generated from morea-framework/morea
-
Notifications
You must be signed in to change notification settings - Fork 0
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
- Loading branch information
Showing
5 changed files
with
76 additions
and
7 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,40 @@ | ||
--- | ||
title: "Autoencoders" | ||
published: true | ||
morea_id: experience-auto | ||
morea_type: experience | ||
morea_summary: "Autoencoders and SVD" | ||
morea_start_date: "2021-07-15T23:00" | ||
morea_labels: | ||
--- | ||
|
||
First review the section about eigenvalues and eigenspaces | ||
[here]. Recall that \({\mathbb R}^p\) represents the linear space of | ||
all vectors with \(p\) real coordinates. For a matrix \(n\times p\) | ||
matrix \(X\), one can use spectral decomposition of \(X^TX\) | ||
(respectively \(XX^T\)) to find an orthonormal basis for \({\mathbb | ||
R}^p\) (respectively \({\mathbb R}^n\)) using eigenvectors of \(X^TX\) | ||
(respectively \(XX^T\)), and therefore for the rows (respectively | ||
columns) of \(X\). Assume that \(n \ge p\), and let the eigenvalues | ||
of \(X^TX\) be \(\lambda_1\ge \lambda_2 \cdots \ge \lambda_p\), then | ||
the highest \(p\) eigenvalues of \(XX^T\) are also | ||
\(\lambda_1, \lambda_2 \upto \lambda_p\), while the remaining \(n-p\) | ||
eigenvalues are all 0. | ||
|
||
|
||
Let \(V\) (respectively \(U\)) be the matrix formed | ||
by placing as columns the orthonormal basis obtained by the | ||
eigendecomposition of \(X^TX\) (respectively \(XX^T\)). | ||
Let \(\Sigma\) | ||
be the \(n\times p\) matrix formed with the positive square roots of | ||
the eigenvalues of \(X^TX\) in all the diagonal locations. | ||
The singular value decomposition observes that | ||
$$ X = U \Sigma V^T. $$ | ||
|
||
The following | ||
[notebook](https://uhm-descartes.github.io/ee445/morea/neural-networks/autoencoder.ipynb) | ||
shows you how to build a neural network that recovers the singular | ||
value decomposition, the autoencoder. In the assessment, you will | ||
reuse the code, but with non-linear activations to project the MNIST | ||
dataset into a low dimensional manifold. | ||
|
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1 @@ | ||
[email protected]:1683013381 |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -1,14 +1,28 @@ | ||
--- | ||
title: "CHANGE ME" | ||
published: false | ||
morea_id: assessment-CHANGE-ME | ||
morea_summary: "CHANGE ME" | ||
title: "Assessment" | ||
published: true | ||
morea_id: assessment-nn | ||
morea_summary: "Neural Networks" | ||
morea_outcomes_assessed: | ||
# - outcome-CHANGE-ME | ||
morea_type: assessment | ||
morea_start_date: "2021-07-16T09:00" | ||
morea_labels: | ||
--- | ||
# CHANGE ME | ||
# Problems | ||
|
||
TBD | ||
* Use a feedforward architecture to train and predict on the CIFAR-10 | ||
and Fashion-MNIST dataset. Here, you may need to use dropout to | ||
train better and reduce overfitting---find out how to implement this | ||
technique. We discussed dropout very briefly in class, so you may want | ||
to look up Dropout techniques online for more background. | ||
|
||
* Project the MNIST dataset into as small a manifold as | ||
possible. Meaning, you should come up with two transformations (we | ||
will call them encoder and decoder). The encoder should represent | ||
each \(28\times 28\) test image into a small vector (you can have | ||
this vector have less than 10 coordinates, but it is ok if your | ||
output is slightly larger too), but that doesn't lose | ||
information---namely the decoder can reconstruct the original image | ||
(with negligible loss) from the small vector output from the | ||
encoder. |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,14 @@ | ||
--- | ||
title: "CHANGE ME" | ||
published: false | ||
morea_id: assessment-CHANGE-ME | ||
morea_summary: "CHANGE ME" | ||
morea_outcomes_assessed: | ||
# - outcome-CHANGE-ME | ||
morea_type: assessment | ||
morea_start_date: "2021-07-16T09:00" | ||
morea_labels: | ||
--- | ||
# CHANGE ME | ||
|
||
TBD |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters