20. 責任のあるAIパターン
• 様々な利害関係者へ
のモデルの影響
20
パターン 問題 解決
経験的ベ
ンチマー
クHeuristic
Benchmark
結果の良し悪
しの程度を意
思決定者へ説
明困難
過去との比較や
経験則に基づく
判断
説明可能
な予測
Explainable
Predictions
予測の説明困
難
シンプルなモデ
ル採用、予測結
果における説明
など
公平性レ
ンズ
Fairness
Lens
不均衡データ
に基づく異な
る人々のグ
ループに対す
る問題のある
バイアス
(What-If toolや
Fairness
Indicatorsなどに
よる)訓練前後
のデータセット
の分析、結果比
較、均衡化など
(Short)
using 19990914 build on win98 using
the new account wizard if I add
multiple accounts with the same
server…
(Long)
this mean the problem was with their
web page and its cool now or we still
need to figure out what the actual
bug was and fix that…
Y. Noyori, H. Washizaki, et al., “Extracting features related to
bug fixing time of bug reports by deep learning and gradient-
based visualization,” 2021 IEEE International Conference on
Artificial Intelligence and Computer Applications
23. MLDP および SEP4MLA の整理
Topology Programming Model operations
レジリエ
ント
サービ
ング
再現性
責任・説
明性
モデル
訓練
問題
表現
データ
表現
Hashed Feature Embeddings
Feature Cross Multimodal Input
Reframing Multilabel
Ensembles Cascade
Neutral Class Rebalancing
Useful Overfitting
Checkpoints
Transfer Learning
Distribution Strategy
Stateless Serving Function
Hyperparameter
Tuning
Batch Serving Continued Model Evaluation
Keyed Predictions
Windowed Inference
Repeatable Splitting
Transform
Bridged Schema
Two-Phase Predictions
Feature
Store
Model
Versioning
Heuristic Benchmark
Workflow Pipeline
Fairness Lens
Explainable
Predictions
Different Workloads in Different
Computing Environments
Distinguish Business Logic from ML
Models
ML Gateway Routing Architecture
Microservice Architecture for ML
Lambda
Architecture
Kappa
Architecture
Data Lake for ML
Parameter-
Server Abstraction
Data flows
up, Model
flow down
Secure Aggregation
Separation of Concerns
and Modularization of
ML Components
Discard PoC Code
ML Versioning
Encapsulate ML models
within Rule-base Safeguards
Deployable Canary Model
今後の拡充検討エリア