Course

Machine Learning: from a line of best fit to a foundation model

Name: Machine Learning: from a line of best fit to a foundation model
Price: 1799 BDT

Start by fitting a straight line and finish by predicting the next token, using the same idea the whole way. Every model here is small enough to check by hand, and every chapter ends where a real production system actually stands today.

73 lessons7h 51mEnglish

What you'll learn

Explain what it means for a model to be fitted to data rather than written by hand
Measure a model’s error as a single number, and describe how gradient descent reduces it
Say why a straight boundary fails on XOR, and what k-nearest neighbours does instead
Choose a classification threshold from the cost of each kind of mistake, not from accuracy alone
Test a model honestly: split before fitting, spot leakage, and say how sure a small test score is
Cluster unlabelled data with k-means, choose k and the feature scale, and justify the groups
Explain an embedding as a position, and read cosine and distance as measurements of it
Complete an analogy with vector arithmetic, and explain why direction carries meaning
Compute a next-token probability from counts, and explain temperature as reshaping a distribution
Describe how one autoregressive model can serve both language and recommendation

Requirements

Comfortable with coordinates, a straight line, and reading a graph
Able to work with a percentage and a simple average
No calculus, no linear algebra and no programming are assumed

Course content

73 lessons · 7h 51m

Learning from examples: how many cups tomorrow?

Uncle Rafi wants tomorrow's tea order from the weather forecast. Fit a line to Shumi's ledger, score it with MSE and MAE, train it by gradient descent, pick a safe learning rate, and see what one festival day does.

0/10

A rule learned from examples

Uncle Rafi makes 92 cups whatever the weather. Tania fits a line to ten days instead: about 2.87 fewer cups per degree.

Preview6m 2.

Turning wrongness into one number

Misses above and below the line cancel, so square them first. The 92-cup rule scores an MSE of 456 on three days.

Preview7m

Features, labels, fit and predict

Temperature goes in X, cups go in y. The scikit-learn words for the job, and the same line in six lines of Python.

Following the slope downhill

Gradient descent moves the line a small step at a time, always the way the error falls. One intercept step by hand.

When the step is too big

At a learning rate of 0.035 the error climbs every step; at 0.001 it crawls. Find the rate that settles fastest.

One bad day: MSE against MAE

Eid eve sells 140 cups at 30 °C. The slope drops from −2.87 to −1.71, and MSE jumps far more than MAE.

Tania's quick-fire round

Feature or label, which error punishes big misses, what fit does, and whether a learning rate is safe. Keep the streak.

Mission: the Eid week forecast

Fit the line, forecast an ordinary 25 °C day, then a 38 °C heat wave and Eid eve. Say which numbers to trust.

10m

Exam practice: learning from examples

Eight questions with the trap behind each wrong answer, then an MSE by hand, a slope to interpret and a claim to evaluate.

10m

10.

Learning from examples: quick revision

The chapter on one screen, gradient descent in order, one broken update to fix, and a spaced flashcard deck.

Did it really learn? Testing a model honestly

Dr Arif asks how Tania tested her forecast. Hold back a test set, split by date for a forecast, read over- and underfitting from two errors, hunt the leaks in Shumi's ledger, and say how sure a score from a few orders really is.

0/9

Testing on the exam you already saw

A score on the training days is not a test. Split Shumi's 24 orders: the 6 latest become the test week.

Preview7m 2.

Wiggly or stiff: over- and underfitting

Train error only falls as the curve gains bends. Test error dips to 0.44 at degree 2, then explodes to 26,641 at degree 9.

Preview8m

The leak: when the answer hides in the data

A 100% score is a warning. Find the copied order, the refund note, the repeat customer and the average fitted on every row: 100% falls to 75%.

How sure is eighty per cent?

A model that is right 80% of the time scores a perfect 4 of 4 about 41% of the time. The standard error says how far a small test can wander.

Five tests, not one: cross-validation

Cut the data into five folds, test on each once, and average. One fold scores 50%, another 100%; the mean is 80%.

Dr Arif's quick-fire round

Leak or not, overfit or underfit, which split, and what cross-validation averages. Keep the streak.

Mission: Dr Arif's viva

Tania defends her refund pipeline. Find the line that leaks, redo the average on training rows, and justify whether 75% on four orders is enough.

