Tatsuya Sueki Master thesis

Tatsuya Sueki Master thesis

• 1.
  Explanation and Predictionof False Positives in Object Detection 48-186130
• 2.
  DNN (deep neuralnetwork) (false negative) (false positive)
• 3.
  Variational Autoencoder (VAE)
• 4.
  • (feature attribution) • • • • [1]:D. Erhan et al. “Visualizing Higher-Layer Features of a Deep Network” University of Montreal 2009 [2]: J. T. Springenberg et al. “Striving for Simplicity: The All Convolutional Net” ICLR (workshop track) 2015 [3]: M. Sundararajan et al. “Axiomatic Attribution for Deep Networks” ICML 2017 [4]: D. Smilkov et al. “SmoothGrad: removing noise by adding noise” arXiv 2017 [1] [2] [3] [4]
• 5.
  TCAV (Testing withConcept Activation Vectors) [Kim 18] Contribution: ( ) ( ) ( ) ( =CAV) CAV → [Kim 18]: B. Kim et al. “Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV)” ICML 2018
• 6.
  TCAV (Testing withConcept Activation Vectors) [Kim 18] Contribution: ( ) ( ) ( ) ( =CAV) CAV → [Kim 18]: B. Kim et al. “Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV)” ICML 2018
• 7.
  TCAV (Testing withConcept Activation Vectors) [Kim 18] Contribution: ( ) ( ) ( ) ( =CAV) CAV → [Kim 18]: B. Kim et al. “Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV)” ICML 2018
• 8.
  TCAV (Testing withConcept Activation Vectors) [Kim 18] Contribution: ( ) ( ) ( ) ( =CAV) CAV → CAV [Kim 18]: B. Kim et al. “Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV)” ICML 2018
• 9.
  TCAV (Testing withConcept Activation Vectors) [Kim 18] Contribution: ( ) ( ) ( ) ( =CAV) CAV → • • • [Kim 18]: B. Kim et al. “Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV)” ICML 2018
• 10.
  ACE (Automatic Concept-basedExplanation) [Ghorbani 19] Contribution: [Achanta 12] → TCAV [Kim 18] [Ghorbani 19]: A. Ghorbani et al. “Towards Automatic Concept-based Explanations” arXiv 2019 [Achanta 12]: R. Achanta et al. “SLIC Superpixels Compared to State-of-the-Art Superpixel Methods” IEEE TPAMI 2012 [Kim 18]: B. Kim et al. “Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV)” ICML 2018
• 11.
  bounding box → TCAV [Kim18] [Achanta 12] [Ghorbani 19]: A. Ghorbani et al. “Towards Automatic Concept-based Explanations” arXiv 2019 [Achanta 12]: R. Achanta et al. “SLIC Superpixels Compared to State-of-the-Art Superpixel Methods” IEEE TPAMI 2012 [Kim 18]: B. Kim et al. “Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV)” ICML 2018
• 12.
  → TCAV [Kim 18] 0.9 0.2 0.1 boundingbox : 0.8 [Achanta 12] [Ghorbani 19]: A. Ghorbani et al. “Towards Automatic Concept-based Explanations” arXiv 2019 [Achanta 12]: R. Achanta et al. “SLIC Superpixels Compared to State-of-the-Art Superpixel Methods” IEEE TPAMI 2012 [Kim 18]: B. Kim et al. “Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV)” ICML 2018
• 13.
  → TCAV [Kim 18] 0.9 0.2 0.1 :0.8 bounding box : 0.7 [Achanta 12] [Ghorbani 19]: A. Ghorbani et al. “Towards Automatic Concept-based Explanations” arXiv 2019 [Achanta 12]: R. Achanta et al. “SLIC Superpixels Compared to State-of-the-Art Superpixel Methods” IEEE TPAMI 2012 [Kim 18]: B. Kim et al. “Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV)” ICML 2018
• 14.
  • Benchmark BDD100K[Yu 18] • person • CenterNet [Zhou 19], Faster R-CNN [Ren 15] [Yu 18]: F. Yu et al. “BDD100K: A Diverse Driving Video Database with Scalable Annotation Tooling” arXiv 2018 [Zhou 19]: X. Zhou et al. “Objects as Points” arXiv 2019 [Ren 15]: S. Ren et al. “Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks” NeurIPS 2015
• 16.
  2 7 11 14
• 17.
  14
• 18.
  0.3 0.3 0.3 0.9 0.8 0.6 0.7 0.50.5 0.5 0.2 0.1 0.9 0.9
• 20.
  1 1
• 21.
  10.24%
• 22.
  • • • • • 10.24% Future Work •3 • • •
• 24.
  18.23% 10.24% (IoU = 0)
• 25.
  Average Precision (AP) http://sucrose.hatenablog.com/entry/2017/02/26/224559
• 27.
  [Tao 18]: G.Tao et al. “Attacks Meet Interpretability: Attribute-steered Detection of Adversarial Samples” NeurIPS 2018 AmI (Attacks meet Interpretability) [Tao 18] Contribution:
• 28.
  • Variational Autoencoder(VAE) • J. An and S. Cho “Variational Autoencoder based Anomaly Detection using Reconstruction Probability” Special Lecture on IE 2015 • D. Park et al. “A Multimodal Anomaly Detector for Robot-Assisted Feeding Using an LSTM-based Variational Autoencoder” IEEE Robotics and Automation Letters 2018 • • P. W. Koh and P. Liang “Understanding Black-box Predictions via Influence Functions” ICML 2017 • • C. Yeh et al. “Representer Point Selection for Explaining Deep Neural Networks” NeurIPS 2018 • logit