not just a black box interpretable deep learning for
play

Not Just a Black Box: Interpretable Deep Learning for Genomics - PowerPoint PPT Presentation

Sep 20, 2022 •163 likes •984 views

Not Just a Black Box: Interpretable Deep Learning for Genomics Presented by: AvanA Shrikumar 1 With great power comes really poor interpretability Deep Power Learning Traditional machine learning Classical statistics 2

  1. Not Just a Black Box: Interpretable Deep Learning for Genomics Presented by: AvanA Shrikumar 1

  2. With great power comes really poor interpretability… Deep Power Learning Traditional machine learning Classical statistics 2 Interpretability

  3. With great power comes really poor interpretability… Deep Interpretable Deep Power Learning Learning Traditional machine learning Classical statistics 2 Interpretability

  4. QuesAons for the model • Which parts of the input are the most important for making a given predic:on? 3

  5. QuesAons for the model • Which parts of the input are the most important for making a given predic:on? • What are the recurring pa<erns in the input? 3

  6. QuesAons for the model • Which parts of the input are the most important for making a given predic:on? • What are the recurring pa<erns in the input? 4

  7. How can we find the important parts of the input for a given predic:on? Idea 1: perturbaAon Output … … … … Yellow = inputs 5

  8. How can we find the important parts of the input for a given predic:on? Idea 1: perturbaAon Output … … … … Yellow = inputs 5

  9. How can we find the important parts of the input for a given predic:on? Idea 1: perturbaAon Output … … … … Yellow = inputs 5

  10. How can we find the important parts of the input for a given predic:on? Idea 1: perturbaAon Output … … … … Yellow = inputs 5

  11. How can we find the important parts of the input for a given predic:on? Idea 1: perturbaAon Output … … … … Yellow = inputs 5

  12. How can we find the important parts of the input for a given predic:on? Idea 1: perturbaAon Output … … … … Yellow = inputs 5

  13. How can we find the important parts of the input for a given predic:on? Idea 1: perturbaAon Output … … … … Yellow = inputs 5

  14. How can we find the important parts of the input for a given predic:on? Idea 1: perturbaAon Output Drawbacks … 1) Computa:onal efficiency - … requires one forward prop for each perturba:on … … Yellow = inputs 5

  15. How can we find the important parts of the input for a given predic:on? Idea 1: perturbaAon Output Drawbacks … 1) Computa:onal efficiency - … requires one forward prop for each perturba:on 2) Satura:on … … Yellow = inputs 5

  16. Satura:on problem illustrated y 1 y 2 i 1 1 0 i 2 i 1 + i 2 6

  17. Satura:on problem illustrated =1 y 1 y 2 i 1 1 =1 0 i 2 =1 i 1 + i 2 6

  18. Satura:on problem illustrated =1 y 1 y 2 i 1 1 =1 0 i 2 =1 i 1 + i 2 0 6

  19. How can we find the important parts of the input for a given predic:on? Idea 2: backpropagate importance Output … … … … Yellow = inputs 7

  20. How can we find the important parts of the input for a given predic:on? Idea 2: backpropagate importance Output … … … … Yellow = inputs 7

  21. How can we find the important parts of the input for a given predic:on? Idea 2: backpropagate importance Output … … … … Yellow = inputs 7

  22. How can we find the important parts of the input for a given predic:on? Idea 2: backpropagate importance Output Examples: - Gradients (Simonyan et al.) … - Deconvolu:onal Networks (Zeiler & Fergus) - Guided Backpropaga:on (Springenberg et al.) … - Layerwise Relevance Propaga:on (Bach et al.) - Integrated Gradients (Sundararajan et al.) … … Yellow = inputs 7

  23. How can we find the important parts of the input for a given predic:on? Idea 2: backpropagate importance Output Examples: - Gradients (Simonyan et al.) … - Deconvolu:onal Networks (Zeiler & Fergus) - Guided Backpropaga:on (Springenberg et al.) … - Layerwise Relevance Propaga:on (Bach et al.) - Integrated Gradients (Sundararajan et al.) - DeepLIFT (Learning Important FeaTures) … - h<ps://github.com/kundajelab/ deepli^, ICML 2017 … - With Peyton Greenside and Anshul Yellow = inputs Kundaje 7

