Deadline: January 10th, 9am
NOTE The deadline will likely be moved to January!
As usual, there’s a notebook on E-Learning with some starter code.
In this assignment, we want to explore self-supervised methods for extracting features from unlabeled data, and then use those features for supervised tasks.
No matter the exact kind of model, we usually do something like this:
For a dataset of your choice, implement the above pipeline. Try at least two
different kinds of self-supervised models; for each, train the model and then
use the features for a classification task.
Also train a model directly on classification (no pre-training) and compare the
performance to the self-supervised models. Also compare the different self-supervision
methods with each other.
To make these comparisons fair, your models should have the same number of parameters. E.g. you might want to use the same “encoder” architecture for each task, and add a small classification head on top; then, the network that you train directly on classification should have the same architecture as the encoder and the classification head combined.
The remainder of this text discusses some issues to keep in mind when building autoencoders or similar models.
Building autencoders in Tensorflow is pretty simple. You need to define an encoding based on the input, a decoding based on the encoding, and a loss function that measures the distance between decoding and input. An obvious choice may be simply the mean squared error (but see below). To start off, you could try simple MLPs. Note that you are in no way obligated to choose the “reverse” encoder architecture for your encoder; e.g. you could use a 10-layer MLP as an encoder and a single layer as a decoder if you wanted. As a start, you should opt for an “undercomplete” architecture where the encoding is smaller than the data.
Note: The activation function of the last decoder layer is very important, as it needs to be able to map the input data range. Having data in the range [0, 1] allows you to use a sigmoid output activation, for example. Experiment with different activations such as sigmoid or linear (i.e. no) activation and see how it affects the model (do not use relu in the output layer!). Your loss function should also “fit” the output function, e.g. a sigmoid output layer goes well with a binary (!) cross-entropy loss.
Note that you can use the Keras model APIs to build the encoder and decoder
as different models, which makes it easy to later use the encoder separately.
You can also have sub-models/layers participate in different models at the same time,
e.g. an encoder model can be part of an autoencoder model together with a decoder,
and of a classification model together with a classifer_head.
Next, you should switch to a convolutional encoder/decoder to make use of the
fact that we are working with image data. The encoding should simply be one or
more convolutional layers, with any filter size and number of filters, plus
downsampling (strided convolutions or pooling). You can also
optionally apply fully-connected layers at the end. As an “inverse” of a
Conv2D, Conv2DTranspose is commonly used. However,
you could also use UpSampling2D along with regular convolutions.
Again, there is no
requirement to make the parameters of encoder and decoder “fit”, e.g. you don’t
need to use the same filter sizes. However, you need to take care when choosing
padding/strides such that the output has the same dimensions as the input. This
can be a problem with MNIST (see the notebook).
It also means that the last
convolutional (transpose) layer should have as many filters as the input
space (e.g. one filter for MNIST or three for CIFAR).
Even other self-supervised models are often similar to autoencoders.
For example, in a denoising
autoencoder, the input is a noisy version of the target (so input and target are not the same
anymore!), and the loss is computed between the output and this “clean” target.
The architecture can remain the same, however.
Similarly, if the input has parts of the image removed and the task is to
reconstruct those parts, the target is once again the full image, but an autoencoding
architecture would in principle be appropriate once again.
Say, you trained some encoder network on a self-supervised task and now build a classification head on top for labeled data. Now you want to train this model. But which parts do you actually train? It could be
There is a trade-off here:
Experiment! You can easily “freeze” layers or whole models by setting their
trainable argument to False.
Self-supervised models are useful in that they can learn from unlabeled data. This can significantly improve performance in settings where large amounts of data are available, but few labels. We can artificially evoke such a situation by just “pretending” that parts of our data has no labels. Try this:
As before, compare to a model that is trained directly on the classification task, but only on the labeled subset. If everything works as expected, your self-supervised model should outperform the directly trained one (on the test set)! This is because the direct training massively overfits on the small dataset, whereas the self-supervised model was able to learn features on all available data. You will most likely want to freeze the encoder model, i.e. not fine-tune it – if you did, the self-supervised model would overfit, as well.