On AI and ML: Looking Beyond the Surface
When exploring new technologies, it's easy to fall into the habit of treating them as just another API-driven platform—something that speaks Python, C++, or Java for the die-hards. But that mindset can miss what’s genuinely groundbreaking.
To avoid that trap, I start by mapping real-world analogies to the technology. It helps me understand the core concepts faster and accelerates my learning curve.
Medical frameworks are especially apt when it comes to AI—diagnostics, pattern recognition, and decision-making all resonate. On the business side, the analogy of financial cycles—from the heavy lift of annual accounting to the rhythm of monthly reporting—grounds abstract tech in the realities of real-world operations.
By translating the unfamiliar into the familiar, you will get to what really matters: how this tech changes the tactics of the game.
The Purpose of AI and ML: More Than Code, It’s About Capability
Artificial Intelligence (AI) is not a single tool or algorithm—it’s a broad ambition: to build systems that exhibit behaviors we associate with human intelligence. That could mean understanding language, making decisions, recognising images, or even learning from experience. If AI is the “goal” of making machines smart, then Machine Learning (ML) is the “method” that has gotten us closest to that goal.
Machine Learning is the engine room of modern AI. Rather than manually programming logic for every situation, ML lets us feed data into algorithms and allow the machine to identify patterns, adjust internal weights, and improve performance on its own. It’s the difference between giving a car turn-by-turn instructions for every journey, versus teaching it how to drive by watching and learning.
The relationship between AI and ML is layered:
-
AI is the umbrella concept: machines doing “intelligent” things.
-
ML is a core enabler: algorithms learning from data to make decisions or predictions.
-
Other branches of AI (e.g. symbolic logic, expert systems) exist, but ML dominates because of its adaptability and success in real-world applications.
A Key to Understanding: Map It to What You Know
Think of ML like the diagnostic engine in medicine: it doesn’t know what lupus is—but it’s seen thousands of similar patient cases and can suggest probabilities. Or imagine AI as the seasoned accountant: it doesn’t just total receipts, it spots irregularities, forecasts trends, and recommends actions.
The takeaway? AI isn’t just about replacing human effort—it’s about reimagining how decisions get made and how patterns get surfaced. In businesses, AI reveals what Excel can’t: dynamic insights. In medicine, it’s the assistive second opinion. In operations, it's the tool that learns faster than any SOP.
The magic of AI and ML isn't in the syntax—they're not just Python-flavored functions. The power lies in what they let us model, automate, and optimise. They turn messy, complex inputs into structured decisions—and with the right framing, they turn technologists into translators of insight.
AI and ML IN CONTEXT
| Aspect | Artificial Intelligence (AI) | Machine Learning (ML) |
|---|---|---|
| Definition | Broad field aiming to create systems that can simulate human intelligence | A subset of AI focused on systems that learn from data and improve over time |
| Scope | Encompasses reasoning, planning, perception, language, robotics, etc. | Primarily focuses on pattern recognition, prediction, and data-driven learning |
| Goal | Enable machines to perform tasks requiring human-like intelligence | Develop models that can learn from data without being explicitly programmed |
| Methods Used | Logic, rule-based systems, search algorithms, expert systems, ML, deep learning | Supervised learning, unsupervised learning, reinforcement learning |
| Human Interaction | May require more complex input (knowledge engineering, symbolic logic) | Relies heavily on data availability and computational models |
| Example Applications | Chatbots, autonomous vehicles, language translation, game AI | Spam detection, image recognition, stock prediction, medical diagnosis |
| Learning Capability | Not always capable of learning; can be hardcoded (e.g., rule-based systems) | Learns from historical data to generalize to new data |
| Dependency | Can exist without ML (e.g., expert systems) | Depends on AI as a broader concept |
FIRST A REAL WORLD TOPIC: I CHOSE MS BECAUSE IT IS A DISEASE I KNOW A BIT ABOUT
| Multiple Sclerosis Use Case | ML Algorithm / Mathematical Foundations | Azure Candidate Services | AWS Candidate Services |
|---|---|---|---|
| Predict MS Progression (e.g., RRMS → SPMS) |
Convolutional Neural Networks (CNNs) – Linear algebra (tensors, matrix multiplications) – Gradient descent – Activation functions |
Azure Machine Learning Azure Databricks Azure Kubernetes Service (AKS) Azure Blob Storage |
Amazon SageMaker AWS Lambda Amazon S3 AWS Batch |