  24. Satura:on revisited y 1 y i 1 i 2 1 2 0 i 1 + i 2 8

  25. Satura:on revisited When (i 1 + i 2 ) >= 1, y gradient is 0 1 y i 1 i 2 1 2 0 i 1 + i 2 8

  26. The DeepLIFT solu:on: difference from reference Reference: i 1 =0 & i 2 =0 y 1 y i 1 i 2 1 2 0 i 1 + i 2 9

  27. The DeepLIFT solu:on: difference from reference Reference: i 1 =0 & i 2 =0 y y=0 when (i 1 + i 2 ) = 0 (reference) 1 y i 1 i 2 1 2 0 i 1 + i 2 9

  28. The DeepLIFT solu:on: difference from reference Reference: i 1 =0 & i 2 =0 y y=0 when (i 1 + i 2 ) = 0 (reference) At (i 1 + i 2 ) = 2, 1 y the “difference from reference” is +1, NOT 0 i 1 i 2 1 2 0 i 1 + i 2 9

  29. The DeepLIFT solu:on: difference from reference Reference: i 1 =0 & i 2 =0 y y=0 when (i 1 + i 2 ) = 0 (reference) At (i 1 + i 2 ) = 2, 1 y the “difference from reference” is +1, NOT 0 i 1 i 2 1 2 0 i 1 + i 2 DeepLIFT addresses other failure modes besides saturaAon (see paper) 9

  30. Reference ma<ers! CIFAR10 model, class = “ship” Original 10

  31. Reference ma<ers! CIFAR10 model, class = “ship” DeepLIFT Reference Original scores 10

  32. Reference ma<ers! CIFAR10 model, class = “ship” SuggesAons on how to pick a DeepLIFT Reference Original scores reference : - MNIST: all zeros (background) 10

  33. Reference ma<ers! CIFAR10 model, class = “ship” SuggesAons on how to pick a DeepLIFT Reference Original scores reference : - MNIST: all zeros (background) - Consider using a distribuAon of references - E.g. mul:ple references generated by shuffling a genomic sequence 10

  34. Eg: morphing 8 to a 3 or a 6 original 8->3 8->6 Backprop Guided Integrated gradients DeepLIFT 11

  35. Example biological problem: understanding stem cell differen:a:on liver cells cardiac cells fer:lized egg blood cells Cell-types are different because different genes are turned on 12

  36. Example biological problem: understanding stem cell differen:a:on liver cells cardiac cells fer:lized egg blood cells Cell-types are different because different genes are turned on How is cell-type-specific gene expression controlled? 12

  37. Example biological problem: understanding stem cell differen:a:on liver cells cardiac cells fer:lized egg blood cells Cell-types are different because different genes are turned on How is cell-type-specific gene expression controlled? Ans: “control elements” act like switches to turn genes on 12

  38. “Control Elements” are switches that turn genes DNA sequence of a gene Control element 13

  39. “Control Elements” are switches that turn genes Sequence contain “DNA words” that controller proteins bind to ACGTGTAACTGATAATGCCGATATT DNA sequence of a gene Control element 13

  40. “Control Elements” are switches that turn genes Sequence contain “DNA words” that controller proteins bind to ACGTGTAACTGATAATGCCGATATT DNA sequence of a gene Control element Controller proteins bind to DNA words 13

  41. “Control Elements” are switches that turn genes DNA sequence of a gene Control element + controller proteins loop over… 13

  42. “Control Elements” are switches that turn genes …and ac:vate nearby genes DNA sequence of a gene Control element + controller proteins loop over… 13

  43. 89%* of disease-associated mutaAons are outside genes! DNA sequence of a gene Controller proteins 14 *Stranger et al ., Genet. , 2011

  44. 89%* of disease-associated mutaAons are outside genes! Control element has “DNA words” that controller proteins bind to ACGTGTAACTGATAATGCCGATATT DNA sequence of a gene Controller proteins 14 *Stranger et al ., Genet. , 2011

  45. 89%* of disease-associated mutaAons are outside genes! Control element has “DNA words” that controller proteins bind to ACGTGTAACTGATAATGCCGATATT DNA sequence of a gene Controller proteins 14 *Stranger et al ., Genet. , 2011

