# Deep learning

The present tutorials covers the basic implementation towards deep learning. Most of this young research field is quite extensively based on the neural networks (that we implemented in a previous tutorial). Therefore, a good knowledge of the previous implementations is required. We will first construct an auto-encoder to perform an unsupervised learning directly from any data. Then, we will see how we can transfer this knowledge to a more extensive supervised classifier. Finally, we will extend these implementations with the use and visualization of convolutional filters. This tutorial is an adaptation of the UFLDL tutorials for audio data.

# Reference slides

The corresponding slides cover

• Deep learning
• Applications

# Tutorial

In this tutorial, we will use the self-taught learning paradigm with a sparse autoencoder in order to build an unsupervised audition system. Then, we will see how to transfer this learning and also rely on the softmax classifier (implemented in the neural networks tutorial) to build a classifier for spectral windows.

First, you will train your sparse autoencoder on an unlabeled training dataset of audio data (in this case, you can use all the datasets from the tutorials, and even add your own audio files), which will be represented as a set of concatenated spectral windows (to keep the temporal information).

Then, we will extract these learned features from the weights of the network trained in an unsupervised fashion, in order to apply this knowledge to a labeled dataset of audio files (by using these features as inputs to a softmax classifier). Finally, we will extend this classifier by using convolutional filters.

## 8.1 - Sparse Auto-Encoders

The autoencoder aims at learning both an encoding function $e\left(\mathbf{x}\right)$ and a decoding function $d\left(\mathbf{x}\right)$ such that ${\textstyle d\left(e\left(\mathbf{x}\right)\right)=\tilde{\mathbf{x}}\approx\mathbf{x}}$. Therefore, the AE is intended to learn a function approximating the identity function, by being able to reconstruct an approximation ${\textstyle\tilde{\mathbf{x}}}$ similar to the input $\mathbf{x}$ via a hidden representation. In the case of entirely random input data (for instance a set of IID Gaussian noise), this task would not only be hard but also quite meaningless. However, based on the assumption that there might exist an underlying hidden structure in the data (where part of the input features are consistently correlated), then this approach might be able to uncover and exploit these statistical regularities. The encoding function $e:\mathbb{R}^{d_{x}}\rightarrow\mathbb{R}^{d_{h}}$ maps an input $\mathbf{x}\in\mathbb{R}^{d_{x}}$ to an hidden representation $\mathbf{h}_{\mathbf{x}}\in\mathbb{R}^{d_{h}}$ by producing a deterministic mapping

where $s_{e}$ is a nonlinear activation function (usually the sigmoid function), $\mathbf{W}_{e}$ is a $d_{h}\times d_{x}$ weight matrix, and $\mathbf{b}_{e}\in\mathbb{R}^{d_{h}}$ is a bias vector.

The decoding function $d:\mathbb{R}^{d_{h}}\rightarrow\mathbb{R}^{d_{x}}$ then maps back this encoded representation $\mathbf{h_{x}}$ into a reconstruction $\mathbf{y}$ of the same dimensionnality as $\mathbf{x}$

where $s_{d}$ is the activation function of the decoder. Usually the weight matrix of the decoding layer $\mathbf{W}_{d}$ is tied to be the transpose of the encoder weight matrix $\mathbf{W}_{d}=\mathbf{W}_{e}^{T}$, in which case the AE is said to have tied weights.

Hence, training an auto-encoder can be summarized as finding the optimal set of parameters $\theta=\left\{\mathbf{W}_{e},\mathbf{W}_{d},\mathbf{b}_{e},\mathbf{b}_{d}\right\}$ (or $\theta=\left\{ \mathbf{W},\mathbf{b}\right\}$ in the case of tied weights) in order to minimize the reconstruction error on a dataset of training examples $\mathcal{D}_{n}$

Usual choices for the reconstruction error function $\mathcal{L}$ are either the squared error $\mathcal{L}(x,y)=\left\Vert x-y\right\Vert ^{2}$ (often used for linear reconstruction) or the cross-entropy loss of the reconstruction $\mathcal{L}(x,y)=-\sum_{i=1}^{d_{x}}x_{i}log\left(y_{i}\right)+\left(1-x_{i}\right)log\left(1-y_{i}\right)$ (if the input is interpreted as vectors of probabilities and a sigmoid activation function is used).

