Tag: attention
In neural networks, "attention" is a mechanism that allows the model to focus on specific parts of the input data when making predictions or decisions. It's particularly important in sequence-to-sequence tasks, like machine translation or natural language processing. Attention mechanisms enable the model to assign different levels of importance to different elements in the input sequence, allowing it to consider and weigh information selectively. This selective focus enhances the model's ability to capture relevant context and dependencies, leading to more accurate and context-aware predictions. The concept is inspired by how humans pay attention to different aspects of information when processing data.- mPLUG-Owl2 Revolutionizing Multi-modal Large Language Model with Modality Collaboration (11 Nov 2023)
This is my reading note for mPLUG-Owl2: Revolutionizing Multi-modal Large Language Model with Modality Collaboration. This paper proposes a method to unify visual and text data for multi modal model. To this end, it uses QFormer to extract visual information and concatenate to text and feed to LLM. However, it separates the projection layer and layer norm for visual and text. This paper is similar to COGVLM.
- CogVLM Visual Expert for Pretrained Language Models (10 Nov 2023)
This is my reading note for CogVLM: Visual Expert for Pretrained Language Models. This paper proposes a vision language model similarly to mPLUG-OWL2. To avoid impacting the performance of LLM, it proposes a visual adapter which adds visual specific projection layer to each attention and feed forward layer.
- Tell Your Model Where to Attend Post-hoc Attention Steering for LLMs (08 Nov 2023)
This is my reading note for Tell Your Model Where to Attend: Post-hoc Attention Steering for LLMs. This paper proposes to improve LLM instruction follow performance by changes the attention weight to emphasize contents highlighted by user. The attention head to model is found by profiling the model on a small scale set of data.
- GIT A Generative Image-to-text Transformer for Vision and Language (16 Oct 2023)
This is my reading note for GIT: A Generative Image-to-text Transformer for Vision and Language. This paper proposes a image-text pre-training model. The model contains visual encoder and text decoder; the text decoder is based on self-attention, which takes concatenated text tokens and visual tokens as input.
- Efficient Streaming Language Models with Attention Sinks (05 Oct 2023)
This is my reading note for Efficient Streaming Language Models with Attention Sinks. This paper proposes a method to extend a LLM to infinite length text. This method is based on sliding attention plus prepending four sink tokens to aggregate global information. This paper shares similar idea as Vision Transformers Need Registers, which adds addition token to capture global information in attention.
- DeepSpeed-VisualChat Multi-Round Multi-Image Interleave Chat via Multi-Modal Causal Attention (01 Oct 2023)
This is my reading note for DeepSpeed-VisualChat: Multi-Round Multi-Image Interleave Chat via Multi-Modal Causal Attention. This paper proposes a method for multi round multi-image multi modality model. The paper utilizes a frozen LLM and visual encoder. The contribution of the paper includes: 1. Casual cross attention method to combine image and multiround text; 2. A new dataset.
- Vision Transformers Need Registers (29 Sep 2023)
This is my reading note for Vision Transformers Need Registers. This paper analyzes the attention map of transformer and find too large scale transformer and trained after a long iteration, some token show exceptionally high norm. Those tokens usually correspond to patches in uniform background. Analysis indicates that those tokens are used to store global information. Thus at would heart dense prediction tasks like image segmentation. To tackle this, the paper proposes add additional tokens during trains and inference, but rejecting for outputs.
- LongLoRA Efficient Fine-tuning of Long-Context Large Language Models (27 Sep 2023)
This is my reading note on LongLoRA: Efficient Fine-tuning of Long-Context Large Language Models. The paper proposes a method to fine tune a pretrained LLM to handle long context. To this end, it divide the data into different groups and performed attention within group; for half of heads, it shift the groups by half to enable attention across the groups.
- Coarse-to-Fine Vision-Language Pre-training with Fusion in the Backbone (22 Sep 2023)
This is my reading note for Coarse-to-Fine Vision-Language Pre-training with Fusion in the Backbone. This papers propose a two-stage pre-training strategy: (i) coarse-grained pre-training based on image-text data; followed by (ii) fine-grained pre-training based on image-text-box data.
- MagiCapture High-Resolution Multi-Concept Portrait Customization (11 Sep 2023)
This is my reading note on MagiCapture High-Resolution Multi-Concept Portrait Customization. This paper proposes a diffusion method to apply a style to a specific face image. Both the style and face are given as images. To do this, this paper fine tune existing model with LORA given several new loss functions: one is face identity loss for the face region given a face recognition model; another one is background similarity for the style. The two loss are applied to the latent vector.
- Key-Locked Rank One Editing for Text-to-Image Personalization (07 Sep 2023)
This is my reading note on Key-Locked Rank One Editing for Text-to-Image Personalization. This paper proposes a personalized image generation method base on controlling attention module of the diffusion model. Especially key captures the layout of concept and value captures the identity of the new concept. A rank one update is applied to the attention weight to this purpose.
- Multimodal Learning with Transformers A Survey (02 Sep 2023)
This is my reading note on Multimodal Learning with Transformers A Survey. This a paper provides a very nice overview of the transformer based multimodality learning techniques.
- DualToken-ViT Position-aware Efficient Vision Transformer with Dual Token Fusion (03 Aug 2023)
This is my reading note for DualToken-ViT Position-aware Efficient Vision Transformer with Dual Token Fusion. The paper discuss efficient transformer, which is based on combining convolution with attention: where convolution extracts local information and then fused with global information via attention.
- Jointly Training Large Autoregressive Multimodal Models (28 Jul 2023)
This is my reading note for Jointly Training Large Autoregressive Multimodal Models. This paper proposes a multimodality model for generating images. The paper is not just dilution based method but instead auto regressive method.it argues to initialize the model from the weight of frozen models.
- Multi-head or Single-head? An Empirical Comparison for Transformer Training (30 Jun 2023)
This is my reading note for Multi-head or Single-head? An Empirical Comparison for Transformer Training. This paper shows that multi head attention is the same as deeper single head attention, but the later is more direct to train and need special care to initialize.
- What Does BERT Look At? An Analysis of BERT's Attention (28 Jun 2023)
This is my reading note for What Does BERT Look At? An Analysis of BERT’s Attention. This paper studies the attention map of Bert.it found that the attention map captures information such as syntax and co reference . It also found there is a lot of redundancy in heads of the same layer.
- ViT AN IMAGE IS WORTH 16X16 WORDS TRANSFORMERS FOR IMAGE RECOGNITION AT SCALE (28 Mar 2021)
Vision Transformer (ViT) is a pure transformer architecture (no CNN is required) applied directly to a sequence of image patches for classification tasks. The order of patches in sequence capture the spatial information of those patches, similar to words in sentences.
- Self Attention (01 Aug 2019)