  46. 89%* of disease-associated mutaAons are outside genes! Control element has “DNA words” that controller proteins bind to ACGTGTAACTGATAATGCCGATATT DNA sequence of a gene Controller proteins 14 *Stranger et al ., Genet. , 2011

  47. 89%* of disease-associated mutaAons are outside genes! Control element has “DNA words” that controller proteins bind to ACGTGTAACTGATAATGCCGATATT DNA sequence of a gene Controller proteins Many posi:ons in a control element are not essen:al for its func:on! 14 *Stranger et al ., Genet. , 2011

  48. 89%* of disease-associated mutaAons are outside genes! Control element has “DNA words” that controller proteins bind to ACGTGTAACTGATAATGCCGATATT DNA sequence of a gene Controller proteins Many posi:ons in a control element are not essen:al for its func:on! à Which posiAons in controller elements maber? 14 *Stranger et al ., Genet. , 2011

  49. Q: Which posiAons in control elements maber? 15

  50. Q: Which posiAons in control elements maber? Experimentally measure control elements in different :ssues 15

  51. Q: Which posiAons in control elements maber? Predict :ssue- Experimentally specific ac:vity of measure control control elements elements in from sequence using deep different :ssues learning 15

Recommend

From ML Successes to Applications ICIP18 Tutorial on Interpretable Deep Learning 2 Black Box

From ML Successes to Applications ICIP18 Tutorial on Interpretable Deep Learning 2 Black Box

From ML Successes to Applications ICIP18 Tutorial on Interpretable Deep Learning 2 Black Box Models Huge volumes of data Solve task Computing power Deep Neural Network Information (implicit) ICIP18 Tutorial on Interpretable Deep

540 views • 33 slides
Not Just a Black Box: Interpretable Deep Learning for Genomics Avan&gt; Shrikumar, Peyton

Not Just a Black Box: Interpretable Deep Learning for Genomics Avan&gt; Shrikumar, Peyton

Not Just a Black Box: Interpretable Deep Learning for Genomics Avan&gt; Shrikumar, Peyton Greenside, Anshul Kundaje Peyton Anshul 1 With great power comes really poor interpretability Deep Interpretable Deep Power Learning Learning

783 views • 41 slides
Towards Interpretable Deep Learning for Natural Language Processing Roy Schwartz University of

Towards Interpretable Deep Learning for Natural Language Processing Roy Schwartz University of

Towards Interpretable Deep Learning for Natural Language Processing Roy Schwartz University of Washington &amp; Allen Institute for Artificial Intelligence December 2018 1 / 37 ( Deep-Learning -Based) AI Today 2 / 37 3 / 37 Deep Learning

1.63k views • 140 slides
Interpretable &amp; Transparent Deep Learning Fraunhofer HHI, Machine Learning Group Wojciech

Interpretable &amp; Transparent Deep Learning Fraunhofer HHI, Machine Learning Group Wojciech

Fraunhofer Image Processing Heinrich Hertz Institute Interpretable &amp; Transparent Deep Learning Fraunhofer HHI, Machine Learning Group Wojciech Samek Northern Lights Deep Learning Workshop (NLDL19) Troms, Norway 8th January 2019

1.15k views • 51 slides
Toward In Interpretable De Deep Re Reinforcement Lea Learning g wi with Li Linea ear Model

Toward In Interpretable De Deep Re Reinforcement Lea Learning g wi with Li Linea ear Model

Toward In Interpretable De Deep Re Reinforcement Lea Learning g wi with Li Linea ear Model el U-Tr Trees Guiliang Liu, Oliver Schulte, Wang Zhu, Qingcan Li Machine Learning Lab, ECML-PKDD 2018 Presentation PROBLEM DEFINITION

372 views • 15 slides
SDRL: Interpretable and Data-efficient Deep Liu Reinforcement Learning Introduction Background

SDRL: Interpretable and Data-efficient Deep Liu Reinforcement Learning Introduction Background

SDRL: Symbolic Deep Reinforcement Learning SDRL: Interpretable and Data-efficient Deep Liu Reinforcement Learning Introduction Background Leveraging Symbolic Planning Method Experiment Conclusion Bo Liu and Future Work Auburn

1k views • 20 slides
Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable

Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable

Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable Models Instead Cynthia Rudin, Duke University Presenters : Sreya Dutta Roy, Ziqian Lin 1 Overview Introduction Explainable ML Vs

