Skip to main content

Paper blog

Entry-level AI/ML developers

Agent Planning with World Knowledge Model

Agent Planning with World Knowledge Model

Basic Information

  • 2024/05/13 發布 (尚未正式於 Conf. 發表)
  • Shuofei Qiao, Runnan Fang, Ningyu Zhang et al. @ Zhejiang University, National University of Singapore, Alibaba Group

問題描述

近年來大型語言模型(LLM)在許多自然語言處理的問題有很快速的成長,而近期開始出現一些使用 LLM 作為 agent model 來處理物理環境中的規劃問題。然而由於當前 SOTA 的 LLM 幾乎都是 autoregressive model,模型實際上會做的事情是去預測下一個 output token 要是什麼,實際上他們對於物理環境是沒有任何理解的。


About 20 minNotePaper ReadReinforcement LearningLLM
Attention is all you need

Attention is all you need

Basic Information

  • NIPS 2017 (former NeuralPS)
  • Ashish Vaswani, Noam Shazeer, Niki Parmar et al. from Google Brain and Google Research

問題描述

RNN

近年來自然語言處理(Natural Language Processing, NLP)與機器翻譯等任務上時常使用 Recurrent Neural Network(RNN), Long Short-Term Memory(LSTM), Gated Recurrent Neural Network 等模型架構,我們也看到使用 Recurrent 模型以及 Encoder-Decoder 架構蔚為流行。


KoiosAbout 16 minNotePaper ReadNLPComputer VisionNeurIPS
PiPa: Pixel- and Patch-wise Self-supervised Learning for Domain Adaptative Semantic Segmentation

PiPa: Pixel- and Patch-wise Self-supervised Learning for Domain Adaptative Semantic Segmentation

Basic Information

  • Mu Chen, Zhedong Zheng, Yi Yang, Tat-Seng Chua
  • 2022 ACM Multimedia

問題描述

這一篇與過去看過的 DACS, ProDA, DAFormer, HRDA 同樣都是以 Unsupervised 的方式解決 Semantic Segmentationb 的 Domain Adaptation問題。


KoiosAbout 7 minNotePaper ReadDomain AdaptationComputer VisionACM Multimedia
Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Basic Information

  • Sergey Ioffe, Christian Szegedy from Google Reserch, @google.com
  • 2015 ICML

Problem Description

在訓練Deep Neural Network的過程中,每層layer的input distribution都會隨著前一層的parameter而改變(Internal Covariate Shift),容易導致Model收斂緩不容易train,而Batch Normalization就是透過在每一層的input都做一次normalization,來解決Internal Covariates Shift,減緩Vanishing Gradient的問題,同時也達到regularization的效果。


XavierAbout 13 minNotePaper ReadDeep Neural Networks
Dropout: A Simple Way to Prevent Neural Networks from Overfitting

Dropout: A Simple Way to Prevent Neural Networks from Overfitting

Basic Information

  • Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, Ruslan Salakhutdinov @ Toronto University
  • 2014 JMLR

問題描述

在近年來發現到 Neural Network 參數越多就有越強大的表達能力,並且通常會有更好的表現。不過隨著參數量的上升,我們也發現到模型越來越會傾向於 Overfitting。


KoiosAbout 10 minNotePaper ReadRegularizationJMLR
HRDA: Context-Aware High-Resolution Domain-Adaptive Semantic Segmentation

HRDA: Context-Aware High-Resolution Domain-Adaptive Semantic Segmentation

Basic Information

  • Lukas Hoyer, Dengxin Dai, Luc Van Gool @ ETH Zurich & MPI for Informatics
  • 2022 ECCV

問題描述

這篇 paper 如同 DAFormer 關注在 UDA for semantic segmentation 。


KoiosAbout 14 minNotePaper ReadDomain AdaptationComputer VisionECCV
AlexNet: ImageNet Classifications with Deep Convolutional Neural Networks

AlexNet: ImageNet Classifications with Deep Convolutional Neural Networks

Basic Information

  • Alex Krizhevsky, Ilya Sutskever, Geoffrey E. Hinton @ University of Toronto
  • 2012 NeurIPS

Problem Description

這是一篇將Deep Learning(深度學習)與Convolutional Neural Networks(卷積神經網路,以下簡稱CNN)運用在Computer Vision(計算機視覺)領域的開拓性論文。作者們train了一個Deep Convolutional Neural Network來分類ImageNet ILSVRC-2010資料集中的120萬張高解析度圖像,並得到了相較前人方法顯著優異許多的表現。


XavierAbout 10 minNotePaper ReadSupervised LearningConvolutional Neural NetworksComputer Vision
Agent57: Outperforming the Atari Human Benchmark

Agent57: Outperforming the Atari Human Benchmark

Basic Information

  • Adrià Puigdomènech Badia, Bilal Piot, Steven Kapturowski, et al. @ Google DeepMind
  • 2020 ICML

問題描述

在 RL 當中,Atari games 是一個相當重要的 benchmark。過去的 RL 模型已經能夠在大多的 atari games 當中獲得相當不錯的 performance,例如 MuZero、R2D2,分別在 57 個遊戲當中有 51 和 52 個遊戲是 outperform 人類的。不過可惜的是,在剩下的遊戲當中這些 SoTA 就通常完全沒辦法學習。


KoiosAbout 20 minNotePaper ReadReinforcement LearningICML