Module Overview
Introduction
This e-portfolio documents my learning journey throughout the Machine Learning module. It showcases my development in understanding machine learning algorithms, ethical considerations, collaborative project work, and professional growth.
The portfolio contains artefacts from each unit, including discussion posts, project contributions, team meeting notes, and peer feedback, culminating in a comprehensive reflection on my learning experience.
Learning Journey
Module Units
Introduction to Machine Learning
Week 1 • August 2025
Learning Outcomes
- Developed understanding of machine learning's role in future industry
- Identified skill sets required to become proficient in machine learning
- Learned about the pitfalls of machine learning and ways to address them
- Explored the interplay among big data analytics, ML, and AI
Collaborative Discussion: The 4th Industrial Revolution
Industry 4.0 ushered in cyber-physical systems, IoT, big data, and AI, revolutionizing factories with instant monitoring, smart maintenance planning, and streamlined supply chains (Lu, 2017). As we advance to Industry 5.0, the focus is on blending human-centered design, resilience, and ecological responsibility, fostering seamless teamwork between skilled workers and intelligent machines (European Commission, 2021).
A vivid illustration comes from the LockerGoga incident that hit Norsk Hydro, a giant in the aluminium sector. The malware severely disrupted its international networks, leaving the firm with no option but to revert to paper and spreadsheets in dozens of plants (Greenberg, 2019).
Key Impacts Identified:
- Customer & operational disruption: Production came to a standstill or progressed at reduced speed, compounding delivery backlogs and eroding customer trust.
- Economic cost: Preliminary estimates put revenue and remediation losses at USD 75–100 million (Greenberg, 2019).
- Reputational cost: Coverage by national and financial press exposed soft spots in essential industrial systems, diminishing confidence among regulators and investors.
- Additional fallout: Elevated stress among workers, intensified pressure on crisis response squads, and a sprint toward stronger cyber-defensive spending.
This episode spotlights Industry 4.0's paradox: digital linking boosts productivity yet exposes latent weak points. Transitioning to Industry 5.0 requires design that prioritizes resilience, retains effective human oversight, and operates within a clear ethical framework to close the gaps revealed by this breach.
References
- European Commission (2021) Industry 5.0: Towards a sustainable, human-centric and resilient European industry. Brussels: European Commission.
- Greenberg, A. (2019) 'Aluminum Producer Norsk Hydro Still Reeling From LockerGoga Ransomware Attack', Wired, 19 March.
- Lu, Y. (2017) 'Industry 4.0: A survey on technologies, applications and open research issues', Journal of Industrial Information Integration, 6, pp. 1–10.
Exploratory Data Analysis
Week 2 • August 2025
Learning Outcomes
- Ability to undertake basic Exploratory Data Analysis (EDA)
- Enhanced understanding of dataset characteristics
- Skills in preparing datasets for machine learning
- Experience with data visualization techniques
Exploratory Data Analysis (EDA) is a critical step in data science and AI projects, providing the foundation for valid and correctly interpreted results. This unit emphasized the importance of validating raw data, checking for anomalies, and ensuring data sets are collected without errors.
Activities Completed
- Collaborative Discussion (Continued): Engaged with peer contributions from Unit 1, providing constructive feedback on their Industry 4.0 discussions following the department's peer review guidelines.
- EDA Tutorial Seminar: Participated in live seminar sessions focusing on practical EDA techniques and applications.
- Team Project Meetings: Regular team meetings to review progress and discuss project direction.
- E-Portfolio Development: Created summary of discussion post and peer responses for portfolio inclusion.
Peer Review Contributions
Provided constructive feedback to colleagues on their Industry 4.0 discussion posts, focusing on:
- Technical accuracy of Industry 4.0 concepts presented
- Quality of case study analysis and supporting evidence
- Connections made between theoretical concepts and practical implications
- Suggestions for strengthening arguments with additional references
Correlation and Regression
Week 3 • August 2025
Learning Outcomes
- Understanding of correlation coefficients and their interpretation
- Knowledge of linear regression fundamentals
- Ability to quantify relationships between variables
- Skills in applying statistical techniques to ML problems
This unit explored statistical techniques for quantifying relationships between variables—essential foundations for supervised machine learning. Understanding correlation and regression provides the mathematical framework for predictive modeling.
