How it works

Read a short section, check that you understood it, practise in your browser, then prove it in a graded lab.

Reading-first chapters

Chapters are made of short sections written for the page: worked derivations, code you can run, and notes in the margin where a definition or a caution belongs.

You go at your own pace, and you can return to any section whenever you need it.

a · Read

Fitting a line

We want the line that sits closest to the data. Measure each miss vertically, square it, and add the squares up. The line with the smallest total is the least-squares fit.

Squaring makes every miss count, and big misses count most.

b · Check

Which line does least squares choose?

The line through the first and last points
The line with the smallest sum of squared misses Correct
The line with the most points exactly on it

c · Lab

In [3]:

X = np.c_[np.ones(len(x)), x]
w, *_ = np.linalg.lstsq(X, y)
print(w.round(2))   # runs on a CPU

[1.98 0.51]

Lab 1 · Passed

Fig. 1A section, its check, and a lab. Illustration.

A check after every section

Each section ends with a few questions. They’re there so you know you’ve understood before you move on.

Check · 1 of 3

A model scores 99% on its training data and 70% on new data. What is the most likely problem?

Underfitting Try again
Overfitting
Too little training data to load

A high training score rules out underfitting. The gap between training and new data is the sign to look for.

Fig. 2A check, with a wrong answer circled in red pencil and the correction in a note. Illustration.

Exercises in your browser

Short exercises run in the page as you read, so you practise each idea straight away, with nothing to install.

Exercise · 2 of 4

Return the mean of each column of X.

def column_means(X):
    return X.mean(axis=0)
column_means(np.eye(2))

Run

array([0.5, 0.5])

Correct

Fig. 3An exercise in the page: write the code, run it, and see the result checked. Illustration.

Graded labs on your own lab machine

Every learner gets a private cloud lab machine with a notebook and a terminal. Each subject has graded labs that apply its chapters to a real or synthetic dataset. Labs are designed for a CPU, so you need no GPU.

your lab machine

In [1]:

import numpy as np
data = np.loadtxt("data/train.csv", delimiter=",")

In [2]:

x, y = data[:, 0], data[:, 1]
x.shape

(200,)

In [ ]:

def fit(x, y):

Lab 1

Fit the line, then submit.

Write fit()
Plot the residuals
Answer the report questions

Submit for grading

learner@lab:~$ ls data
test.csv  train.csv
learner@lab:~$

Fig. 4The lab workspace: a notebook and a terminal on your own machine. Illustration.

How grading works

Labs are graded automatically by tests you don’t see, so passing means your code really works.

The capstone is graded in five milestones, ending in a report, a runnable demo and a viva.

your notebook

run on submit

tests you don’t see

Passed

Fig. 5Grading: on submit, your notebook is run against tests you don’t see. Illustration.

Credits and time

1 credit = 30 hours of learner effort.

Credits by semester, on one track
Semester	Credits	Hours
Semester 0: Bridge	–	45 h
Semester 1: Foundations	14	420 h
Semester 2: Deep learning core	12	360 h
Semester 3: Domain track	8	240 h
Semester 4: Domain track and capstone	16	480 h

S045 h

S1420 h

S2360 h

S3240 h

S4480 h

Fig. 6Hours per semester on one track: the table above, as bars. Illustration.

At 10 hours a week, a 4-credit subject takes about 12 weeks. Go faster or slower as your time allows.

What we don’t do

No videos. No live classes. No AI tutor.

We leave these out on purpose. The time goes into the writing, the checks and the labs.

See the programme Questions