AD 4-7

4 EU AI Act & Standards

4

Categories of Risk

Unacceptable Risk
These systems are prohibited.
Ex.: social scoring, mass surveillance, other harmful practices.

High Risk
These systems are permitted but subject to strict compliance requirements.
Ex.: recruitment systems, credit assessment, insurance assessment.

Limited Risk
Transparency obligations apply.
Ex.: generative AI, emotion recognition systems.

Minimal Risk
Mostly unaffected by regulation.

Definition of an AI System

Machine-based system that:
- Operates with varying levels of autonomy.
- May adapt after deployment.
- Infers outputs from received inputs.
- Produces predictions, recommendations, content, or decisions.
- Influences physical or virtual environments.

Exemptions and Actors

Exemptions:
National security
Defense applications
Scientific research
AI systems in research and development phase
Non-professional private use
Open-source software (unless classified as forbidden or high-risk)

Actors:
AI providers
Deployers
Importers
Distributors
Authorized representatives
Affected persons within the EU
Non-EU entities whose AI outputs are used in the EU

Provider vs Deployer

Provider
Entity that:
- Develops an AI system
- Places the AI system on the market
Examples: AI company, software vendor, autonomous driving software supplier.
Responsibilities:
- Conduct conformity assessments
- Maintain quality management systems
- Establish post-market monitoring
- Report incidents
- Cooperate with authorities
- Provide technical information

Deployer
Entity that uses an AI system under its authority.
Examples: hospital using diagnostic AI, bank using credit-scoring AI, fleet operator using autonomous vehicles.
Responsibilities:
- Monitor operation
- Supply relevant input
- Maintain logs
- Cooperate with providers

Provider → responsibility for development and market placement.
Deployer → responsibility for operational use.

Forbidden AI Applications

Manipulation
Exploitation of Vulnerable Groups
Biometric Categorization
Social Scoring
Predictive Policing
Untargeted Facial Image Scraping
Emotion Recognition

Mandatory General Requirements

Risk management
Data governance
Technical documentation
Record keeping
Transparency
Human oversight
Accuracy
Robustness
Cybersecurity

Detailed Requirements for High-Risk AI

Risk Management
Data Governance
Technical Documentation
Record Keeping
Transparency
Human Oversight
Robustness, Accuracy, and Cybersecurity

Governance Structure of the AI Act

AI Office
- GPAI supervision
- Monitoring implementation
- International cooperation

European AI Board
- Expertise sharing
- Best practices

Guidance Advisory Forum
- Stakeholder involvement
- Technical advice

Scientific Panel
- Risk assessment
- Evaluation methodologies
- Model capability assessment

Important Standards

Foundations
ISO/IEC 22989: AI concepts and terminology
ISO/IEC 23053: ML framework
ISO/IEC 5338: AI lifecycle processes
ISO/IEC 25059: AI quality model

Governance
ISO/IEC 42001: AI management systems
ISO/IEC 23894: AI risk management
ISO/IEC 38507: AI governance

Trustworthiness
ISO/IEC TR 24028: Trustworthy AI
ISO/IEC TR 24027: Bias in AI
ISO/IEC TS 4213: ML classification performance
ISO/IEC TS 6254: Explainability

Data Quality
ISO/IEC 5259 series
ISO/IEC 8183 data lifecycle framework

ISO/IEC TS 22440-1

— Read in the Slides —
AI components are classified according to Application Usage Levels (AUL).
Software Technology Classes (SWTC)
Hazard Analysis and Risk Assessment (HARA)
AI Fault Analysis
AI Fault Avoidance
AI Fault Control

5 ML in AD introduction

5

Machine Learning Principle

Input Data
↓
Learning Algorithm
↓
Optimization of Weights
↓
Prediction Model
↓
Output

ML Category

Supervised Learning (Learning from labeled examples (x,y))

Unsupervised Learning (Discovering patterns without labels)

Reinforcement Learning (Learning through rewards and penalties)

Self-Supervised Learning (Learning representations directly from large amounts of data)

Tesla Perception Architecture

Multi-Camera Images
↓
Neural Network Backbone
(Extract meaningful features from camera images)
↓
Feature Extraction
↓
Multi-Scale Feature Fusion
↓
Detection Heads
(Convert extracted features into object predictions)
↓
Occupancy and Vector Space Representation
(Transform perception into a common vector-space representation)
↓
Video Processing Modules
↓
Integrated Perception Output

Labeling types

Offline Labeling
(Offline reconstruction using past and future observations)

Static and Dynamic Environment Labeling