The links of courses on 2021 and 2022 are not corresponding in the README. Someone should change it into the correct order.

Check out this blog for the learning rate decay that might be useful for your later experiments.

Link Or: https://medium.com/towards-data-science/learning-rate-schedules-and-adaptive-learning-rate-methods-for-deep-learning-2c8f433990d1

The intermediate steps in formulas (5) and (6) are missing the minus sign given by the derivation of w0 and w1.

Hi. The link between the tanh and the sigmoid functions is wrong, it should be 2*phi(2x)-1.

Since many students are asking for importing from different folders to reuse the code. Here is a useful link that can solve this problem.

https://stackoverflow.com/questions/4383571/importing-files-from-different-folder

Try the following:

sys.path.append('/path/to/application/app/folder')
import file

e.g.

sys.path.append('./scripts') # which means go to the scripts folder in the current folder
import helpers # import the helpers functions in the scripts folder

Hi All,

The note:
https://see.stanford.edu/materials/lsoeldsee263/08-min-norm.pdf

gives another view of least squares as to minimize the l2 norm of the unknown - weight vector which I have not seen before and is very interesting to me. Maybe some of you already have seen it anyway.

Best

Lecture 5b page 2 defines the logistic function as e^z / (1+e^z). This formulation is equivalent to 1/(1+e^(-z)), or 1 - 1/(1 + e^z).

Lecture 12a page 5 says that the sigmoid function 1/(1+e^(-x)) is one minus the logistic function, which is contradictory to lecture 5b.

Wikipedia's definition of the sigmoid function seems to say that the sigmoid function and logistic function often are referred to interchangeably. Therefore, I believe that saying that the sigmoid is one minus the logistic is wrong.

You say at the beginning of Exercise 5:

Reload the data through function load data() by setting sub sub_sample=False to keep only a few data
examples.
It should be sub_sample=True. It's a small typo.. But just in case..

Thanks @jmuth

In Task C, the domain of the assignment vector α is {0,1}, right?
I mean, in order to be aligned with the python code.
Thanks!

This refers to Problem Set 2, Exercise 1
In the question - 'What would X32 represent?' I think it should be X_(31) instead of X_(32). X_32 would not be a matrix element as per the notations used in the exercise.

Hello,
Just a quick typo on the MAP for ridge regression
In the equation 3 i guess the last term is p(\omega ] \lamba) not p(\omega ] \thetha)

Hi,

I think I've spotted a few typos in the lecture notes for lecture 6a on Generalized Mixture Models:

• Page 9, the probability of Gaussian distribution needs to be conditioned on both parameters , not just on mu . Alternatively, maybe a bold mu could be used to indicate the vector of usual parameters?
• Page 14, μ and σ² are not the natural parameters, they're the usual parameters. η is the natural parameter.
• Page 15, the log is missing a closing parenthesis, which should be right after h(y_n)

In Problem Set 7, in the definition of strictly convex function, I think it should be :

for all x,y in the domain of f such that x ≠ y

instead of

for all x,y in the domain of f

🤓

Currently, glove_solution.py computes a wrong value for ys, which is demonstrated by the following example code:

# example matrices parameter
xs = np.eye(5)
ys = np.eye(5)

# example indices parameter
ix = 0
jy = 0

# example scale parameter
scale = 1

# code from glove_solution.py
x, y = xs[ix, :], ys[jy, :]

xs[ix, :] += scale * y
ys[jy, :] += scale * x

print(ys)

Outputs

 [[ 3.  0.  0.  0.  0.]
 [ 0.  1.  0.  0.  0.]
 [ 0.  0.  1.  0.  0.]
 [ 0.  0.  0.  1.  0.]
 [ 0.  0.  0.  0.  1.]]

The problem with the ys value is that on the update it uses already modified variable x (modified by changing xs)

Possible solution: add data copying

Hi,

For the exercise 2 in lab03, the file plots.py is importing a file called build_polynomial. We are missing this file..

From what I've seen, we are using a function called build_poly in this file. Is it the build_poly function we wrote just before testing the polynomial regression?

Thanks

In the slidedeck 12b, the hand-written image Figure 1 could be written in LaTeX using this simple template: https://goo.gl/DZ2QTQ

Miss the "grid_search.py" file for 2nd assignment.
@glederrey

import numpy as np

def generate_w(num_intervals = 10):
    w0 = np.linspace(-100.0, 200.0, num_intervals)
    w1 = np.linspace(-150.0, 150.0, num_intervals)
    return (w0, w1)

def get_best_parameters(grid_w0, grid_w1, grid_losses):
    flat_index = np.argmin(grid_losses)
    index = np.unravel_index(flat_index, grid_losses.shape)

    return (grid_losses.flat[flat_index], grid_w0[index[0]], grid_w1[index[1]])

Would it be possible to add the solutions for exercise 5/6 in lab02?

Lecture5b, last page, ToDo 3: it should point to chapter 4 of JWHT book 😀

In order to get the same figure as in figure 1 in lab03
I think that lambdas in ex03.ipynb should be set
to np.logspace(-5, 1, 15)
instead of np.logspace(-5, 0, 15).
Am I right?

Hi,

Could you please do some clean-up in the repository history before next semester starts?

A fresh clone gives me:

--- /home/git/ML_course -------------------------
  758.9 MiB [##########] /.git                                                                                                                                                                            
   33.6 MiB [          ] /exam
    1.5 MiB [          ] /labs
    4.0 KiB [          ]  .gitignore
    4.0 KiB [          ]  README.md

You can use git-filter-branch or bfg for that purpose.

Our users (on noto.epfl.ch) have 253 copies (and counting) of this repository, using already 230 GB of disk and about 2 TB of s3 backup/archive space.

Thank you!