As can be seen from the definition of the objective functions, by solely minimizing the reconstruction error, nothing prevents an auto-encoder with an input of $n$ dimensions and an encoding of the same (or higher) dimensionnality to simply learn the identity function. In this case, the AE would merely be mapping an input to a copy of itself. Surprisingly, it has been shown that non-linear autoencoders in this over-complete setting (with a hidden dimensionality strongly superior to that of the input) trained with stochastic gradient descent, could still provide useful representations, even without any additional constraints

We can see that the framework defined by AEs fit the overarching goal of unsupervised and self-taught learning, as it tries to exploit statistical correlations of the data structure to find a non-linear representation aimed at decomposing and then reconstructing the input. In the starting code, we provide the basic functions to perform this learning.

File Explanation
base-toolboxes Toolboxes to perform learning
display_network.m Network visualization function
initializeParameters.m Random initialization of network weights
params2stack.m Transform parameters as a stack
pcaWhitening.m Perform PCA whitening on input data
sampleSpectrums.m Sample from a set of spectrums
sparseAutoencoderCost.m Cost and gradient function for an SAE
stack2params.m Transform stack to a parameter
stackedAECost.m Cost and gradient for multiple layers
stackedAEPredict.m Prediction function for logistic regression

In order to perform the various computations, we will use the minFunc package as an advanced optimizer and the unlabeled data (any audio files) to train a sparse autoencoder. However, we first need to prepare an input set, so that the given data is formatted to a common size (slices of audio input). In the following problem, you will implement the sparse autoencoder algorithm, and show how it discovers an optimal representation for spectral windows.

Specifically, in this exercise you will implement a sparse autoencoder, trained with 4 consecutive spectral distributions (FFT, Mel, Bark or Cepstrum) using the L-BFGS optimization algorithm (this algorithm is provided in the minFunc subdirectory, which is a 3rd party CCA software).

Inputs

The first step is to generate a training set. To get a single training example $x$, we need to compute the spectral transform from a sound and then subsample a set of a given number of consecutive spectral frames. This will allow the network to learn from the complete spectro-temporal information. However, the sampled parts will need to be converted into vectors.

Exercise

1. Update the sampleSpectrums function to generate examples
2. Test your function on different transforms to observe the outputs
3. Check the random intialization of parameters

Expected output []

The learning of an autoencoder is based on a cost function that tries to reconstruct the input from a combination of hidden units. Hence, as underlined in the previous question, we will start by implementing the simplest reconstruction cost, given by

However, as discussed previously, solely minimizing the reconstruction error, nothing prevents an auto-encoder with an input of $n$ dimensions and an encoding of the same (or higher) dimensionnality to simply learn the identity function. One way to avoid this situation is to enforce a sparsity constraint on the learning. The idea is to force the network to make this reconstruction from fewer data. If we denote $a^{h}_{j}\left(x\right)$ as the activation of hidden unit $j$ in the hidden layer of the autoencoder when given a specific input $x$. Furthermore, let

be the average activation of hidden unit $j$ (averaged over the training set). We would like to (approximately) enforce the constraint

where $\rho$ is a given sparsity parameter, typically a small value close to zero. In other words, we would like the average activation of each hidden neuron $j$ to be small. To satisfy this constraint, the hidden unit’s activations must mostly be near 0. To achieve this, we will add an extra penalty term to our optimization objective that penalizes $\hat\rho_j$ deviating significantly from $\rho$. Many choices of the penalty term will give reasonable results. We will choose the following

Here, $n_{h}$ is the number of neurons in the hidden layer, and the index $j$ is summing over the hidden units in our network. If you are familiar with the concept of KL divergence, this penalty term is based on it, and can also be written

where $KL(\rho \mid \mid \hat\rho_j) = \rho \log \frac{\rho}{\hat\rho_j} + (1-\rho) \log \frac{1-\rho}{1-\hat\rho_j}$ is the Kullback-Leibler (KL) divergence

Hence, we can simply define the cost function $\mathcal{J}_{sparse}\left(\theta\right)$ and the corresponding derivatives of $\mathcal{J}_{sparse}$ with respect to the different parameters as the original cost with the added constraint

In order to test the validity of your implementation, you can use the method of gradient checking, which allows you to verify that your numerically evaluated gradient is very close to the true (analytically computed) gradient.

Implementation tip: If you are debugging your code, perform the learning on smaller models and smaller training sets (using few training examples and hidden units) to speed things up.

Exercise

1. Update the sparseAutoencoderCost to perform the simple Euclidean cost.

1. Test your algorithm and visualize the learned filters.
2. Update the sparseAutoencoderCost to include the sparsity constraint.
3. Test your algorithm and compare the results.
4. Also add the weight decay to further regularize learning.