Summary Post: Industry 4.0 Reflection
Through engaging with peers on the Industry 4.0 discussion, several key themes emerged that deepened my understanding of machine learning's role in modern industry:
- Security as Foundation: The Norsk Hydro case highlighted how ML systems require robust security infrastructure. Peers contributed additional examples of cyber-physical vulnerabilities in smart manufacturing.
- Ethical Considerations: Discussions explored the balance between automation efficiency and workforce impacts—a recurring theme in responsible AI deployment.
- Resilience by Design: Industry 5.0's emphasis on human-centricity and sustainability emerged as a framework for addressing Industry 4.0's limitations.
Key Insight: Machine learning in industrial contexts must balance three pillars—efficiency, security, and human oversight—to achieve sustainable digital transformation.
Technical Concepts Mastered
- Pearson Correlation: Measuring linear relationships between continuous variables
- Spearman Correlation: Assessing monotonic relationships for ordinal data
- Simple Linear Regression: Predicting outcomes from single features
- Multiple Linear Regression: Incorporating multiple predictors
- R-squared Interpretation: Understanding variance explained by models
Linear Regression with Scikit-Learn
Week 4 • August 2025
Learning Outcomes
- Practical implementation of linear regression using Scikit-Learn
- Understanding of the fit-predict paradigm
- Skills in train-test splitting and model evaluation
- Experience with feature scaling and preprocessing
Team Project Formation
Project Scope Definition: The team convened to finalize our project direction, discussing various ML applications and datasets. We collectively decided to focus on NYC Airbnb market analysis.
Role Distribution:
- Data collection and preprocessing: Kanan Akparov
- Exploratory data analysis and visualization: Ade Putra Tio Aldino (My role)
- Model development and implementation: Dwane Lay
- Model evaluation and documentation: Nelson Okolie
Technical Stack Agreement
Communication Protocol:
- Weekly video meetings every Monday at 7 PM GMT
- Daily asynchronous updates via dedicated Slack channel
- Shared Google Drive for documentation and resources
- GitHub issues for tracking tasks and progress
Scikit-Learn Implementation Practice
Key Technical Learning:
- Scikit-Learn API Consistency: Understanding the fit-predict paradigm that unifies all estimators
- Train-Test Split: Implementing proper data splitting to avoid overfitting
- Feature Scaling: Learning when and how to apply standardization and normalization
- Model Evaluation Metrics: Using R², MSE, RMSE, and MAE to assess regression model quality
- Cross-Validation: Introduction to k-fold cross-validation for robust model evaluation
Clustering
Week 5 • September 2025
Learning Outcomes
- Understanding of clustering algorithm logic and techniques
- Knowledge of distance measurements (Euclidean, Manhattan, Cosine)
- Skills in K-means and agglomerative clustering methods
- Ability to evaluate clustering quality (Silhouette Score, Elbow Method)
Clustering is a fundamental unsupervised learning technique that groups objects based on similarity. This unit explored the theoretical foundations of clustering algorithms and their applications across diverse fields including pattern recognition, image analysis, bioinformatics, and machine learning.
