neural Networks Second it stacks multiple LSTM layers rather than one by having multiple hidden states for a single timestep. The increase of the parameter numbers help the…
  22. September 2022 Sequence to Sequence Learning with Neural Networks First is using separated LSTMs for encoder and decoder rather than a straight end to end design. This indirect idea makes it possible to train multiple…