MSP-AI Meeting

周惶振

Fri, Apr 2, 2021 5:47 AM

Meeting Link

• Server Setup

• Meeting

Paper

2021~2022 Fall

• [] 12/02 Labelling unlabelled videos from scratch with multi-modal self-supervision

• [] 11/04 Online Continual Learning from Imbalanced Data

• [] [10/14 Self-supervised Co-training for Video Representation Learning]

  • https://www.robots.ox.ac.uk/~vgg/research/CoCLR/

• [] [10/07]

• [] 09/30 Automatic Shortcut Removal for Self-Supervised Representation Learning

  • Summary
  • Contribtuion
  • Strength
    • Novelty
    • Contribtuion
    • I can use
      *
  • Weakness
    • What needs more

• 09/09 Audio-Visual Instance Discrimination with Cross-Modal Agreement

  • Strength
    • Novelty
    • Contribtuion
    • I can use
      *
  • Weakness
    • What needs more
      *

• [ ]

2020~2021 Spring

• 08/05 CVPR' 2018 End-to-End Incremental Learning

  • Background
  • Novelty
    • Distilla loss

• [07/29]

• 07/01 CVPR 2020': S3VAE: Self-Supervised Sequential VAE for Representation Disentanglement and Data Generation

  • Background Knowledge:
    • Time-varying
      • motion
    • Time-invariant
      • objects
      • timbre of speakers
    • Benefits of learning disentangled representations
      • More explainable
      • Easier and more efficient to manipulate data gen- eration
    • Disentangled Sequential Data Generation
    • Self-Supervised Learning
      • Intrinsic labels
      • Auxiliary data
      • Off-the-shelf tools
    • Variational inference
  • Main Challenges:
    • Label is too expensive
    • Unsurvised learning is hard to perform well
  • Main Novelty:
    • Leverage supervisory signals from intrinsic la- bels to regularize the static representation and off-the-shelf tools to regularize the dynamic representation
    • Model both the prior and the posterior of zt by recurrent models independently -> consistent dynamic information in synthetic sequences
    • Ensure full disentanglement of zf and zt by posterior factorization
  • Contributions:
    • Fewer works have explored representation disentanglement for sequential data generation
  • Proposed Approach:
    *

• 06/24 NeurIPS 2020': Causal Intervention for Weakly-Supervised Semantic Segmentation

  • Background Knowledge:

    • Classification Activation Map (CAM)
    • Pseudo-Masks
    • associates non-causal but positively correlated pixels to labels
    • disassociates causal but negatively correlated ones
    • multi-label soft-margin loss: A mixed objective optimization network for the recognition of facial attributes

  • Graphical model vs Graph model

    • Graphical model

  • Main Challenges:

  • Main Novelty and Contributions:

    • Context Adjustment (CONTA)

  • Proposed Approach:

• 06/17 WACV 2021': Class-incremental Learning via Deep Model Consolidation

• Main Challenges:

• Main Novelty and Contributions:

• Proposed Approach:

  • Step1: To train a (t − s)-class classifier using training data Dnew, which we refer as the new model
    $f_{new}(x;\Theta new)$
  • Step2: To consolidate the old model and the new model (Deep Model Consolidation (DMC) for image classification)

• Similarity: Federated learning

• 06/10 CVPR 2019': Learning a Unified Classifier Incrementally via Rebalancing

  • Main Challenges:

    • Incremental Learning:
      • Preserve parameters of the orignal model
      • Catastropic forgetting
        

  • Main Novelty and Contributions:

    • For catastropic forgetting:
      • Cosine-norm
      • Less-forget constraint
      • Inter-class separation

• 06/03 ICLR 2021': RETHINKING THE ROLE OF GRADIENT-BASED ATTRIBUTION METHODS FOR MODEL INTERPRETABILITY

  • Simpler:
    • Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps
  • Interpretability vs prior
  • Prior 、Posterior 和Likelihood

• [ ]05/28 On the uncertainty of self-supervised monocular depth estimation

  • Uncertainty modeling methods (how reliable on your predictions)
  • Mean and Variance
  • Cosine Scheduling
  • Uncertainty metric:
    • Area Under the Sparsification Error (AUSE)
    • Area Under the Random Gain (AURU) (higher is better)

• Attention-Based Models for Speech Recognition

  • Adaptive Learning Rate Algorithm
  • Use sigmoid to apply on attention
  • Optimizer:
    • Older: AdaDelta
    • Adam
  • [ ]
  • 中文
  • 其他參考資料:
    • Neural Turing Machines
    • 李宏毅老師

• Teaching a GAN what not to learn

  • 知乎
  • Github
  • Frechet inception distance (FID) socres

• 紀錄Miro

• [Bootstrap Your Own Latent: A New Approach to Self-Supervised Learning]

  • Novelty
    • No negative smaples
  • Self-supervised
  • Bootstrap Method，Bootstrapping

• [] [Deep Adversarial Metric Learning]

  • 
  • [ ]強baseline-A Metric Learning Reality Check
  • [ ]Metric Learning Survey
  • [ ]A tutorial on distance metric learning: Mathematical foundations,algorithms, experimental analysis, prospects and challenges
  • [ ]Learning to Self-Train for Semi-Supervised Few-Shot Classification
  • Use non-linear distance
    • Linear Mahlanobis distance
    • Large margin nearest neighbor
  • [] [Novelty]
    • Generate synthetic negative examples
    • Goal
    • 
    • 
    • 
  • Other Refs.
    • Introduce
    • [Survey](file:///C:/Users/User/Desktop/symmetry-11-01066-v2.pdf)
    • Deep Metric Learning with Graph Consistency
    • FOCAL: EFFICIENT FULLY-OFFLINE METAREINFORCEMENT LEARNING VIA DISTANCE METRIC LEARNING AND BEHAVIOR REGULARIZATION
    • UNSUPERVISED DOMAIN ADAPTATION FOR DISTANCE METRIC LEARNING
    • Fewer is More: A Deep Graph Metric Learning
      Perspective Using Fewer Proxies

• RANet: Region Attention Network for Semantic Segmentation

  • Supplemental
  • Main Goal:
    • Classifying each pixel in an image to a class label
    • Important task for understanding images
  • Main idea:
    • 
  • Method:
    • 
    • Region Construction Block
      • 
      • 
      • 
        • $G_{ij}$ = -1,0,1

• Stand-Alone Self-Attention in Vision Models
  

  

  • Another Paper:Attention Augmented Convolutional Networks

  • Novelty:

    • Replace CNN with attention
      • Specialized Layer for images
        • Spatially localized features
        • Translation invariance
      • Less memory and computations
    • Create a stand-alone model using only self-attetnion
      • 
      • 

  • Results:

    • To show the efficiency of models
    •