Technology
MInference 1.0: Accelerating Pre-filling for Long-Context LLMs via Dynamic Sparse Attention Huiqiang Jiang, Yucheng Li
Existing methods for speeding up pre-filling often fail to maintain acceptable accuracy or efficiency when applied to long-context LLMs.
Deep Generative Models with Learnable Knowledge Constraints
Zhiting Hu, Zichao Yang, Russ R. Salakhutdinov, LIANHUI Qin, Xiaodan Liang, Haoye Dong, Eric P. Xing
Neural Information Processing Systems http://nips.cc/
How to Initialize your Network? Robust Initialization for WeightNorm & ResNets
Devansh Arpit, Vรญctor Campos, Yoshua Bengio
Neural Information Processing Systems http://nips.cc/
[R1/R2] Infinite width assumption: the infinite width assumption is needed due to the technical detail that the norm
We thank reviewers for their valuable comments. We respond to the main concerns below. Similar to that in Zhang et al. [31], we chose 10k block ResNet to stress the We will rephrase L243 to better express this. Derivative of weights depend on this term due to the chain rule. We will make this explicit in the revised manuscript.
Improving Diffusion-Based Image Synthesis with Context Prediction Ling Y ang
Diffusion models are a new class of generative models, and have dramatically promoted image generation with unprecedented quality and diversity.
On the Curved Geometry of Accelerated Optimization
Neural Information Processing Systems http://nips.cc/
Virtual Class Enhanced Discriminative Embedding Learning
Binghui Chen, Weihong Deng, Haifeng Shen
Neural Information Processing Systems http://nips.cc/
Predicting the Politics of an Image Using Webly Supervised Data
Christopher Thomas, Adriana Kovashka
Neural Information Processing Systems http://nips.cc/
e4dd5528f7596dcdf871aa55cfccc53c-AuthorFeedback.pdf
We show that predicting bias from images alone is more challenging than prediction from12 text. The features are then applied on visual-only data,23 e.g. We tried a variant of this approach of predicting latent text topics from images24 and obtained 0.681 on our full dataset (much lower than our method; compare to Tab. 1 in main). The model achieved 0.626 acc. Further balancing of loss hyperparameters could potentially improve the result.