Artifact 3 – ML Training Methods Interactive Session

Artifact Overview

This artifact documents an interactive coaching session with the AIML-500 Machine Learning Training Methods Coach, an AI-powered chatbot built on SchoolAI. The session covered seven core questions about how machine learning models are trained, spanning supervised learning, unsupervised learning, reinforcement learning, algorithm selection, training pipelines, iteration, and the role of data.

I selected this artifact because it demonstrates a completely different learning modality compared to Artifacts 1 and 2. Where the AI Lab (Artifact 1) showcased tool exploration and the ML vs. DL Presentation (Artifact 2) highlighted collaborative communication, this artifact captures real-time interactive learning, critical thinking through follow-up questions, and the ability to apply ML concepts to practical scenarios under guided questioning.

ML Training Concepts Interactive AI Learning Critical Thinking Technical Problem Solving Self-Directed Inquiry Concept Application

Topics Covered

The session was structured around seven guided questions, each building on the previous one to create a comprehensive understanding of how ML models learn from data.

1 Supervised Learning

How labeled data, loss functions, and weight adjustments enable models to learn input-output mappings.

2 Unsupervised Learning

Clustering, dimensionality reduction, density models, and autoencoders for finding hidden patterns.

3 Reinforcement Learning

Agent-environment interaction, reward signals, and the exploration vs. exploitation tradeoff.

4 Algorithm Selection

Why choosing the right algorithm matters and the tradeoffs between SGD and closed-form solutions.

5 Training Pipeline Steps

The full 9-step process from data collection through deployment and monitoring.

6 Iteration & Epochs

How repeated passes through data help convergence, and when to stop to prevent overfitting.

7 Role of Data

Data quality, quantity, bias, representativeness, and how diversity drives generalization.

Highlight: Supervised Learning Deep Dive

The session began with the coach asking how a supervised learning model learns from training data. I provided an explanation using spam detection as a real-world example, then the coach walked through a detailed house price prediction scenario showing the full training loop.

Mohammed (Student)

From what I understand so far, supervised learning works by giving the model labeled examples where each input has a known correct answer. The model makes a prediction, checks how far off it was from the actual label, and then adjusts its internal weights to get closer next time. Like in spam detection, the model sees thousands of emails tagged as spam or not spam, and gradually learns the patterns.

AI Coach

Great — your summary is correct. Here's a concise step-by-step walkthrough using house-price prediction: Define features and label, initialize model, forward pass, compute error (loss), backpropagate, update weights, repeat across many houses until the model learns sensible weights.

Highlight: SGD vs. Closed-Form Solutions

After discussing algorithm selection, the coach challenged me with a follow-up question about why SGD is preferred over closed-form solutions for large datasets. This required applying computational complexity knowledge to a practical ML scenario.

Mohammed (Student)

SGD is preferred over closed-form solutions on big datasets because of computational efficiency. A closed-form solution like the Normal Equation requires inverting a matrix, which has O(n^3) time complexity. When you have millions of data points, that becomes extremely expensive. SGD processes data in small batches, scales well regardless of dataset size, uses less memory, and can work in an online learning setting where new data keeps coming in.

AI Coach

Confirmed. No single theory fits all, partly empirical. Start with mini-batches (32-256). Use learning-rate scaling rules: if you multiply batch size by k, multiply LR by k with warmup. Monitor training/validation loss and pick the best trade-off.

Highlight: Self-Directed Inquiry

Beyond the seven required questions, I asked independent follow-up questions about transfer learning, batch size selection, semi-supervised learning, and concept drift. This demonstrated initiative and the ability to connect topics across the ML landscape.

Mohammed (Student)

How do transfer learning techniques change the traditional training pipeline? When we use a pre-trained model like ResNet or BERT, we skip the early training steps and just fine-tune on our specific task. Does that fundamentally change how we think about the role of data, since we need much less of it?

AI Coach

Transfer learning changes the pipeline significantly. Phase A: Freeze all backbone layers, train only the new classifier for 5-10 epochs. Phase B: Unfreeze last 1-2 blocks, reduce LR, and train 5-10 more epochs. You still need representative task-specific data for fine-tuning and to avoid domain mismatch.

Mohammed (Student)

What happens when the data itself changes over time, like in a recommendation system where user preferences shift? How do practitioners handle concept drift during the training and retraining process?

Tools & Platform

SchoolAI — AI-powered coaching platform used to host the interactive training methods session
AIML-500 ML Training Methods Coach — Custom chatbot configured with structured questions and adaptive follow-ups
Real-time chat interface — Text-based conversation format enabling iterative question-and-answer learning

Value Proposition

This artifact demonstrates that I can engage deeply with ML training concepts in an interactive, adaptive learning environment. Unlike static assignments, this session required real-time critical thinking: each answer was evaluated by the coach, and follow-up questions tested whether I truly understood the material or was just reciting definitions. The fact that I went beyond the required questions to explore transfer learning, batch size optimization, and concept drift shows genuine intellectual curiosity and the ability to connect individual topics into a broader understanding of the ML training landscape.

For a potential employer or collaborator, this artifact shows that I do not just learn passively. I ask questions, challenge assumptions, and seek to understand how concepts apply in real production environments.

Reflection

This artifact represents a unique addition to my portfolio because it captures learning in action. Artifacts 1 and 2 showcase finished products (a tool exploration report and a polished presentation), but this artifact shows how I think through problems in real time. The interactive format pushed me to articulate my understanding precisely because the coach would immediately challenge vague or incomplete answers.

The most valuable lesson from this session was understanding that ML training is not a linear process but an interconnected system. Data quality affects model selection, which affects algorithm choice, which affects how you iterate, which circles back to data. Seeing these connections through guided questioning was far more effective than reading about them in isolation.

One thing I would do differently is prepare specific real-world scenarios before entering the session. While I was able to think on my feet with examples like spam detection and recommendation systems, having pre-planned case studies would have allowed even deeper exploration of each topic.

If I were presenting this artifact to a technical hiring manager, I would emphasize the self-directed questions about transfer learning and concept drift, since those show awareness of practical challenges that go beyond textbook ML training. For a non-technical audience, I would focus on the analogies used throughout the session (studying for exams, post office sorting) to demonstrate my ability to translate complex concepts into accessible language.

The feedback from the AI coach throughout the session helped me refine my explanations in real time. For instance, when I explained validation sets, the coach confirmed my answer but then pushed me further with the overfitting scenario exercise, which deepened my understanding of early stopping as a practical technique rather than just a theoretical concept.