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 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

🤓

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.

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]])

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!

The end of the readme for text-classification project was hard to read due to some text being recognized as code by github. I added spacing between existing code block and text to fix this.

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?

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

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.

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)