## 8.2 - Training and visualizing

Once you have coded and verified your objective and derivatives, you can train the parameters of the model and use it to extract features from the spectral windows. Equiped with the code that computes $\mathcal{J}_{sparse}$ and its derivatives, we will minimize $\mathcal{J}_{sparse}$ by using the L-BFGS algorithm, provided in the minFunc package (code provided by Mark Schmidt) included in the starter code. (For the purpose of this assignment, you can use its default parameters). The minFunc code assumes that the parameters to be optimized are a long parameter vector, so we will use the $\theta$ parameterization rather than passing each parameter separately.

In the starter code, we have provided a function for initializing the parameters. We initialize the biases $b^{h}_i$ to zero, and the weights $W^{h}_{ij}$ to random numbers drawn uniformly from the interval $\left[-\sqrt{\frac{6}{n_{\rm in}+n_{\rm out}+1}},\sqrt{\frac{6}{n_{\rm in}+n_{\rm out}+1}}\,\right]$, where $n_{in}$ is the fan-in (the number of inputs feeding into a node) and $n_{out}$ is the fan-out (the number of units that a node feeds into).

Visualization
After training the autoencoder, you can use display_network to visualize the learned weights.

Exercise

1. Run your full implementation on a reduced dataset.
2. Perform the visualization in order to see the learned weights.
3. Compare the results with using the full dataset.

## 8.3 - Logistic regression

As we have seen in the Neural Networks tutorial, in order to perform multi-class predictions, we cannot rely on simply computing the distance between desired patterns and the obtained binary value. The idea here is to rely on the softmax regression, by considering classes as a vector of probabilities. If you have implemented the logistic regression in the previous tutorial, you can simply copy and paste your code here. Otherwise, we recall the mathematical bases behind these computations.

The desired answers will be considered as a set of probabilities, where the desired class is $1$ and the others are $0$ (called one-hot representation). Then, the cost function will rely on the softmax formulation

Therefore, we compute the output of the softmax by taking

By taking derivatives, we can show that the gradient of the softmax layer is

In softmaxCost, implement the softmax cost function $J(\theta)$. Remember to include the weight decay term in the cost as well. Your code should also compute the appropriate gradients, as well as the predictions for the input data (which will be used in the cross-validation step later).

Implementation Tip
Computing the ground truth matrix - In your code, you may need to compute the ground truth matrix $M$, such that $M(r, c)$ is $1$ if $y(c) = r$ and $0$ otherwise. This can be done quickly, without a loop, using the MATLAB functions sparse and full. Specifically, the command M = sparse(r, c, v) creates a sparse matrix such that $M(r(i), c(i)) = v(i)$ for all i. This code for using sparse and full to compute the ground truth matrix is already provided in softmaxCost.m.

Implementation tip: Preventing overflows
In the softmax regression, we compute the unbounded hypothesis of the input belonging to each class, which can lead to overflow. However, given the definition of the logistic function, the overall (relative) probabilities remain equivalent if we substract the same quantity from each of the $\theta_j^T x^{(i)}$. Hence, to prevent overflow, we shall simply subtract some large constant value from each of the $\theta_j^T x^{(i)}$ terms before computing the exponential.

Exercise

• Update softmaxCost to compute the softmax cost
• Train the softmax model by using softmaxTrain and the L-BFGS algorithm.
• Evaluate the accuracy of training the softmax on different architectures.

## 8.4 - Transfer learning

Up to now, we have seen how to train a single AE, which will basically learn one abstraction over the raw input data. Hence, we would like to perform learning of several AEs, each time by feeding the output of the previous one, in order to learn higher-level abstractions. The greedy layerwise approach for pretraining a deep network works by training each layer in turn. We will see autoencoders can be “stacked” in a greedy layerwise fashion for pretraining (initializing) the weights of a deep network.

A stacked autoencoder is simply a neural network consisting of multiple layers of sparse autoencoders in which the outputs of each layer is wired to the inputs of the successive layer. Formally, consider a stacked autoencoder with $n$ layers. The encoding step for the stacked autoencoder is given by running the encoding step of each layer in forward order

The decoding step is given by running the decoding stack of each autoencoder in reverse order

The information of interest is contained within $a^{(n)}$, which is the activation of the deepest layer of hidden units. This vector gives us a representation of the input in terms of higher-order features. The features from the stacked autoencoder can be used for classification problems by feeding $a^{(n)}$ to a softmax classifier.