Wiki Activity: Clustering Applications
Understanding Clustering in Business Context:
- Customer Segmentation: Grouping customers based on purchasing behavior, demographics, and preferences for targeted marketing
- Anomaly Detection: Identifying outliers in clusters can reveal fraudulent transactions or unusual patterns
- Document Organization: Text clustering helps organize large document collections, improving information retrieval
- Image Segmentation: Clustering pixels based on color and spatial features enables object detection
Team Project Development - Week 5
Project Focus Finalized: NYC Airbnb Market Analysis
Week 5 Accomplishments:
- Dataset Acquired: 48,895 listings with 16 features including price, location, room type, reviews, and availability
- Data Preprocessing: Reduced to 36,753 valid records after cleaning (75% retention)
- My EDA Contributions:
- Neighborhood distribution analysis revealing Manhattan and Brooklyn dominance
- Price variation analysis by borough using box plots and violin plots
- Correlation matrix visualization identifying key relationships (reviews vs. availability: r=0.56)
- Price distribution analysis showing left-skewed pattern with mode at $60-80
Clustering with Python
Week 6 • September 2025
Learning Outcomes
- Successfully conducted clustering analysis on large datasets (36K+ records)
- Utilized appropriate Python libraries (sklearn.cluster, matplotlib, seaborn)
- Evaluated clustering quality using multiple metrics
- Interpreted results in business context for stakeholder communication
Team Project Submission: NYC Airbnb Analysis
Executive Summary: Our team analyzed Airbnb's NYC market to understand pricing dynamics, location preferences, and market segmentation patterns. Using exploratory data analysis, linear regression, and K-means clustering, we identified actionable insights for pricing optimization and targeted marketing strategies.
K-Means Clustering Results - Three Market Segments
Cluster 0 - Budget-Friendly Segment:
- Price Range: $50-100
- Reviews: 20-50 (moderate engagement)
- Characteristics: Cost-sensitive travelers; less popular areas
- Strategy: Focus on occupancy rates and positive reviews
Cluster 1 - Mid-Range Segment:
- Price Range: $100-175
- Reviews: 50-150 (high engagement)
- Characteristics: Well-established, frequently booked properties
- Strategy: Maintain quality and leverage review reputation
Cluster 2 - Premium Segment:
- Price Range: $175-350+
- Reviews: Fewer than 20 (lower turnover)
- Characteristics: Luxury properties in prime locations
- Strategy: Emphasize exclusivity and unique amenities
Linear Regression Insights
Room Type Impact on Price:
- Private rooms: -$77 vs entire homes (baseline)
- Shared rooms: -$103 vs entire homes
Location Premium:
- Manhattan: +$52
- Brooklyn: +$23
- Queens: +$11
Key Insight: Room type and location are dominant price drivers; host characteristics have smaller but significant effects.
My Individual Contributions
- Exploratory Data Analysis Lead:
- Created comprehensive visualizations (Figures 1-7 in report)
- Conducted correlation analysis identifying key relationships
- Analyzed price distributions across neighborhoods and room types
- Data Preprocessing: Collaborated on outlier detection and treatment strategy
- Visualization and Presentation: Designed all figures using Matplotlib and Seaborn
- Report Writing: Co-authored EDA findings section
Key References
- Bishop, C.M. (2016) Pattern recognition and machine learning. Springer.
- Carvache-Franco, M. et al. (2023) 'Market segmentation in urban tourism', PLoS ONE, 18(5).
- Jiao, J. and Bai, S. (2019) 'Cities reshaped by Airbnb', Environment and Planning A, 52(1).
Introduction to Artificial Neural Networks
Week 7 • September 2025
Learning Outcomes
- Critical understanding of ANN architecture and functioning
- Ability to design and implement simple ANN artefacts
- Understanding of biological inspiration behind artificial neurons
- Knowledge of activation functions (Sigmoid, ReLU, Tanh)
This unit introduced Artificial Neural Networks (ANNs), biology-inspired algorithms that learn from their environment and adapt to make increasingly accurate decisions. ANNs represent a cornerstone of Industry 4.0, enabling machines to operate with minimal human intervention.
