VIP20

本课程属于：029-AI for CAE Simulation VVUQ & Geometric Deep Learning English Technical Courses · 共 20 门相关课程，组合学习更完整。

查看所属VIP全部课程 →

← 返回所属目录咨询 / 购买

Out-of-Distribution Detection and AI Failure Diagnosis

30 Chapters

View Course Samples Online

Course overview

scope, key concepts, and the role of OOD detection in AI failure diagnosis.

Foundations of distribution shift

covariate shift, label shift, concept shift, and semantic shift with real-world examples.

Foundationsdistribution

Failure modes of machine learning models under distribution shift

overconfidence, silent errors, and degradation patterns.

distributionoverconfidence

Taxonomy of OOD detection methods

discriminative, generative, distance-based, and Bayesian approaches.

Baseline OOD scoring functions

softmax confidence, max logit, entropy, and their limitations.

Distance-based OOD detection

Mahalanobis distance, cosine similarity, and feature-space density estimation.

Distance-basedOOD

Generative model-based OOD detection

likelihood ratios, variational autoencoders, and normalizing flows.

Generativemodel-based

Bayesian and ensemble approaches for uncertainty quantification and OOD detection.

Bayesian and ensemble approaches for uncertainty quantification and OOD detection.

Bayesianensemble

Outlier exposure and contrastive learning for robust OOD detection.

Outlier exposure and contrastive learning for robust OOD detection.

exposurecontrastive

Evaluation metrics and benchmarks for OOD detection

AUROC, FPR@95TPR, OOD datasets, and pitfalls.

Evaluationbenchmarks

Hyperparameter tuning and calibration for OOD detection in production pipelines.

Hyperparameter tuning and calibration for OOD detection in production pipelines.

Hyperparametercalibration

OOD detection in computer vision

image classification, object detection, and segmentation use cases.

OOD detection in natural language processing

text classification, question answering, and generative models.

OOD detection in time-series and sensor data

anomaly detection in IoT and industrial systems.

OOD detection in tabular data and healthcare applications

challenges and case studies.

Diagnosing AI failures

root cause analysis workflow from symptom to distribution shift identification.

Diagnosingfailures

diagnosing a silent failure in a deployed image classifier using OOD scores and feature analysis.

diagnosingdeployed

failure diagnosis in a recommendation system due to concept drift and OOD user behavior.

diagnosisrecommendation

OOD detection for autonomous driving perception systems and safety-critical diagnosis.

Integrating OOD detection into MLOps

monitoring, alerting, and retraining triggers.

Human-in-the-loop diagnosis

interactive tools and visual analytics for OOD investigation.

Human-in-the-loopdiagnosis

Adversarial OOD detection

robustness to adversarial outliers and evasion attacks.

Open-set recognition and OOD detection

similarities, differences, and combined approaches.

Open-setrecognition

Generalized OOD detection

multi-modal, multi-domain, and continual learning scenarios.

Scalable OOD detection for large-scale deployment

efficiency, latency, and resource constraints.

Regulatory and ethical considerations in AI failure diagnosis and OOD detection.

Regulatory and ethical considerations in AI failure diagnosis and OOD detection.

Regulatoryconsiderations

Practical project

build an OOD detection pipeline for a real-world dataset and diagnose simulated failures.

Practical project

implement a monitoring dashboard for OOD detection and failure diagnosis in a production-like setting.

implementmonitoring

Advanced topics

foundation models, zero-shot OOD detection, and open challenges.

Advancedfoundation

best practices, pitfalls, and future directions in OOD detection and AI failure diagnosis.

practicespitfalls