Pereyra et al. propose an entropy regularizer for penalizing over-confident predictions of deep neural networks. Specifically, given the predicted distribution $p_\theta(y_i|x)$ for labels $y_i$ and network parameters $\theta$, a regularizer

$-\beta \max(0, \Gamma – H(p_\theta(y|x))$

is added to the learning objective. Here, $H$ denotes the entropy and $\beta$, $\Gamma$ are hyper-parameters allowing to weight and limit the regularizers influence. In experiments, this regularizer showed slightly improved performance on MNIST and Cifar-10.

Also find this summary at [davidstutz.de](https://davidstutz.de/category/reading/).

## Task
Add '**rejection**' output to an existing classification model with softmax layer.

## Method
1. Choose some threshold $\delta$ and temperature $T$
2. Add a perturbation to the input x (eq 2),
let $\tilde x = x - \epsilon \text{sign}(-\nabla_x \log S_{\hat y}(x;T))$
3. If $p(\tilde x;T)\le \delta$, rejects
4. If not, return the output of the original classifier

$p(\tilde x;T)$ is the max prob with temperature scailing for input $\tilde x$

$\delta$ and $T$ are manually chosen.

Madry et al. provide an interpretation of training on adversarial examples as sattle-point (i.e. min-max) problem. Based on this formulation, they conduct several experiments on MNIST and CIFAR-10 supporting the following conclusions:
- Projected gradient descent might be “strongest” adversary using first-order information. Here, gradient descent is used to maximize the loss of the classifier directly while always projecting onto the set of “allowed” perturbations (e.g. within an $\epsilon$-ball around the samples). This observation is based on a large number of random restarts used for projected gradient descent. Regarding the number of restarts, the authors also note that an adversary should be bounded regarding the computation resources – similar to polynomially bounded adversaries in cryptography.
- Network capacity plays an important role in training robust neural networks using the min-max formulation (i.e. using adversarial training). In particular, the authors suggest that increased capacity is needed to fit/learn adversarial examples without overfitting. Additionally, increased capacity (in combination with a strong adversary) decreases transferability of adversarial examples.

Also view this summary at [davidstutz.de](https://davidstutz.de/category/reading/).

Very efficient data augmentation method. Linear-interpolate training set x and y randomly at every epoch.
```python
for (x1, y1), (x2, y2) in zip(loader1, loader2):
    lam = numpy.random.beta(alpha, alpha)
    x = Variable(lam * x1 + (1. - lam) * x2)
    y = Variable(lam * y1 + (1. - lam) * y2)
    optimizer.zero_grad()
    loss(net(x), y).backward()
    optimizer.step()
```
- ERM (Empirical Risk Minimization) is $\alpha = 0$ version of mixup, i.e. not using mixup.
- Reduces the memorization of corrupt labels.
- Increases robustness to adversarial examples.
- Stabilizes the training of GAN.

The model learns to translate using a seq2seq model, an autoencoder objective, and an adversarial objective for language identification.
The system is trained to correct noisy versions of its own output and iteratively improves performance.
Does not require parallel corpora, but relies on a separate method for inducing a parallel dictionary that bootstraps the translation.

https://i.imgur.com/6uXNAgo.png

Inducing word translations using only monolingual corpora for two languages. Separate embeddings are trained for each language and a mapping is learned though an adversarial objective, along with an orthogonality constraint on the most frequent words. A strategy for an unsupervised stopping criterion is also proposed.

https://i.imgur.com/HmME09P.png

* Semi-supervised method
* There is a teacher net and a student net, with identical architecture.
* The teacher makes predictions on unlabeled data, which are used as ground-truth for training the student net.
* After each gradient descent update on the student, the teacher's weights are updated so that it becomes an exponential moving average of the weights of the student at previous timesteps. It's called a "mean teacher" because of this moving average.

# Really Short

A method for training a recurrent network that avoids doing backpropagation through time by instead approximately forward-propagating the derivatives of the recurrent-state variables with respect to the parameters.

