EVALUATE
Check two later readings
The 35- and 40-minute observations are excluded from training.
Train first, then inspect the errors.
JIGYASA / GUEST LECTURE
Machine Learning
in the Age of AI
From Classroom Concepts to Real-World Decisions
A visual journey through equations, data, learning algorithms and agentic systems. Follow one complete experiment from a mathematical model to a prediction you can calculate.
Open the one-page lecture launchpad ↗
Third-year engineering · Basic Python is enough · Synthetic teaching data
OPENING FILM / YOUTUBE ORIGINALS
How Far Is Too Far?
Start with BabyX in The Age of A.I., hosted by Robert Downey Jr. Then ask what the system senses, what was designed, and what is learned.
Source: YouTube Originals, 18 December 2019. Embedded playback needs internet; the direct YouTube links remain available if the venue blocks the player. The documentary is a discussion opener, not a benchmark for current models.
THE COMPLETE EXPERIMENT
A differential equation
becomes a learning model
Hot liquid starts at 80°C in a 25°C room. The equation describes its behaviour; seven synthetic observations teach it the unknown cooling rate. All times below are in minutes.
- Choose equation
- Observe data
- Fit / train
- Check later data
- Predict a new value
INITIAL GUESS
Seven training readings · MSE objective · 180 gradient updates
Blue curve: current model. Copper dots: synthetic training observations. Fitting and training mean the same parameter-estimation step here.
Training interval: 0–30 min. New prediction: beyond that interval.
INFERENCE
Calculate a new value
Fit k first; then substitute a new time.
Inspect the observations, training method and assumptions
| Time (min) | Observed °C | Role |
|---|---|---|
| 0 | 80 | Training |
| 5 | 70.2 | Training |
| 10 | 62.4 | Training |
| 15 | 54.9 | Training |
| 20 | 50.1 | Training |
| 25 | 44.9 | Training |
| 30 | 41.8 | Training |
| 35 | 38.4 | Held out |
| 40 | 36.2 | Held out |
Gradient descent optimises log(k) to keep the rate positive, using mean squared error, a learning rate of 0.001 and 180 updates. A new input changes t while fitted k stays fixed. The room temperature and effective vessel properties are assumed constant. A 45-minute estimate extrapolates beyond the training interval. These synthetic observations illustrate the method; they do not validate a real vessel.
Download observations CSV · Open the complete notebookTHE LECTURE
The lecture, at a glance
Begin with BabyX, learn a cooling rate from data, compare model families, and finish by inspecting intelligent systems and their evidence.
What is learning?
BabyX prompts the question. A cooling equation makes the mechanism precise.
Train a small model
Watch an unknown cooling rate update from seven synthetic readings.
Understand model families
Data and EDA lead into regression, KNN, SVM, trees, clustering and sequence models.
Judge intelligent systems
Evaluate scores, inspect hallucinations and trace an agent’s tool loop.
Equation→Data→Training→Evaluation→Inference
DATA / EDA
Read the dataset before training
The downloadable CSV has 600 fictional independent campus scenarios. Temperature and expected occupancy each have 24 missing values. Energy spans 34.02–157.30 kWh; median 91.83 kWh.
Inspect rows, units, missing values, duplicates and implausible ranges.
Explore the target distribution and feature relationships; correlation is not cause.
Check what is known before the decision. The end-of-day meter is deliberately leaked.
MODEL FAMILIES
First identify the task. Then choose a method.
These names refer to different learning setups, algorithms, architectures and model products. They are not stages on a single intelligence scale.
Numeric estimate or class probability; a useful transparent baseline.
Predict from nearby labelled examples; distance and scaling matter.
Find a separating margin; kernels can express nonlinear boundaries.
Feature splits and boosted trees; often strong for tabular data.
Group unlabelled examples; clusters require interpretation.
Gated recurrent architecture for ordered sequences; evaluate on later periods.
Foundation models for language and multimodal tasks; generated claims need checking.
LIVE EXAMPLE
Follow the evidence, not the score alone
The 600 rows below are fictional independent campus-energy scenarios. Every comparison shown here uses the same 120 validation rows; the held-out test remains closed.