E-Portfolio Activity: Perceptron Implementations
Completed three progressive perceptron exercises in Jupyter Notebook:
1. Simple Perceptron (simple_perceptron.ipynb):
- Created basic single-layer perceptron
- Implemented formula: output = activation(Σ(weights × inputs) + bias)
- Recognized that single perceptrons can only solve linearly separable problems
2. Perceptron AND Operator (perceptron_AND_operator.ipynb):
- Trained perceptron to learn logical AND operation
- Input patterns: [0,0], [0,1], [1,0], [1,1]
- Observed weight adjustment through iterative learning
- Limitation: XOR problem revealed single-layer perceptron's inability to solve non-linearly separable problems
3. Multi-Layer Perceptron (multi-layer_Perceptron.ipynb):
- Implemented multi-layer network with hidden layers
- Sigmoid Activation: σ(x) = 1 / (1 + e^(-x))
- Overcomes single perceptron limitations for complex patterns
Extension Activity: Google Colab Exploration
- Platform Advantages: Free GPU access, pre-installed ML libraries, collaborative features
- TensorFlow Tutorials Completed: Basic neural network creation, MNIST digit classification
- Future Application: Recognized Colab's potential for team project work in upcoming units
Training an Artificial Neural Network
Week 8 • September-October 2025
Learning Outcomes
- Understanding of error handling mechanisms in ANNs
- Knowledge of backpropagation and gradient computation
- Experience optimizing network hyperparameters
- Practical skills in weight adjustment through gradient descent
E-Portfolio Activity: Gradient Descent Cost Function
Following the Mayo (2017) article, I implemented and experimented with gradient descent:
Experiment 1: Varying Learning Rate
- LR = 0.001 (Conservative): Slow but steady, required 5000+ iterations
- LR = 0.1 (Aggressive): Rapid initial decrease, oscillation after ~500 iterations
- LR = 1.0 (Too High): Divergence observed—cost increased rather than decreased
- Optimal LR = 0.01-0.05: Balanced speed and stability
Experiment 2: Varying Iteration Count
- 100 Iterations: Insufficient—network barely started learning
- 500 Iterations: Moderate improvement, cost still decreasing
- 1000 Iterations: Good convergence for simple problems
- 5000 Iterations: Diminishing returns after 2000 iterations
Key Insight: Early stopping based on validation error is more efficient than fixed iterations.
Weight Update Rule:
new_weight = old_weight - (learning_rate × gradient)
Introduction to Convolutional Neural Networks
Week 9 • October 2025
Learning Outcomes
- Understanding of CNN architecture for image processing
- Knowledge of convolutional layers, pooling, and feature extraction
- Critical evaluation of ethical implications in computer vision
- Awareness of both benefits and risks of CNN applications
CNNs have revolutionized computer vision, enabling applications from medical imaging to autonomous vehicles. This unit explored both the technical foundations and ethical considerations of deploying CNN-based systems.
Ethical Analysis: Computer Vision Applications
Critical Ethical Issues:
- Surveillance & Privacy: Real-time facial recognition in public spaces raises questions about consent and privacy rights
- Algorithmic Bias: Studies show facial recognition systems have significantly higher error rates for underrepresented groups (Buolamwini & Gebru, 2018)
- Deepfakes: GANs creating synthetic media for misinformation and personal harm
- Autonomous Systems: Accountability questions when CNN-based systems make errors
Positive Applications:
- Healthcare: Early disease detection in medical imaging (cancer screening, diabetic retinopathy)
- Accessibility: Computer vision assists visually impaired individuals in navigation
- Environmental: Wildlife monitoring, deforestation detection from satellite imagery
- Safety: Finding missing persons, threat detection in security screening
Recommendations for Responsible CNN Deployment:
- Establish clear legal boundaries for facial recognition in public spaces
- Develop diverse, representative training datasets to reduce bias
- Implement bias testing and mitigation throughout development lifecycle
- Create explainable AI (XAI) methods for CNN decision transparency
Natural Language Processing
Week 10 • October 2025
Learning Outcomes
- Understanding of NLP evolution from rule-based to deep learning
- Knowledge of Transformer architectures (BERT, GPT, T5)
- Skills in evaluating NLP models using domain-specific metrics
- Exploration of self-supervised learning and transfer learning in NLP
This unit explored Natural Language Processing (NLP), one of the most rapidly evolving fields in machine learning. The advent of Transformer-based architectures has revolutionized language understanding.
