Computer Vision for Business

Session 5 Overview

Duration: 3 hours | Responsible AI and final projects

Why Ethics Matter in Computer Vision

CV systems make consequential decisions about people

The CV Ethics Landscape

Four interconnected areas of concern

Ethical Challenges in Computer Vision

Facial recognition as a case study

Key Ethical Issues

Common challenges in facial recognition systems

Case Study 1: Clearview AI

Scraped billions of photos from social media

Case Study 2: Amazon Rekognition

Facial recognition sold to police departments, later paused

Case Study 3: Emotion Recognition in Hiring

AI analyzing candidate facial expressions during video interviews

Bias in Computer Vision Systems

Bias can enter at every stage of the ML pipeline

Sources of Bias

Three main categories of bias in CV systems

The Gender Shades Study

Joy Buolamwini & Timnit Gebru (2018)

Bias Mitigation: The Audit Process

Systematic approach to detecting and measuring bias

Bias Audit Implementation

Measure performance across demographic groups

def audit_model_fairness(model, test_data, demographic_labels, true_labels):
    """Audit model performance across demographic groups."""

    results = {}

    for group in set(demographic_labels):
        # Get predictions for this group
        group_indices = [i for i, d in enumerate(demographic_labels) if d == group]
        group_data = [test_data[i] for i in group_indices]
        group_labels = [true_labels[i] for i in group_indices]

        predictions = [model.predict(x) for x in group_data]

        # Calculate metrics
        accuracy = sum(p == l for p, l in zip(predictions, group_labels)) / len(predictions)

        results[group] = {
            "accuracy": accuracy,
            "sample_size": len(predictions)
        }

    # Flag disparities
    accuracies = [r["accuracy"] for r in results.values()]
    if max(accuracies) - min(accuracies) > 0.05:
        print("WARNING: Accuracy disparity > 5% across groups")

    return results

Bias Mitigation Strategies

Interventions at different stages

Privacy and Regulatory Landscape

Global regulations affecting CV systems

GDPR Requirements for CV

Europe's General Data Protection Regulation

EU AI Act Risk Categories

World's first comprehensive AI regulation (2024)

High-Risk AI System Requirements

What the EU AI Act requires for CV applications

GDPR-Compliant Processing

Building compliance into your pipeline

@dataclass
class DataProcessingRecord:
    """GDPR-compliant data processing record."""
    processing_id: str
    timestamp: datetime
    purpose: str
    legal_basis: str  # consent, legitimate_interest, contract
    data_subject_id: str  # pseudonymized
    data_categories: list
    retention_period: str
    automated_decision: bool
    human_review: bool
    recipients: list

def process_image_with_compliance(image_path, purpose, consent_record):
    """Process image with GDPR compliance measures."""
    record = DataProcessingRecord(
        processing_id=generate_uuid(),
        timestamp=datetime.now(),
        purpose=purpose,
        legal_basis="consent",
        data_categories=["biometric_data", "image"],
        retention_period="30_days",
        ...
    )
    log_processing(record)
    result = cv_model.process(image_path)
    return filter_to_purpose(result, purpose)  # Data minimization

Framework for Responsible CV

A holistic approach to ethical AI deployment

Pre-Deployment Checklist

Before launching any CV system

Transparency Practices

Making CV systems understandable and accountable

Generating Explainable Outputs

Code for transparent predictions

def generate_prediction_explanation(model, image, prediction):
    """Generate explainable output for CV predictions."""

    explanation = {
        "prediction": prediction["class"],
        "confidence": prediction["confidence"],
        "model_version": model.version,
        "processing_timestamp": datetime.now().isoformat(),

        # Confidence breakdown
        "alternative_classes": prediction["all_probabilities"],

        # Limitations disclosure
        "known_limitations": [
            "Model trained on Western retail products",
            "Performance may vary with unusual lighting",
            "Confidence threshold for automated decisions: 0.85"
        ],

        # Recourse information
        "appeal_process": "Contact support@company.com for review"
    }

    return explanation

Implementing Human Oversight

Keeping humans in the loop for critical decisions

Final Project Presentations

Putting it all together

Presentation Format

10 minutes per team + 5 minutes Q&A

Section 1: Business Problem

2 minutes - Setting the context

Section 2: Technical Approach

3 minutes - How you built it

Section 3: Live Demo

3 minutes - Show, don't tell

Section 4: Results & Analysis

2 minutes - Honest assessment

Evaluation Criteria

How projects will be assessed

Final Project Deliverables

What you need to submit

Suggested Project Ideas

Choose based on your industry interest

Recommended Resources

Continue your learning journey

Course Journey Recap

What we covered over 18 hours

Key Takeaways

The essential lessons

The Future of Computer Vision

Emerging trends and opportunities

Thank You!

Computer Vision for Business - Course Complete