Inspect the data fields and frozen experiment
Inputs assumed available before the day
- Temperature
- Expected occupancy
- Planned operating hours
- Floor area
Outputs and audit
- Observed energy (kWh)
- High-demand label: observed energy ≥ 125 kWh
- End-of-day meter: deliberately invalid as a before-day input
- Split: 360 train / 120 validation / 120 held-out test
RUNNING ML / NO ACCOUNT NEEDED
See the methods work
These small experiments compute locally in your browser on synthetic data. Change an input, inspect the result, and explain which part of the method made it happen.
01 · Parameter training02 · KNN classification03 · K-means training04 · Evaluation & leakage05 · Attention calculation
SUPERVISED / KNN
A new point borrows its neighbours’ labels
Click the plot to move the black query point. KNN finds the k nearest labelled observations and takes their majority vote.
Both axes use the same synthetic 0–100 scale. The vote share is an empirical neighbour count, not a calibrated probability.
Slate: Pattern A · Gold: Pattern B · Black cross: query · Lines: nearest neighbours
UNSUPERVISED / K-MEANS
Watch clusters form through iteration
Assign each point to its nearest centre, move the centres to their groups’ means, and repeat. No outcome labels are supplied.
Synthetic points and deterministic initial centres. Inertia is the sum of squared distances to the nearest centre. Useful groups still need domain interpretation.
Circles: observations · Crosses: centres · Colours: current assignments
FROM ML TO AGENTS
Capabilities are measured task by task
Official catalogues list GPT-6 Astra and Gemini 4 Argon as of 17 October 2026. A model score needs its benchmark, metric, test conditions and date. It is not a universal intelligence score.
A fluent answer can contain invented citations or false facts. Open the cited source and check the claim.
Goal → plan → tool → observe → revise, within permissions and stop rules. Human review remains essential where errors matter.
AI can assist with bounded tasks; people frame problems, verify evidence and take responsibility for deployment.
Current model references: OpenAI models · Google DeepMind model cards.
A PAPER BECOMES AN EXPERIMENT
How to visualise ML research
Start with the question and the evidence. Turn the paper’s mechanism into an inspectable computation, and keep the published claims separate from your own teaching example.
QuestionModel / equationDataBaselineExperimentFigures & limits
PRIMARY PAPER / 2017
Attention Is All You Need
Vaswani and colleagues introduced the Transformer architecture. Below, calculate one small attention operation using deliberately chosen vectors.
Read the original paper ↗Paper idea
Attention combines information using weights derived from query–key comparisons.
Our teaching exampleFour tokens and two-dimensional illustrative vectors. These are not learned GPT/Gemini weights or a reproduction of the paper’s translation results.
CHOOSE A QUERY TOKEN
weights = softmax(QKᵀ / √d)
All query-to-key weights
Inspect the inputs and calculation
Each token has an illustrative query Q, key K and value V, with d = 2. Compute the four dot products for the selected query, divide by √2, apply softmax so the weights sum to 1, then form the weighted sum of V.
The input vectors are supplied for explanation, not trained on text. Attention weights in this toy operation do not prove understanding or provide a complete explanation of a language model.
CONNECT TO THE DE EXAMPLE
Physics Informed Deep Learning
Raissi, Perdikaris and Karniadakis describe neural networks informed by differential equations. Use it to discuss how data and physical laws can enter learning.
Our cooling lab fits one parameter in a chosen equation. It is not a physics-informed neural network or a reproduction of this paper.
Read the authors’ paper ↗OUR RESEARCH WORKFLOW
GFIS research observatory
Open the public research status page, or use the protected observatory when presenting with an authorised account.
Public research status ↗Research observatory · sign-in required ↗Make one figure explain the paper. Write the question, identify the data and baseline, name the plotted variables and units, reproduce a result when code and rights allow, then show where your result agrees or differs. A new illustration alone is not a reproduction.
KEEP LEARNING
Take the example with you
Use these original teaching files for revision. The notebooks may run in Jupyter or Google Colab after the CSV is placed beside them.
