### Summary
Knowing when a model is qualified to make a prediction is critical to safe deployment of ML technology. Model-independent / Unsupervised Out-of-Distribution (OoD) detection is appealing mostly because it doesn't require task-specific labels to train. It is tempting to suggest a simple one-tailed test in which lower likelihoods are OoD (assigned by a Likelihood Model), but the intuition that In-Distribution (ID) inputs should have highest likelihoods _does not hold in higher dimension_. The authors propose to use the Watanabe-Akaike Information Criterion (WAIC) to circumvent this problem and empirically show the robustness of the approach.

### Counterintuitive Properties of Likelihood Models:
https://i.imgur.com/4vo0Ff5.png
So a GLOW model with Gaussian prior maps SVHN closer to the origin than Cifar (but never actually generates SVHN because Gaussian samples are on the shell). This is bad news for OoD detection.

### Proposed Methodology:
Use the WAIC criterion for OoD detection which gives an asymptotically correct estimate of the gap between the training set and test set expectations:
https://i.imgur.com/vasSxuk.png
Basically,  the correction term subtracts the variance in likelihoods across independent samples from the posterior. This acts to robustify the estimate, ensuring that points that are sensitive to the particular choice of posterior are penalized. They use an ensemble of generative models as a proxy for posterior samples i.e. the ensembles acts as approximate posterior samples.
Now, OoD can be detected with a Likelihood Model:
https://i.imgur.com/M3CDKOA.png
### Discussion
Interestingly, GLOW maps Cifar and other datasets INSIDE the gaussian shell (which is an annulus of radius $\sqrt{dim} = \sqrt{3072} \approx 55.4$
https://i.imgur.com/ERdgOaz.png
This is in itself quite disturbing, as it suggests that better flow-based generative models (for sampling) can be obtained by encouraging the training distribution to overlap better with the typical set in latent
space.

Adversarial examples and defenses to prevent them are often presented as a case of inherent model fragility, where the model is making a clear and identifiable mistake, by misclassifying a label humans would classify correctly. But, another frame on the adversarial examples research is that they're a way of imposing a certain kind of prior requirement on our models: that they be sensitive to certain scales of perturbation to their inputs. One reason to want to do this is because you believe the model might reasonably need to interact with such perturbed inputs in future. But, another is that smoothness of model outputs, in the sense of an output that doesn't change sharply in the immediate vicinity of an example, can be a useful inductive bias that improves generalization. In images, this is often not the case, as training on adversarial examples empirically worsens performance on normal examples.  In text, however, it seems like you can get more benefit out of training on adversarial examples, and this paper proposes a specific way of doing that. 

An interesting up-front distinction is the one between generating adversarial examples in embeddings vs raw text. Raw text is generally harder: it's unclear how to permute sentences in ways that leave them grammatically and meaningfully unchanged, and thus mean that the same label is the "correct" one as before, without human input. So the paper instead works in embedding space: adding a delta vectors of adversarial noise to the learned word embeddings used in a text model. One salient downside of generating adversarial examples to train on is that doing so is generally costly: it requires calculating the gradients with respect to the input to calculate the direction of the delta vector, which requires another backwards pass through the network, in addition to the ones needed to calculate the parameter gradients to update those. It happens to be the case that once you've calculated gradients w.r.t inputs, doing so for parameters is basically done for you for free, so one possible solution to this problem is to do a step of parameter gradient calculation/model training every time you take a step of perturbation generation. However, if you're generating your adversarial examples via multi-step Projected Gradient Descent, doing a step of model training at each of the K steps in multi-step PGD means that by the time you finish all K steps and are ready to train on the example, your perturbation vector is out of sync with with your model parameters, and so isn't optimally adversarial. To fix this, the authors propose actually training on the adversarial example generated by each step in the multi-step generation process, not just the example produced at the end. So, instead of training your model on perturbations of a given size, you train them on every perturbation up to and including that size. This also solves the problem of your perturbation being out of sync with your parameters, since you "apply" your perturbation in training at the same step where you calculate it. 

The authors sole purpose in this was to make models that generalize better, and they show reasonably convincing evidence that this method works slightly better than competing alternatives on language modeling tasks. More saliently, in my view, they come up with a straightforward and clever solution to a problem, which could potentially be used in other domains.

Summary of this post

 - How does this minibatch discrimination heuristic work and how does it change the behaviour of the GAN algorithm? Does it change the underlying objective function that is being minimized?
 - the answer is: for the original GAN algorithm that minimises Jensen-Shannon divergence it does change the behaviour in a non-trivial way. One side-effect is assigning a higher relative penalty for low-entropy generators.
 - when using the blended update rule from here, the algorithm minimises the reverse KL-divergence. In this case, using minibatch discrimination leaves the underlying objective unchanged: the algorithm can still be shown to miminise KL divergence.
 - even if the underlying objectives remain the same, using minibatch discrimination may be a very good idea practically. It may stabilise the algorithm by, for example, providing a lower variance estimator to log-probability-ratios.

[Here is the ipython/jupyter notebook](https://gist.github.com/fhuszar/a91c7d0672036335c1783d02c3a3dfe5) I used to draw the plots and test some of the things in this post in practice.

