cvpaper.challenge の Meta Study Group 発表スライド
cvpaper.challenge はコンピュータビジョン分野の今を映し、トレンドを創り出す挑戦です。論文サマリ・アイディア考案・議論・実装・論文投稿に取り組み、凡ゆる知識を共有します。2019の目標「トップ会議30+本投稿」「2回以上のトップ会議網羅的サーベイ」
http://xpaperchallenge.org/cv/
This document discusses methods for automated machine learning (AutoML) and optimization of hyperparameters. It focuses on accelerating the Nelder-Mead method for hyperparameter optimization using predictive parallel evaluation. Specifically, it proposes using a Gaussian process to model the objective function and perform predictive evaluations in parallel to reduce the number of actual function evaluations needed by the Nelder-Mead method. The results show this approach reduces evaluations by 49-63% compared to baseline methods.
【DL輪読会】Efficiently Modeling Long Sequences with Structured State SpacesDeep Learning JP
This document summarizes a research paper on modeling long-range dependencies in sequence data using structured state space models and deep learning. The proposed S4 model (1) derives recurrent and convolutional representations of state space models, (2) improves long-term memory using HiPPO matrices, and (3) efficiently computes state space model convolution kernels. Experiments show S4 outperforms existing methods on various long-range dependency tasks, achieves fast and memory-efficient computation comparable to efficient Transformers, and performs competitively as a general sequence model.
The document summarizes a research paper that compares the performance of MLP-based models to Transformer-based models on various natural language processing and computer vision tasks. The key points are:
1. Gated MLP (gMLP) architectures can achieve performance comparable to Transformers on most tasks, demonstrating that attention mechanisms may not be strictly necessary.
2. However, attention still provides benefits for some NLP tasks, as models combining gMLP and attention outperformed pure gMLP models on certain benchmarks.
3. For computer vision, gMLP achieved results close to Vision Transformers and CNNs on image classification, indicating gMLP can match their data efficiency.
【DL輪読会】Efficiently Modeling Long Sequences with Structured State SpacesDeep Learning JP
This document summarizes a research paper on modeling long-range dependencies in sequence data using structured state space models and deep learning. The proposed S4 model (1) derives recurrent and convolutional representations of state space models, (2) improves long-term memory using HiPPO matrices, and (3) efficiently computes state space model convolution kernels. Experiments show S4 outperforms existing methods on various long-range dependency tasks, achieves fast and memory-efficient computation comparable to efficient Transformers, and performs competitively as a general sequence model.
The document summarizes a research paper that compares the performance of MLP-based models to Transformer-based models on various natural language processing and computer vision tasks. The key points are:
1. Gated MLP (gMLP) architectures can achieve performance comparable to Transformers on most tasks, demonstrating that attention mechanisms may not be strictly necessary.
2. However, attention still provides benefits for some NLP tasks, as models combining gMLP and attention outperformed pure gMLP models on certain benchmarks.
3. For computer vision, gMLP achieved results close to Vision Transformers and CNNs on image classification, indicating gMLP can match their data efficiency.
Semi supervised, weakly-supervised, unsupervised, and active learningYusuke Uchida
An overview of semi supervised learning, weakly-supervised learning, unsupervised learning, and active learning.
Focused on recent deep learning-based image recognition approaches.
【DL輪読会】Deep Transformers without Shortcuts: Modifying Self-attention for Fait...Deep Learning JP
The document proposes modifications to self-attention in Transformers to improve faithful signal propagation without shortcuts like skip connections or layer normalization. Specifically, it introduces a normalization-free network that uses dynamic isometry to ensure unitary transformations, a ReZero technique to implement skip connections without adding shortcuts, and modifications to attention and normalization techniques to address issues like rank collapse in Transformers. The methods are evaluated on tasks like CIFAR-10 classification and language modeling, demonstrating improved performance over standard Transformer architectures.
5. 背景
5
(1). Boosting: where the coeffcients associated with the combinations of the
single models are actually trained, instead of simply taking average;
(2). Bootstrapping/Bagging: the training data are different for each single model;
(3). Ensemble of models of different types and architectures;
(4). Ensemble of random features or decision trees.
■アンサンブルの理論解析
• いくつかの状況設定で理論解析はあるが、単純平均のアンサンブルにおける理論解析がない
単純平均のアンサンブル学習の理論解析に着目
■単純平均のアンサンブル学習の理論解析
• 初期化乱数のみ異なるモデル(学習データ、学習率、アーキテクチャ固定)における以下の現
象を
理論的に説明することを試みる
Training average does not work: 学習前にモデルをアンサンブルしても効果
なし
Knowledge distillation works:単一モデルに複数モデルから蒸留できる
Self-distillation works:単一モデルから別の単一モデルへの蒸留でも性能が向上