Equation → prediction
Executed notebook with observations, parameter training, loss curve, held-out check and inference.
Download notebook ↗Lecture field guide
The flow, model-family comparison, worked example and links in one reading page.
Open the guide ↗Expanded learner notes
Cooling equation, EDA, model families, scores and agents.
Open notes PDF ↗Presentation
The guest lecture deck as a PDF for review.
View slide PDF ↗Download editable PowerPoint ↗Learner sheet
Cooling equation, data/EDA prompts, model map and evidence chain.
Download PDF ↗Regression notebook
Compare a simple baseline with a linear model.
Download notebook ↗Classification notebook
Read probabilities and high-demand labels.
Download notebook ↗Leakage investigation
Audit a suspicious feature and freeze a final choice.
Download notebook ↗Revision podcasts
Two existing explainers on supervised learning and evaluation.
Listen below ↓Supervised learning · 10 min 34 sec
Model evaluation · 18 min 22 sec
Continue in Workshop 01. The full learner workspace has six notebooks, a 600-row dataset, notes and a model-card exercise. Open Workshop 01 ↗
WHAT ML CAN DO
Inputs become useful outputs
Each application needs appropriate data, an honest test and someone responsible for acting on the result.
Predict a quantity
Sales history → demand estimate → replenishment planning.
Classify a case
Message content → routing label → support-team review.
Find similar groups
Usage features → clusters → investigate behaviour.
Flag an anomaly
Sensor history → unusual signal → technician inspection.
Forecast what follows
Ordered observations → later value → plan ahead.
Draft and assist
Verified context → generated draft or bounded tool action → human approval.
Predictions can fail when inputs change. Correlation does not establish cause; clusters need interpretation; generated claims need verification.
FROM CASE STUDY TO SYSTEMS
Explore the work behind the lecture
These public pages show research, prototypes and teaching systems at their stated maturity. A polished page does not replace field or clinical validation.
THE CATALOGUE
13 MindForgeAI projects
See the broader portfolio and choose a domain to examine.
Explore projects ↗AGRICULTUREApollo AgriIntel
Ask how an input becomes a useful agricultural decision.
Open project ↗EDGE SYSTEMSApollo ABS-100
An interactive reference assembly, pending physical validation.
Explore assembly ↗INDUSTRIAL RESEARCHGFIS Upgrade
A simulation and evidence workflow, not plant control.
Open lab ↗CYBERSECURITYAI Deception Lab
A bounded synthetic experiment, not a live honeypot.
Open lab ↗NEXT TOPICMLOps Academy
Versioning, experiment evidence, delivery and monitoring.
Explore course ↗Our platforms, ready to open
Public introductions and resources. Individual workspaces may require sign-in.
ORIGINAL UNIT DECKS / PUBLIC ACCESS
Your original five-unit library
The complete Gamma/NotebookLM college decks retain their original visual design. All ten files open directly from this website, without a Drive access request.
Private teaching cues are stored in the same restricted CEO faculty folder. Sign in with your authorised faculty account.
01ORIGINAL UNIT / 30 SLIDES
Introduction to Machine Learning
Learning types and the original foundation story.
Public PDF and editable PowerPoint
02ORIGINAL UNIT / 30 SLIDES
Data Preprocessing
Missing values, scaling, encoding and data splits.
Public PDF and editable PowerPoint
03ORIGINAL UNIT / 30 SLIDES
Feature Selection & Extraction
Feature selection, extraction, PCA and LDA.
Public PDF and editable PowerPoint
04ORIGINAL UNIT / 30 SLIDES
Supervised Learning Algorithms
Regression, classification, trees and evaluation.
Public PDF and editable PowerPoint
05ORIGINAL UNIT / 35 SLIDES
Unsupervised Learning Algorithms
Clustering, patterns and unsupervised methods.
Public PDF and editable PowerPointA QUESTION TO TAKE HOME
What would you build,
and how would you know it works?
Define the decision. Establish a baseline. Preserve an honest test. Explain where the system could fail.