# Short

This paper deals with learning **Online** setting, where we have an infinite sequence of points $<(x_t, y_t): t \in \mathbb N>$, where at each time $t$ we would like to predict $y_t$ given $x_1, ... x_t$.  We'd like do this by training a **Recurrent** model: $o_t , s_t := f_\theta(x_t, s_{t-1})$ to **Optimize** parameters $\theta$ such that we minimize the error of the next prediction: $\mathcal L_t := \ell(o_t, y_t)$

The standard way to do this is is Truncated Backpropagation through Time (TBPTT).  We "unroll" the network for T steps (the truncation window), every T steps, and update $\theta$ to minimize $\sum_{\tau=t-T+1}^t \mathcal L_\tau$ given the last T data points: $< (x_{\tau}, y_\tau) : \tau \in [t-T+1 .. t]>$ and the previous recurrent state $s_{t-T}$.  This has the disadvantage of having to store T intermediate hidden states and do 2T sequential operations in the forward/backward pass.  Moreover it gives a biased gradient estimate because it ignores the effect of $\theta$ on $s_{t-T}$.

Another option which usually is even more expensive is Real-Time Recurrent Learning (RTRL).  RTRL is the application of [forward mode automatic differentiation](https://en.wikipedia.org/wiki/Automatic_differentiation#The_chain_rule,_forward_and_reverse_accumulation) to recurrent networks (Backpropagation Through Time is reverse-mode automatic differentiation).  Instead of first computing the loss and then backpropagating the gradient, we forward-propagate the jacobian defining the derivitive of the state with respect to the parameters: $\frac{\partial s_t}{\partial \theta} = \frac{\partial s_t}{\partial s_{t-1}}  \frac{\partial s_{t-1}}{\partial \theta} +  \frac{\partial s_t}{\partial \theta}|_{s_{t-1}} \in \mathbb R^{|S| \times |\Theta|}$ (where the second term is "the derivative with $s_{t-1}$ held constant"), and updating $\theta$ using the gradient  $\frac{\partial \mathcal L_t}{\partial \theta} = \frac{\partial \mathcal L_t}{\partial s_{t-1}}  \frac{\partial s_{t-1}}{\partial \theta} +  \frac{\partial \mathcal L_t}{\partial \theta}|_{s_{t-1}}$.  This is very expensive because it involves computing, storing and multiplying large Jacobian matrices.  

This paper proposes doing an approximate form of RTRL where the state-derivative is stochastically approximated as a Rank-1 matrix: $\frac{\partial s_t}{\partial \theta} \approx \tilde s_t \otimes \tilde \theta_t: \tilde s_t \in \mathbb R^{|S|}, \tilde \theta_t \in \mathbb R^{|\Theta|}$ where $\otimes$ denotes the outer-product.  The approximation uses the "Rank-1 Trick"*****.

They show that this approximation is **Unbiased**** (i.e. $\mathbb E[\tilde s_t \otimes \tilde \theta_t] = \frac{\partial s_t}{\partial \theta}$), and that using this approximation we can do much more computationally efficient updates than RTRL, and without the biasedness and backtracking required in TBPTT.

They demonstrate this result on some toy datasets.  They demonstrate that it's possible to construct a situation where TBPTT fails (due to biasedness of the gradient) and UORO converges, and that for other tasks they achieve comparable performance to TBPTT.

----

\* The "Rank-1 Trick", is that if matrix $A \in \mathbb R ^{M\times N}$ can be decomposed as $A = \sum_k^K v_k \otimes w_k: v_k \in \mathbb R^M, w_k \in \mathbb R^N$, then we can define $\tilde A :=\left( \sum_k^K \nu_k v_k \right) \otimes \left( \sum_k^K \nu_k w_k \right) \approx A$, with $\nu_k = \{1 \text{ with } p=\frac12 \text{ otherwise } -1\}$, and show that $\mathbb E[\tilde A] = A$.  This trick is applied twice to approximate the RTRL updates: First to approximate $s'\otimes \theta' :\approx \frac{\partial s_t}{\partial \theta}|_{s_{t-1}}$, then $\tilde s_t \otimes \tilde \theta_t :\approx \frac{\partial s_t}{\partial s_{t-1}}  (\tilde s_{t-1} \otimes \tilde \theta_{t-1}) + s' \otimes \theta'$.  (Note that in the paper they add the additional terms $\rho_k$ to this equation to reduce the variance of this estimator).


** The "unbiasedness" is a bit of a lie, because it is only true if $\theta$ is not changing over time, which it is during training.  This is the case for RTRL in general, and not just this work.  

A NLP paper.

> "conditional computation, achieving greater than 1000x improvements in model capacity with
only minor losses in computational efficiency on modern GPU clusters. We introduce
a Sparsely-Gated Mixture-of-Experts layer (MoE), consisting of up to
thousands of feed-forward sub-networks"

## Evaluation
* 1 billion word language modeling benchmark
* 100 billion word google news corpus