10m

Exam practice: did it really learn?

Eight questions with the trap behind each wrong answer, then choose a degree from two errors, name a leak and compute a cross-validation mean.

10m

Did it really learn? Quick revision

The chapter on one screen, an honest evaluation in order, and a spaced flashcard deck.

Late or on time? Drawing a line between classes

Some Doorstep orders reach customers late. Split late from on time with a perceptron, see why one line fails, let the nearest orders vote, scale km against taka, choose k honestly, and turn a score into a probability with logistic regression.

0/10

A boundary that learns

A perceptron nudges a line after every mistake. On sixteen Doorstep orders it never stops, ending at 14 of 16.

Preview7m 2.

When one line is not enough

Dr Arif's XOR puzzle: clean clusters, opposite corners matching, and no straight line that splits them.

Preview6m

Let the neighbours vote

k-nearest neighbours copies the most similar past orders. The 7 km order to Shyamoli is voted Late, 2 to 1.

Scale before you measure

In raw units taka swamps km, so a 6 km order is voted On time. Standardised, it is Late, 2 to 1.

Jagged or oversmoothed

k = 1 overfits, k = 15 calls every order On time, and k = 5 scores best held out: 13 of 16.

A probability, not a verdict

Logistic regression bends the line into an S-curve: a 17% chance of lateness at 3 km, 92% at 6 km.

Uncle Rafi's quick-fire round

Which model, does it need scaling, over- or underfitting, and which scikit-learn call. Keep the streak.

Mission: Friday's call-ahead list

Fit the S-curve, read P(Late) for five waiting orders, choose who the rider phones, and say how far to trust 14 of 16.

10m

Exam practice: drawing a line

Eight questions with the trap behind each wrong answer, then a sigmoid, two distances and a 100% claim to evaluate.

10m

10.

Drawing a line: quick revision

The chapter on one screen, a kNN prediction in order, one leaking scaler to fix, and a spaced flashcard deck.

Being wrong is a choice: which orders to call?

Fake cash-on-delivery orders are refused at the door. Turn the app’s fraud scores into a decision: count the four kinds of right and wrong, trade precision against recall, see why accuracy and a "90% accurate" flag mislead, check each area is treated fairly, and set the threshold in taka.

0/9

The threshold is a decision

The app scores each order from 0 to 10. Refused and honest scores overlap from 4.8 to 6.3, so every threshold makes some mistake.

Preview6m 2.

Counting the mistakes: the confusion matrix

At threshold 5: 5 refusals caught, 1 missed, 5 honest customers called, 31 cooked correctly. Build the matrix by hand, then read it in scikit-learn.

Preview7m

Precision against recall

Catch 5 of the 6 refusals with the fewest calls: threshold 6.2 gives precision and recall of 83%. Lower the line and precision pays for recall.

Why accuracy can lie

Never flagging anything scores 86%. A flag that catches 90% of refusals is right only 60% of the time, because refusals are 6 in 42.

Fair to every area

One threshold, two areas: Mirpur recall 66.7% with FPR 27.8%, Gulshan 100% and 0%. What a second threshold fixes, what it cannot, and who decides.

Shumi's quick-fire round

Which box, which metric, which denominator, and which mistake costs more. Metrics from counts at speed. Keep the streak.

Mission: set the threshold in taka

A false alarm costs ৳100 and a miss ৳300: the cheapest line is 6.2 at ৳400. Charge ৳500 for a miss and it drops to 4.8. Then say how sure that is.

10m

Exam practice: being wrong is a choice

Eight questions with the trap behind each wrong answer, then a defect detector at a Gazipur factory: matrix, precision, recall, accuracy and cost.

10m

Being wrong is a choice: quick revision

The chapter on one screen, the threshold choice in order, one broken recall to fix, and a spaced flashcard deck.

Corner Bakery's regulars: structure without labels

Uncle Rafi wants three offers for three kinds of regular, but nobody labelled the customers. Group them with k-means, watch it converge, beat a bad start, choose k with the elbow, standardise the features, and recommend from the nearest neighbours.

0/10

Grouping with no answers to check against

Fifteen regulars, two numbers each, no labels. k-means finds three groups, and Tania names them from their centroids.

Preview6m 2.

Assign, update, repeat until nothing moves

From poor starting rings, seven steps take the inertia from 16.91 to 0.85. One update worked by hand.