1.04k views • 39 slides
Interpretable Multimodal Deep Learning for Objective Diagnosis, Prognosis and Biomarker Discovery

Interpretable Multimodal Deep Learning for Objective Diagnosis, Prognosis and Biomarker Discovery

Interpretable Multimodal Deep Learning for Objective Diagnosis, Prognosis and Biomarker Discovery Faisal Mahmood, PhD Postdoctoral Fellow Department of Biomedical Engineering Johns Hopkins University faisalm@jhu.edu | http://faisal.ai March

1.08k views • 62 slides
with Interpretable Deep Learning Presented by: Avanti Shrikumar Kundaje Lab Stanford University

with Interpretable Deep Learning Presented by: Avanti Shrikumar Kundaje Lab Stanford University

Understanding Genome Regulation with Interpretable Deep Learning Presented by: Avanti Shrikumar Kundaje Lab Stanford University Example biological problem: understanding stem cell differentiation liver cells Lung cells fertilized egg

1.24k views • 74 slides
Layer-wise Relevance Propagation in Neural Neural Networks Deep Learning Shortcomings Networks

Layer-wise Relevance Propagation in Neural Neural Networks Deep Learning Shortcomings Networks

LRP Ariyan Zarei Motivation Having More interpretable Layer-wise Relevance Propagation in Neural Neural Networks Deep Learning Shortcomings Networks to have more interpretable Papers and Demo Introduction Machine Learning models

373 views • 35 slides
EMBC Tutorial on Interpretable and Transparent Deep Learning Wojciech Samek Grgoire Montavon

EMBC Tutorial on Interpretable and Transparent Deep Learning Wojciech Samek Grgoire Montavon

EMBC Tutorial on Interpretable and Transparent Deep Learning Wojciech Samek Grgoire Montavon Klaus-Robert Mller (Fraunhofer HHI) (TU Berlin) (TU Berlin) 13:30 - 14:00 Introduction KRM 14:00 - 15:00 T echniques for Interpretability GM

545 views • 54 slides
EMBC Tutorial on Interpretable and Transparent Deep Learning Wojciech Samek Grgoire Montavon

EMBC Tutorial on Interpretable and Transparent Deep Learning Wojciech Samek Grgoire Montavon

EMBC Tutorial on Interpretable and Transparent Deep Learning Wojciech Samek Grgoire Montavon Klaus-Robert Mller (Fraunhofer HHI) (TU Berlin) (TU Berlin) 13:30 - 14:00 Introduction KRM 14:00 - 15:00 Techniques for Interpretability GM

1.03k views • 23 slides
2019.07.04 Xie, P., Zuo, K., Zhang, Y., Li, F., Yin, M., &amp; Lu, K. (2019). Interpretable

2019.07.04 Xie, P., Zuo, K., Zhang, Y., Li, F., Yin, M., &amp; Lu, K. (2019). Interpretable

2019.07.04 Xie, P., Zuo, K., Zhang, Y., Li, F., Yin, M., &amp; Lu, K. (2019). Interpretable Classification from Skin Cancer Histology Slides Ryan Chen Using Deep Learning: A Retrospective Multicenter Study. arXiv preprint arXiv:1904.06156.

488 views • 15 slides
Tutorials Interpretable Deep Learning: Towards Understanding &amp; Explaining DNNs P a r t 3

Tutorials Interpretable Deep Learning: Towards Understanding &amp; Explaining DNNs P a r t 3

Tutorials Interpretable Deep Learning: Towards Understanding &amp; Explaining DNNs P a r t 3 : V a l i d a t i n g E x p l a n a t i o n s W o j c i e c h S a m e k , G r g o i r e M o n t

805 views • 28 slides
Tutorials Interpretable Deep Learning: Towards Understanding &amp; Explaining DNNs P a r t 2

Tutorials Interpretable Deep Learning: Towards Understanding &amp; Explaining DNNs P a r t 2

Tutorials Interpretable Deep Learning: Towards Understanding &amp; Explaining DNNs P a r t 2 : M e t h o d s o f E x p l a n a t i o n W o j c i e c h S a m e k , G r g o i r e M o n t a v

475 views • 36 slides
Bayesian Deep Learning Mohd Adnan Problems With Deep Learning What does a model not know?

