Explainable AI that Can be Used for Judgment with Responsibility

Transcription

1 Explain Workshop Explainable AI that Can be Used for Judgment with Responsibility FUJITSU LABORATORIES LTD. 5th April 08

2 About me Name: Hajime Morita Research interests: Natural Language Processing Summarization Morphological Analysis Information Extraction Recent jobs: I received my Ph.D. from Tokyo Institute of Technology (05) Kyoto University (05-07) Fujitu Laboratory (07 -)

3 Outline Introduction of explainable AI Knowledge Graph (short) Deep Tensor Explainable AI

4 Why we need explainable AI If we applied IN-explainable AI to clinical decision. She has cancer. score = 0.99 She looks well. I can't believe AI. Let's see a state for a while. I found a small cancer in her stomach! xx month ago If we applied explainable AI to clinical decision. Explainable AI Output Inference result Verify Cancer? really? Let's carry out more detailed examination Clinical evidence Cooperation based on the evidence report. 3

5 Application to Genomic Medicine We developed a prototype of AI that infers which gene mutation of the patient is associated with disease. Gene analysis (sequencer) A diagnosis (drawing blood) Gene mutation Analyzing mutation Investigating DB, literature Report creation Identification of the driver mutation Making clinical decision Gene mutation Explainable AI Inference of driver mutation Evidence a b c d e f Web Literature Bio DB The presentation of the therapeutic drug to personality (genetic abnormality) PubMed Medical articles 7million Bio DB 3million records Gene Ontology Disease Ontology Ref. Hokkaido Univ. Hospital( 4

6 Formation of Evidence Path by Knowledge Graph Mutation Gene Drug Disease g g>a Losartan Ventricular hypertrophy AGTR Tachycardia Renal tubular dysgenesis Cited references and databases dbnsfp:chr3:g g>a PubMed: PubMed: PubMed: PubMed: The mutation cause an abnormality in the gene (AGTR) and the abnormality can causes diseases such as tachycardia. 5

7 Overview Two key components Deep Tensor is a neural classifier of graph data Knowledge graph is an extremely large graph knowledge base () Explaining the important factor for the inference Outputting the factors which strongly influenced the inference result through Deep Tensor Input graph Inference factor identification Outputs Inference factors Evidence a b c d e g Inference result Inference factors Inference result Deep Tensor () Explaining the evidence for inference result a b c d e g f Knowledge graph generates a evidence path from input to the inference result based on the inference factors. Evidence formation Knowledge Graph 6

8 Knowledge Graph 7

9 What is knowledge graph? Knowledge Nodes and a edge that connects the nodes. subject relation Francis Harry Compton Crick found object DNA double helix Knowledge graph Nodes Gene, mutation, drug, disease, etc... Edges Drug A is responsive to disease B Gene C has an function D Mutation E is located on Gene F... c.74c>g Mutation Gene PTPN LEOPARD syndrome Disease Knowledge Graph 8

10 How to make knowledge graph From existing database Extraction of relational data from public databases Relations between gene and its attributes (name, ID, function, etc.) Relations between protains From literature Knowledge extraction using Natural Language Processing technologies some mutations in PTPN have been identified in LEOPARD syndrome Named Entity Recognition Relation Extraction c.74c>g PTPC PTPN Identification * PTPC is a synonym of PTPN c.74c>g Mutation LEOPARD syndrome Gene PTPN LEOPARD syndrome Disease Knowledge Graph 9

11 Deep Tensor Koji Maruhashi, Masaru Todoriki, Takuya Ohwa, Keisuke Goto, Yu Hasegawa, Hiroya Inakoshi, Hirokazu Anai, Learning Multi-way Relations via Tensor Decomposition with Neural Networks, AAAI 08 (AAAI 08), February 08. 0

12 Naïve Idea Linearly convert graph data as close as a target pattern Optimize the target pattern along with Neural Networks Graph data Target Pattern expressing important features Red: features used by conventional methods Linear conversion Class A Tensor Decomposition Class B Target pattern optimization Extended Backpropagation Backpropagation Classification errors Classify graph data accurately with high interpretability

13 Difficulties Using Graph for Deep Leaning Image The pixel coordinate number is fixed Input (,) (,) (,3) (,4) ~ 0 Graph data Exists a lot of Numbering possibilities Input ~ Input ~ Coordinate label caused results unstabiliy

14 Tensor Representation R R R3 R4 Multi-way data can be represented as a tensor Multi-way data Tensor representation Sub Relation Obj S R O S R3 O S3 R O O OO3O4 S S S3 S4 3

15 Tensor Decomposition R R R3 R4 R R R3 R4 Approximate a tensor X by a core tensor X multiplied by factor matrices {C k } Results are easy to interpret in terms of nodes and edges. sbj rel obj S R O S R3 O S3 R O O OO3O4 S S S3 S4 X X, C k C k O OO3O4 S S S3 S4 = argmin X X k k C k T Can we leverage tensor decomposition? 4 X

16 Leveraging Tensor Decomposition Analyze graph data more efficiently Graph data Target Pattern Graph isomorphism determination Neural Network (NN) Too much cost! Tensor-based expression Structure Restricted Tensor Decompsition approximation Core tensor A new type of tensor decomposition Target core tensor close Input of NN Optimize using Extended Backpropagation 5

17 Structure Restricted Tensor Decomposition (SRTD) Details Calculate X which minimize X k I k <J k k C k X k I k J k k C k T by using C k which minimize X k I k <J k k C k V k I k J k k C k T subject to C k T C k = I I k J k, C k C k T = I(I k < J k ) SRTD X I I K Core tensor J J K X V d-layer Neural Network C k W d {b d } Factor matrices Target core tensor y 6

18 How to interpret prediction results? D D D3 D4 Solution: Learn interpretable models that output similar results as Neural Networks P PP3P4 S S S3 S4 X Tensor Decomposition is interpretable! C k X V Neural Network Black Box y Similar results Interpretable Model (Linear regression) y = X, W + b X White Box y 8

19 Local Interpretable Models Details Use perturbed samples X p based on LIME [Ribeiro+ KDD6] Learn interpretable model g A X p = X p, W A + b A which minimize p π p y A p g A X p where π p = exp( X X p σ ) X Class B X 3 X W A π() π() π(3) X X 4 π(4) π(5) π(6) X 6 Class A X 5 Probability that X p classified into A by DeepTensor y A () = 0.9 y A () =.0 y A (3) = 0.9 y A (4) = 0. y A (5) = 0.0 y A (6) = 0. Contribution score is calculated by X W A k k C k T 9 is

20 Recap: Overview Deep Tensor + knowledge graph Knowledge graph generates input graph (extracting subgraph representing about the mutation) Deep Tensor infers which the mutation cause disease or not, and output inference factors. Knowledge graph makes evidence graph based on the inference factors. () Explaining the important factor for the inference Outputting the factors which strongly influenced the inference result through Deep Tensor Input graph Inference factor identification Outputs Inference factors Evidence a b c d e g mutation cause disease or not Inference result mutation Inference factors Inference result Deep Tensor () Explaining the evidence for inference result a b c d e g f Knowledge graph generates a evidence path from input to the inference result based on the inference factors. Evidence formation Knowledge Graph 0

21 Conclusion Explainable AI is a key technology for cooperation of AI and humans We developed a prototype of explainable AI The explainable AI explains important part of input data, and the evidence that explains the important part and inference result. We are now trying to proof of the concept of explainable AI, by cooperating with several medical groups.

22