Understanding Transformer Architectures
Transformer Breakthrough (Vaswani et al., 2017):
- Self-Attention: Allows model to weigh importance of different words regardless of distance
- Parallel Processing: All positions processed simultaneously, dramatically reducing training time
- Positional Encoding: Maintains sequence information without sequential processing
- Multi-Head Attention: Captures different types of relationships simultaneously
BERT (Bidirectional Encoder Representations):
- Architecture: Encoder-only, bidirectional context understanding
- Training: Masked Language Modeling (MLM)
- Applications: Question answering, sentiment analysis, NER
- Strength: Excellent for understanding tasks
GPT (Generative Pre-trained Transformer):
- Architecture: Decoder-only, unidirectional processing
- Training: Autoregressive language modeling
- Applications: Text generation, creative writing, code generation
- Strength: Exceptional generation capabilities, few-shot learning
T5 (Text-to-Text Transfer Transformer):
- Unified Framework: All NLP tasks framed as text-to-text problems
- Architecture: Encoder-decoder structure
- Applications: Translation, summarization, question answering
NLP Evaluation Metrics
- BLEU: N-gram overlap for machine translation (0-100, higher is better)
- ROUGE: Recall-oriented for summarization tasks
- Perplexity: Measures how well model predicts text (lower is better)
MLOps Applications in NLP:
- Data drift detection for language models
- Automated retraining pipelines
- A/B testing for model deployment
Model Selection and Evaluation
Week 11 • October 2025
Learning Outcomes
- Mastery of model evaluation techniques and metrics
- Practical experience with transfer learning
- Implementation of explainable AI methods (Grad-CAM)
- Comprehensive hyperparameter tuning skills
Individual Project: CIFAR-10 Object Recognition with Transfer Learning
Why Transfer Learning?
- Challenge: Small images (32×32) with limited spatial information
- Risk: 50K training samples risk overfitting when training from scratch
- Solution: Reuse ImageNet pretrained knowledge (1.2M images)
- Result: +15% better accuracy vs training from scratch
MobileNetV2 Architecture
Architecture Stack:
- Input: 160×160×3 (resized from 32×32)
- Pretrained MobileNetV2 feature extractor
- Global Average Pooling
- Dense Layer (128 units, ReLU)
- Dropout (0.3)
- Output (10 units, Softmax)
Two-Phase Training Strategy
Phase 1 (5 epochs):
- Only classifier head trainable
- Learning Rate: 1e-3
- Purpose: Warm-up to avoid catastrophic forgetting
Phase 2 (15 epochs):
- Last 30% of backbone unfrozen
- Learning Rate: 1e-4 (reduced for fine-tuning)
- Result: +3-5% accuracy improvement
Performance Results
Overall Metrics:
- Test Accuracy: 94.46%
- Macro Precision: 0.945
- Macro Recall: 0.944
- Macro F1-Score: 0.944
Best Performing Classes:
- Automobile: F1 = 0.9683
- Ship: F1 = 0.9603
- Truck: F1 = 0.9563
- Airplane: F1 = 0.9523
Challenging Classes:
- Cat: F1 = 0.8813 (confusion with dog due to similar textures)
- Dog: F1 = 0.9082 (missed some true positives)
Explainability with Grad-CAM
Implemented Grad-CAM (Gradient-weighted Class Activation Mapping) to visualize which image regions influence predictions:
- Verified model focuses on relevant object features, not background
- Demonstrated meaningful learned representations
- Aligns with EU AI Act requirements for explainability
Individual Project Scope & Responsibilities
- Model Architecture Design: Independently selected MobileNetV2 as the optimal backbone after evaluating alternative architectures
- Training Pipeline Development: Designed and implemented data preprocessing, augmentation strategies, and training callbacks
- Model Evaluation: Conducted performance analysis using confusion matrix, precision–recall curves, and Top-K accuracy
- Model Explainability: Implemented Grad-CAM to interpret and visualise model decision-making
- Communication & Presentation: Produced a comprehensive fifteen-slide technical presentation summarising methodology and results
Industry 4.0 and Machine Learning
Week 12 • November 2025
Learning Outcomes
- Comprehensive understanding of Industry 4.0/5.0 concepts
- Knowledge of ML applications in modern manufacturing
- Awareness of future trends: Edge AI, Federated Learning, XAI
- Understanding of evolving job landscape in ML/AI
This unit completed a full circle from Unit 1, connecting the Norsk Hydro ransomware case to Industry 5.0's emphasis on resilience and human-centricity as frameworks for addressing Industry 4.0's challenges.