Bayesian Deep Learning Mohd Adnan Problems With Deep Learning What does a model not know?

Bayesian Deep Learning Mohd Adnan Problems With Deep Learning What does a model not know? Uninterpretable black-boxes Easily fooled (AI safety) Lacks solid mathematical foundation Crucially relies on big dat Why

297 views • 15 slides
Deep Visual Models with Interpretable Features and Modularized Structures Quanshi Zhang John

Deep Visual Models with Interpretable Features and Modularized Structures Quanshi Zhang John

Deep Visual Models with Interpretable Features and Modularized Structures Quanshi Zhang John Hopcroft Center Shanghai Jiao Tong University Quanshi Zhang, Ying Nian Wu, and Song-Chun Zhu, &quot;Interpretable Convolutional Neural Networks&quot;

694 views • 42 slides
IMLI: An Incremental Framework for MaxSAT-Based Learning of Interpretable Classification Rules

IMLI: An Incremental Framework for MaxSAT-Based Learning of Interpretable Classification Rules

IMLI: An Incremental Framework for MaxSAT-Based Learning of Interpretable Classification Rules Bishwamittra Ghosh Joint work with Kuldeep S. Meel 1 Applications of Machine Learning 2 Example Dataset 3 Representation of an interpretable

1.24k views • 55 slides
Partial Substitute tong-wang@uiowa.edu Poster #67 A black-box model + High predictive

Partial Substitute tong-wang@uiowa.edu Poster #67 A black-box model + High predictive

Tong Wang Gaining Free or Low-Cost University of Iowa Transparency with Interpretable Tippie College of Business Partial Substitute tong-wang@uiowa.edu Poster #67 A black-box model + High predictive performance - non-interpretable An

289 views • 12 slides
Self-supervised Learning of Interpretable Keypoints from Unlabelled Videos

Self-supervised Learning of Interpretable Keypoints from Unlabelled Videos

Self-supervised Learning of Interpretable Keypoints from Unlabelled Videos www.robots.ox.ac.uk/~vgg/research/ Oral presentation, CVPR 2020 unsupervised_pose/ Tomas Jakab Ankush Gupta VGG, University of Oxford DeepMind, London Hakan Bilen

259 views • 14 slides
Deep Learning Theory with Application to Cancer Research Leonie Zeune, Stephan van Gils, Guus van

Deep Learning Theory with Application to Cancer Research Leonie Zeune, Stephan van Gils, Guus van

Deep Learning Theory with Application to Cancer Research Leonie Zeune, Stephan van Gils, Guus van Dalum, Leon Terstappen, Christoph Brune Inverse Problems and Machine Learning, Pasadena, USA, Feb 9-11, 2018 Deep Learning as a Black Box

909 views • 75 slides
Class Activation Map (CAM) Prof. Seungchul Lee Industrial AI Lab. Issues on CNN (or Deep

Class Activation Map (CAM) Prof. Seungchul Lee Industrial AI Lab. Issues on CNN (or Deep

Class Activation Map (CAM) Prof. Seungchul Lee Industrial AI Lab. Issues on CNN (or Deep Learning) Deep learning performs well comparing with any other existing algorithms But works as a black box A classification result is simply

698 views • 11 slides
InfoGAIL: Interpretable Imitation Learning from Visual Demonstrations Chih-Hui Ho, Chun Hu,

InfoGAIL: Interpretable Imitation Learning from Visual Demonstrations Chih-Hui Ho, Chun Hu,

InfoGAIL: Interpretable Imitation Learning from Visual Demonstrations Chih-Hui Ho, Chun Hu, Po-Jung Lai 1 Outline 1. Introduction 2. Related work Generative adversarial imitation learning (GAIL) 3. Proposed method 4. Experiment

692 views • 14 slides
Hao Su July 6, 2017 Outline Overview of 3D deep learning 3D deep learning algorithms

Hao Su July 6, 2017 Outline Overview of 3D deep learning 3D deep learning algorithms

Deep 3D Representation Learning for Visual Computing Hao Su July 6, 2017 Outline Overview of 3D deep learning 3D deep learning algorithms Conclusion 2 Outline Overview of 3D deep learning Background 3D deep learning tasks 3D deep

1.66k views • 122 slides

More recommend