Tag: 算法
All the articles with the tag "算法".
- Parallelizing Linear Transformers with the Delta Rule over Sequence LengthUpdated: at 16:46Published: at 14:43- DeltaNet 
- MLP Memory: Language Modeling with Retriever-pretrained External MemoryPublished: at 14:22- 用MLP学习并代替RAG中kNN输出的概率分布。 
- Swin Transformer: Hierarchical Vision Transformer using Shifted WindowsPublished: at 17:47- 看看Shift-Window Attention。 
- SpikeVideoFormer: An Efficient Spike-Driven Video Transformer with Hamming Attention and O(T) ComplexityPublished: at 16:56- 用汉明距离替换Attention中的点乘操作,避免出现Spike错开的情况。中间的做法比较有趣,但是实验感觉做的一般般,尤其是claim了自己有硬件实现的情况下energy计算还用的是纯算法的计算,并且FPGA的具体实现也没有透露,说了也没有说清楚。精度没有超过ANN2SNN的SOTA。重点还是需要用一些其他的操作替换掉对SNN不适应的算子。 
- SlowFast Networks for Video RecognitionUpdated: at 06:15Published: at 16:57- 多分支CNN,会不会有一些分支能学到更加相似的帧间变化? 
- DeltaCNN: End-to-End CNN Inference of Sparse Frame Differences in VideosUpdated: at 15:07Published: at 12:11- 利用CNN Layer的“线性”特征在帧之间做feature的差分,并且做了CUDA加速。和ViStream几乎一样的思路,能不能解决我们现在的问题? 
- Temporal Flexibility in Spiking Neural Networks: Towards Generalization Across Time Steps and Deployment FriendlinessPublished: at 15:38- ICLR2025 Poster,似乎也在做Elastic inference? 
- A Simple Framework for Contrastive Learning of Visual RepresentationsPublished: at 13:42- 对比学习SimCLR的论文。对比学习能对齐每一层的Feature吗? 
- QKFormer: Hierarchical Spiking Transformer using Q-K AttentionPublished: at 18:09- QKFormer,NIPS2024 Spotlight,把Direct Training SNN在ImageNet和CIFAR上的点刷的特别高,感觉之后要做就避不开它。 
- Transformers without NormalizationPublished: at 16:09- 何恺明新作,用DyT代替Norm,把同步操作变成了Element Wise的操作。新文章里面有用到,学习一下。