8PM - 9PM, Every Thursday

ReferencesATLAS, arXiv 2025TL;DR - Test time context memorization to improve long-term context understandingIssues to solveLimited memory capacity of Transformers, and its fixed-size memory $\rightarrow$ Section 3.1Optimizing the memory only with respect to the last input during online training $\rightarrow$ Section 3.2 ATLAS introduces a novel memory-augmented Transformer framework that learns ..

ReferencesCoZ, arXiv 2025GithubTerminologiesSingle-image super-resolution (SISR): A task to reconstruct a high-resolution (HR) image from a low-resolution (LR) image. It can be represented as: $p(x_H|x_L)$Autoregressive model: A model that predicts the next value from the previous predictions.Vision-language model (VLM): A model results the output from the image input and the text input.ill-pose..

ReferenceLoRA, ICLR 2022GithubTL;DR - Replace 'dense layer' with 'rank decomposition matrices'Full fine-tuning of large language models (LLMs) for the specific downstream tasks is not feasible (e.g., fine-tune GPT-3 175B for document summarization with short due or limited GPUs)"Compared to GPT-3 175B fine-tuned with Adam, LoRA can reduce the number of trainable parameters by a factor of 10,000 ..

2024EfficientAD: Accurate Visual Anomaly Detection at Millisecond-Level Latencies, WACV 2024 [slide][paper] Generative Unlearning for Any Identity, CVPR 2024 [slide][paper] CVPR 2024 virtual attendance [slide] AMI-Net: Adaptive Mask Inpainting Network for Industrial Anomaly Detection and Localization, T-ASE 2024 [slide][paper]Active token mixer, AAAI 2023 [slide][paper]Continuous Memory Repre..

ReferencesControlNet ICCV2023GithubControlNet at HuggingFaceTerminologiesZero-Initialized Layers: A layer initialized with zero valuesText-to-Image: A task of image generation by input textHyperNetwork: A network to generate the weights for another network or parts of the neural networkAdapter: A module to adapt a pretrained model to other tasks (e.g., CLIP ICML2021, T2I-Adapter AAAI2024)Low-Ran..

ReferencesViT-CoMer CVPR2025 HighlightGithubTerminologiesDense prediction: Object-oriented tasks, including object detection, semantic segmentation, etc.Depth-wise convolution: Each channel of the convolutional filters corresponds to the input dimensionTL;DR - Parallelize the ViT and CNN featuresProposed method: A plain, pretraining-free, and feature-enhanced ViT backbone named ViT-CoMer(a) Visi..