Industry 5.0 - Three Pillars
1. Human-Centricity:
- Workers and machines collaborate seamlessly
- Technology adapts to human needs
- Focus on worker wellbeing and augmentation
2. Sustainability:
- Circular economy principles
- Energy efficiency and renewable integration
- ML-driven resource management
3. Resilience:
- Adaptive systems that handle disruptions
- Diversified supply chains
- Embedded cybersecurity with graceful degradation
Future Trends in Machine Learning
- Self-Supervised Learning (SSL): Models learn from unlabeled data, democratizing AI development
- Edge AI: ML inference on edge devices—our MobileNetV2 project exemplifies these principles
- Neural Architecture Search (NAS): Automated network design using ML itself
- Federated Learning: Training without centralizing data for privacy preservation
- Explainable AI (XAI): LIME, SHAP, Grad-CAM for regulatory compliance
- Multimodal Learning: Processing text, images, audio simultaneously
Emerging Roles in Industry 4.0/5.0
- Machine Learning Engineer: Designing, training, and deploying ML models
- Data Scientist: Extracting insights from complex datasets
- AI Ethics Specialist: Ensuring responsible AI development
- MLOps Engineer: Managing ML model lifecycle in production
- Digital Twin Specialist: Creating virtual replicas of physical systems
- Edge AI Developer: Optimizing models for resource-constrained devices
Final Reflection
Connecting First and Final Units:
This unit completes a full circle from Unit 1, where I discussed the Norsk Hydro ransomware attack exemplifying Industry 4.0's paradox—digital connectivity enables unprecedented efficiency yet creates critical vulnerabilities. Now, understanding Industry 5.0's emphasis on resilience and human-centricity provides the framework for addressing these challenges.
Key Insights Gained:
- Technology is a Tool, Not a Solution: ML algorithms are powerful but require thoughtful deployment considering ethical, social, and operational contexts
- Balance is Essential: Efficiency vs. resilience, automation vs. human employment, innovation vs. sustainability
- Interdisciplinary Approach Required: Effective ML deployment requires collaboration between technical experts, domain specialists, ethicists, and policymakers
- Continuous Evolution: Industry 4.0 to 5.0 represents ongoing adaptation—not a fixed destination but a continuous journey
Personal Commitment:
As I transition from student to ML professional, I commit to contributing to Industry 5.0's vision—developing AI systems that are not only technically sophisticated but also human-centric, sustainable, and resilient. This means prioritizing explainability, considering environmental impact, advocating for responsible deployment, and maintaining human oversight in critical applications.
Module Highlights
1 Major Team Project Completed and 1 Major Individual Project
Team Project 1: NYC Airbnb Market Analysis
Unit 6 Submission
- K-Means clustering for market segmentation
- Linear regression for price prediction
- 36,753 records analyzed
- 3 distinct market segments identified
Individual Project 2: CIFAR-10 Image Classification
Unit 11 Submission
- Transfer learning with MobileNetV2
- 94.46% test accuracy achieved
- 60,000 images trained
- Grad-CAM for explainability