A/B. モデリングとMLパイプライン統合の流れ
メタモデル
安全性・高信頼性を含む要求・
アーキテクチャモデリング
DNN訓練・評価・修正
パイプライン
具
体
化・
洗
練
要求分析・設計
DNN評価
問題の可視化
DNN評価
解決の可視化
42
OK
OK OK
NG
OK OK
OK
NG
NG
OK OK OK
[ML.VP1 🡨
AI.VP1]
Provide reliable
real-time object
detection system
for driving
decision making in
highway (incl.
traffic sign
detection and
lane/vehicle
detection)
• [ML.DS1] Procured
datasets
• [ML.DS2] Internal
database from
collection during
operation
• [ML.DC1] Open and
commercial datasets
• [ML.DC2] Data
collected during
operation (image and
identification result)
•[ML.F1 🡨
AI.D1/AI.D3]
Bounding box
for object (incl.
other vehicles
or signs)
•[ML.F2 🡨
AI.D2] Ridge
detection for
lane detection
[ML.BM1]
Models will be
developed,
tested, and
deployed to cars
monthly
• [ML.PT1] Input:
image from sensors
• [ML.PT2 ← AI.D]
Output: traffic signs,
lane marking,
vehicles, and
pedestrians.
[ML.De1] Use
prediction results
for decision-
making in self-
driving system
[ML.IS1]
Using test data,
achieve very high
recall and high
precision in
following condition:
night, rainy, and
general condition
Datasets is split into
80:20 ratio
[ML.MP1]
Prediction should
be made in
batches real
time.
[ML.M1] Input data monitoring
[ML.VP1 🡨
AI.VP1]
Provide reliable
real-time object
detection system
for driving
decision making in
highway (incl.
traffic sign
detection and
lane/vehicle
detection)
• [ML.DS1] Procured
datasets
• [ML.DS2] Internal
database from
collection during
operation
• [ML.DC1] Open and
commercial datasets
• [ML.DC2] Data
collected during
operation (image and
identification result)
•[ML.F1 🡨
AI.D1/AI.D3]
Bounding box
for object (incl.
other vehicles
or signs)
•[ML.F2 🡨
AI.D2] Ridge
detection for
lane detection
[ML.BM1]
Models will be
developed,
tested, and
deployed to cars
monthly
• [ML.PT1] Input:
image from sensors
• [ML.PT2 ← AI.D]
Output: traffic signs,
lane marking,
vehicles, and
pedestrians.
[ML.De1] Use
prediction results
for decision-
making in self-
driving system
[ML.IS1]
Using test data,
achieve very high
recall and high
precision in
following condition:
night, rainy, and
general condition
Datasets is split into
80:20 ratio
[ML.MP1]
Prediction should
be made in
batches real
time.
[ML.M1] Input data monitoring
[ML.VP1 🡨
AI.VP1]
Provide reliable
real-time object
detection system
for driving
decision making in
highway (incl.
traffic sign
detection and
lane/vehicle
detection)
•[ML.DS1] Procured
datasets
•[ML.DS2] Internal
database from
collection during
operation
•[ML.DC1] Open and
commercial datasets
•[ML.DC2] Data
collected during
operation (image and
identification result)
•[ML.F1 🡨
AI.D1/AI.D3]
Bounding box
for object (incl.
other vehicles
or signs)
•[ML.F2 🡨
AI.D2] Ridge
detection for
lane detection
[ML.BM1]
Models will be
developed,
tested, and
deployed to cars
monthly
•[ML.PT1] Input:
image from sensors
•[ML.PT2 ← AI.D]
Output: traffic signs,
lane marking,
vehicles, and
pedestrians.
[ML.De1] Use
prediction results
for decision-
making in self-
driving system
[ML.IS1]
Using test data,
achieve very high
recall and high
precision in
following condition:
night, rainy, and
general condition
Datasets is split into
80:20 ratio
[ML.MP1]
Prediction should
be made in
batches real
time.
[ML.M1] Input data monitoring
修正戦略の追加
DNN訓練
DNN修正
Jati H. Husen, Hironori Washizaki, Hnin Thandar Tun, Nobukazu Yoshioka, Yoshiaki Fukazawa, Hironori Takeuchi, Hiroshi Tanaka, Kazuki Munakata, “Extensible Modeling
Framework for Reliable Machine Learning System Analysis,” 2nd International Conference on AI Engineering – Software Engineering for AI (CAIN’23)