To learn this architecture, we first train the first layer on the raw input. Then, we use the first layer to transform the raw input into a vector consisting of activation of the hidden units (by performing a forward pass), The second second layer is then trained on this vector and we repeat the operation for any subsequent layers, using the output of each layer as input for the subsequent layer. Finally, to produce better results, after this phase of training is complete, we fine-tune the whole network using backpropagation by tuning the parameters of all layers at the same time.

Concretely, for each example in the the labeled training dataset, we forward propagate the example to obtain the activation of the hidden units. This transformed representation is used as the new feature representation with which to train the softmax classifier. Therefore, in the end we will use a single network composed at each layer of the encoding function of each AEs, therefore this can be seen as a way to transfer the knowledge on features learned by each AE.

Exercise

1. Setup a target architecture.
2. Train each autoencoder separately.
3. Perform the weight transfer into a single network.
4. Fine-tune this network to obtain a complete classifier.
5. Evaluate your classifier on audio data.

## 8.4 - Convolutional layers

We have seen the auto-encoder which consider a non-linear transform of its input data. This can be seen as a fully-connected network. However, this implicitly requires a very large number of parameters and also does not consider that most features will essentially focus on local information given that most data have a property of stationarity. A very successful type of transform used in deep learning is convolutional layer. These restrict the connections between hidden and input units, allowing each hidden unit to connect to only a small subset of the input units. Specifically, each hidden unit will connect to only a small contiguous region of pixels in the input.

Convolution The property of stationary data means that the statistics of one part of the data are the same as any other part. This suggests that the features that we learn at one part can also be applied to other parts, and we can use the same features at all locations. More precisely, having learned features over small (say 8x8) patches sampled randomly from the larger set, we can then apply this learned 8x8 feature detector anywhere. Specifically, we can take the learned 8x8 features and convolve them with the larger image, thus obtaining a different feature activation value at each location in the input matrix.

Formally, given some large $r \times c$ input matrix, we first train a sparse autoencoder on small $a \times b$ patches sampled from these inputs, learning $k$ features $f = \sigma\left(Wx_{small} + b\right)$ (where $\sigma$ is the sigmoid function), given by the weights $W$ and biases $b$ from the visible units to the hidden units. For every $a \times b$ patch $x_{s}$ in the large image, we compute $f_{s} = \sigma\left(Wx_{small} + b\right)$, giving us $f_{convolved}$, a $k \times (r - a + 1) \times (c - b + 1)$ array of convolved features.

Implementing convolution First, we want to compute $\sigma\left(W(r,c) + b\right)$ for all valid $(r,c)$ (valid meaning that the entire 8x8 patch is contained within the input data), where $W$ and $b$ are the learned weights and biases from the input layer to the hidden layer, and $x(r,c)$ is the 8x8 patch with the upper left corner at $(r,c)$. To accomplish this, one naive method is to loop over all such patches and compute $\sigma\left(W(r,c) + b\right)$ for each of them; while this is fine in theory, it can very slow. Hence, we usually use Matlab’s built in convolution functions, which are well optimized.

Observe that the convolution above can be broken down into the following three small steps. First, compute $Wx(r,c)$ for all $(r,c)$. Next, add $b$ to all the computed values. Finally, apply the sigmoid function to the resulting values. You can replace the loop in the first step with one of MATLAB’s optimized convolution functions, conv2, speeding up the process significantly.

It should be noted that conv2 performs a 2-D convolution, but we have 4 “dimensions” - input number, feature number, row, column of image, and (color) channel of image - that you want to convolve over. Because of this, you will have to convolve each feature separately for each input, using the row and column as the 2 dimensions you convolve over. Second, because of the mathematical definition of convolution, the feature matrix must be “flipped” before passing it to conv2. The following implementation tip explains the “flipping” of feature matrices when using MATLAB’s convolution functions (using both

Pooling After obtaining features using convolution, we would next like to use them for classification. In theory, one could use all the extracted features with a classifier such as a softmax classifier, but this can be computationally challenging. Furthermore, learning a classifier with inputs having 3+ million features can be unwieldy, and can also be prone to over-fitting. Thus, to describe a large input, one natural approach is to aggregate statistics of these features at various locations. For example, one could compute the mean (or max) value of a particular feature over a region of the image. These summary statistics are much lower in dimension (compared to using all of the extracted features) and can also improve results (less over-fitting). This aggregation operation is called pooling, or sometimes mean pooling or max pooling (depending on the operation).

Exercise

1. Implement the convolution operator
2. Derive the convolution and pooling operations
3. Enhance the auto-encoder with a convolutional layer
4. Train a multi-layer convolutional AE to perform classification
5. Visualize the corresponding filters