Preview7m

Where you start changes where you finish

Shumi's starting rings converge to inertia 9.98, a local minimum. n_init runs many starts and keeps the best.

How many groups? The elbow

Inertia always falls as k grows. The savings tell the story: 19.88, 9.27, then only 0.31. The elbow is at 3.

Taka against visits: scale changes the groups

On raw features spend decides everything: regulars merge with occasional buyers. Standardised, the three kinds appear exactly.

Customers like you

Should Sabbir get the tea card? His raw neighbours say no; his standardised neighbours, all regulars, say yes.

Uncle Rafi's quick-fire round

Labels or none, assign or update, stuck or best, which k and which scale. Keep the streak.

Mission: three offers for the board

Fill the elbow table raw and standardised, justify k and the scaling, name the three groups and design an offer for each.

10m

Exam practice: clustering

Eight questions with the trap behind each wrong answer, then two k-means steps by hand and a claim about k to evaluate.

10m

10.

Structure without labels: quick revision

The chapter on one screen, k-means in order, one flawed plan to fix, and a spaced flashcard deck.

Find me something like…: meaning as geometry

Customers ask the Doorstep app for "something like samosa". Place the menu on a map, measure likeness by distance and by cosine, turn a step into a meaning, build a semantic search, and pick a real model from its Hugging Face card.

0/9

Every item becomes a position

Two numbers per menu item, and the snacks, sweets, drinks and meals group themselves. Patty sits 0.5 from samosa.

Preview6m 2.

Two ways to measure similar

Distance compares position and cosine compares direction. Lentil rice’s nearest changes from butter rice to beef rice, and cos(milkshake, buttermilk) = 0.28.

Preview8m

Direction carries meaning

The step from milkshake to buttermilk means "make it savoury". Take it from doughnut and you land on samosa exactly.

Search by meaning

"A cold sweet drink" becomes one more point, and cosine ranks milkshake first. The same lookup is the retrieval half of RAG.

Choosing a model on the Hugging Face Hub

What a model card tells you, why Bangla requests need a multilingual model, and the rule that one model embeds both sides.

Tania's quick-fire round

Read a cosine, compute one by hand, land an analogy, and pick the right line of a model card. Keep the streak.

Mission: the Doorstep menu search

Uncle Rafi orders in Bangla. Design the search end to end: one model, stored menu vectors, cosine ranking, and a k you can defend.

10m

Exam practice: meaning as geometry

Eight questions with the trap behind each wrong answer, then a cosine and a distance by hand, two rankings to compare and a claim to evaluate.

10m

Meaning as geometry: quick revision

The chapter on one screen, the menu search in order, one broken cosine to fix, and a spaced flashcard deck.

Predicting the next thing: what will they order next?

Doorstep wants to suggest the next item in a Corner Bakery basket. Read baskets as sentences, predict the next item from counts, lengthen the context and back off, tune the temperature, see why a prompt steers a model, and finish the course with one mission across all seven chapters.

0/9

A basket is a sentence

Count what followed tea in twelve baskets and divide: samosa 6/11. The whole model, checkable by hand.

Preview7m 2.

How much of the basket to read

Two items of context stop the button suggesting a second samosa. The cost is evidence, and backoff covers what was never seen.

Preview7m

The temperature slider

Raise each probability to 1/T: at 0.5 samosa takes 36/49. Temperature reshapes the bars, never reorders them, and adds no knowledge.

Why a prompt changes anything

A prompt is the context. A milkshake basket and a tea basket share no suggestion, though the counts never changed.

What a real model changes

Lentil rice defeats the counts. A real model keeps the task and adds embeddings and long context; how is the Deep Learning course.

The suggestion button quick-fire

Probabilities from counts, backoff, temperature and prompts, one card at a time. Keep the streak.

Mission: Corner Bakery end to end

The whole course in one question: forecast, test, classify, set a threshold in taka, segment, search, and suggest the next item.

10m

Exam practice: predicting the next thing

Eight questions with the trap behind each wrong answer, then a canteen corpus: probabilities, temperature, backoff and a claim to evaluate.

10m

Predicting the next thing: quick revision

The chapter on one screen, the suggestion loop in order, one broken temperature line to fix, and a spaced flashcard deck.

About the creator

Classytic