| Classify MS Lesions in MRI Scans | Support Vector Machines (SVMs) – Dot product kernels – Optimization (Lagrange multipliers) – Geometry of hyperplanes |
Azure ML Pipelines Azure Computer Vision Azure Functions |
Amazon SageMaker Ground Truth AWS Rekognition AWS EC2 |
| Discover MS Subtypes via Gene Expression | K-Means Clustering, PCA – Euclidean distance – Vector space geometry – Variance minimization |
Azure Synapse + ML Azure Databricks Azure ML Designer |
Amazon SageMaker Studio Amazon EMR AWS Glue |
| Predict Relapse Events | Random Forests / Gradient Boosted Trees – Entropy – Information gain – Bagging and boosting |
Azure AutoML Azure ML Interpretability SDK Azure Monitor |
Amazon SageMaker Autopilot AWS CloudWatch Amazon Forecast |
| Personalised Treatment Recommendations | Reinforcement Learning – Markov Decision Processes – Bellman equations – Policy optimization |
Azure Personalizer Azure RL Toolkit |
Amazon SageMaker RL AWS DeepRacer |
| Symptom Tracking via Wearable Data | RNNs, LSTMs – Sequence modeling – Time-series forecasting – Nonlinear dynamics |
Azure IoT Hub Azure Stream Analytics Azure ML |
AWS IoT Core Amazon Kinesis Amazon SageMaker |
TL;DR FOR THE 'ML Algorithm / Mathematical Foundations' FORMULA ABOVE
| Algorithm / Group | Mathematical Foundations | Simplified Explanation | Used In |
|---|---|---|---|
| Convolutional Neural Networks (CNNs) | Tensors, Matrix Multiplication Gradient Descent Activation Functions |
Learns patterns in image-like data by applying filters; adjusts weights to minimize error | AI, ML |
| Support Vector Machines (SVMs) | Dot Product Kernels Optimization (Lagrange Multipliers) Geometry of Hyperplanes |
Finds the best boundary between categories using geometric reasoning | ML |
| K-Means Clustering / PCA | Euclidean Distance Vector Space Geometry Variance Minimization |
Groups similar items together (K-Means); finds patterns in data dimensions (PCA) | ML |
| Random Forests / Gradient Boosted Trees | Entropy Information Gain Bagging and Boosting |
Builds many decision trees and combines them for better accuracy | ML |
| Reinforcement Learning | Markov Decision Processes Bellman Equations Policy Optimization |
Learns from actions and rewards, like trial-and-error for machines | AI |
| RNNs, LSTMs | Sequence Modeling Time-Series Forecasting Nonlinear Dynamics |
Works well with sequences like text or sensor data by remembering past information | AI, ML |
TYPICAL MATHS AND STATISTICS FORMULA USED UNDER THE HOOD
| Algorithm | Mathematical Foundation (Simplified) | Non-Technical Explanation | Used In | Azure Services | AWS Services |
|---|---|---|---|---|---|
| Linear Regression | Linear algebra, least squares | Finds a straight line to best predict numbers | ML | Azure Machine Learning, Azure Synapse | Amazon SageMaker, Amazon Forecast |
| Logistic Regression | Probability, sigmoid function | Estimates yes/no outcomes | ML | Azure ML Studio, Azure AutoML | SageMaker Autopilot, Amazon Comprehend |
| Decision Trees | Entropy, information gain | Makes decisions like a flowchart | ML, AI | Azure ML Designer, Azure Databricks | SageMaker, AWS Glue |
| Random Forest | Ensemble of trees, bagging | Combines many small decisions for a better result | ML | Azure AutoML, Azure Synapse | SageMaker, AWS Forecast |
| Gradient Boosted Trees | Gradient descent, boosting | Builds better trees one by one to correct mistakes | ML | Azure ML SDK + XGBoost | SageMaker XGBoost, Amazon Personalize |
| K-Means Clustering | Euclidean distance, vector space | Groups similar items together without labels | ML | Azure ML Designer, Azure Databricks | SageMaker Clustering, Amazon EMR |
| Principal Component Analysis (PCA) | Matrix decomposition, eigenvectors | Shrinks large data into key patterns | ML | Azure Synapse Analytics, Azure ML | SageMaker PCA, AWS Glue |
| Support Vector Machine (SVM) | Geometry, dot products | Draws the best boundary between categories | ML | Azure ML Studio, Azure Databricks | SageMaker SVM, AWS EC2 |
| Neural Networks | Tensors, matrix ops, activation functions | Mimics brain-like patterns to learn complex things | ML, AI | Azure Machine Learning, Azure GPU VMs | SageMaker Neural Network, AWS Inferentia |
| Convolutional Neural Network (CNN) | Linear algebra, filters | Great at seeing patterns in images | AI, ML | Azure Computer Vision, Azure ML + GPUs | SageMaker Vision, Amazon Rekognition |
| Recurrent Neural Network (RNN) | Sequences, time steps | Learns from sequences like text or time-series | AI, ML | Azure ML (LSTM), Azure Stream Analytics | SageMaker RNN, Amazon Kinesis |
| Reinforcement Learning | MDP, Bellman equations | Learns by trying actions and getting rewards | AI | Azure Personalizer, Azure RL Toolkit | SageMaker RL, AWS DeepRacer |
| Naive Bayes | Bayes’ theorem, probability | Classifies things based on past likelihood | ML | Azure ML Designer, Azure Synapse | SageMaker, Amazon Comprehend |