### What is minibatch discrimination?

In the vanilla Generative Adversarial Networks (GAN) algorithm, a discriminator is trained to tell apart generated synthetic examples from real data. One way GAN training can fail is to massively undershoot the entropy of the data-generating distribution, and concentrate all it's parameters on generating just a single or few examples.

To remedy this, the authors play with the idea of discriminating between whole minibatches of samples, rather than between individual samples. If the generator has low entropy, much lower than real data, it may be easier to detect this with a discriminator that sees multiple samples.

Here, I'm going to look at this technique in general: modifying an unsupervised learning algorithm by replacing individual samples with i.i.d. minibatches of samples. Note, that this is not exactly what the authors end up doing in the paper referenced above, but it's an interesting trick to think about.

#### How does the minibatch heuristic effect divergences?

The reason I'm so keen on studying GANs is the connection to principled information theoretic divergence criteria. Under some assumptions, it can be shown that GANs minimise the Jensen-Shannon (JS) divergence, or with a slight modification the reverse-KL divergence. In fact, a recent paper showed that you can use GAN-like algorithms to minimise any $f$-divergence.

So my immediate question looking at the minibatch discrimination idea was: how does this heuristic change the divergences that GANs minimise.

#### KL divergence

Let's assume we have any algorithm (GAN or anything else) that minimises KL divergence $\operatorname{KL}[P\|Q]$ between two distributions $P$ and $Q$. Let's now modify this algorithm so that instead of looking at distributions $P$ and $Q$ of a single sample $x$, it looks at distributions $P^{(N)}$ and $Q^{(N)}$ of whole a minibatch $(x_1,\ldots,x_N)$. I use $P^{(N)}$ to denote the following distribution:

$$ P^{(N)}(x_1,\ldots,x_N) = \prod_{n=1}^N P(x_n) 
$$

The resulting algorithm will therefore minimise the following divergence:

$$ d[P\|Q] = \operatorname{KL}[P^{(N)}\|Q^{(N)}] 
$$

It is relatively easy to show why this divergence $d$ behaves exactly like the KL divergence between $P$ and $Q$. Here's the maths for minibatch size of $N=2$:

\begin{align} d[P\|Q] &= \operatorname{KL}[P^{(2)}\|Q^{(2)}] \\ 
&= \mathbb{E}_{x_1\sim P,x_2\sim P}\log\frac{P(x_1)P(x_2)}{Q(x_1)Q(x_2)} \\ &= \mathbb{E}_{x_1\sim P,x_2\sim P}\log\frac{P(x_1)}{Q(x_1)} + \mathbb{E}_{x_1\sim P,x_2\sim P}\log\frac{P(x_2)}{Q(x_2)} \\ &= \mathbb{E}_{x_1\sim P}\log\frac{P(x_1)}{Q(x_1)} + \mathbb{E}_{x_2\sim P}\log\frac{P(x_2)}{Q(x_2)} \\ &= 2\operatorname{KL}[P\|Q] \end{align}

In full generality we can say that:

$$ \operatorname{KL}[P^{(N)}\|Q^{(N)}] = N \operatorname{KL}[P\|Q] $$

So changing the KL-divergence to minibatch KL-divergence does not change the objective of the training algorithm at all. Thus, if one uses minibatch discrimination with the blended training objective, one can rest assured that the algorithm still performs approximate gradient descent on the KL divergence. It may still work differently in practice, for example by reducing the variance of the estimators involved.

This property of the KL divergence is not surprising if one considers its compression/information theoretic definition: the extra bits needed to compress data drawn from $P$, using model $Q$. Compressing a minibatch of i.i.d. samples corresponds to compressing the samples independently. Their codelengths would add up linearly, hence KL-divergences add up linearly, too.

#### JS divergence

The same thing does not hold for the JS-divergence. Generally speaking, minibatch JS divergence behaves differently from ordinary JS-divergence. Instead of equality, for JS divergences the following inequality holds:

$$ JS[P^{(N)}\|Q^{(N)}] \leq N \cdot JS[P\|Q] 
$$

In fact for fixed $P$ and $Q$, $JS[P^{(N)}\|Q^{(N)}]/N$ is monotonically non-increasing. This can be seen intuitively by considering the definition of JS divergence as the mutual information between the samples and the binary indicator $y$ of whether the samples were drawn from $Q$ or $P$. Using this we have that:

\begin{align} \operatorname{JS}[P^{(2)}\|Q^{(2)}] &= \mathbb{I}[y ; x_1, x_2] \\ &= \mathbb{I}[y ; x_1] + \mathbb{I}[y ; x_2 \vert x_1] \\ &\leq \mathbb{I}[y ; x_1] + \mathbb{I}[y ; x_2] \\ &= 2 \operatorname{JS}[P\|Q] \end{align}

Below I plotted the minibatch-JS-divergence $JS[P^{(N)}\|Q^{(N)}]$ for various minibatch-sizes $N=1,2,3,8$, between univariate Bernoulli distributions with parameters $p$ and $q$. For the plots below, $p$ is kept fixed at $p=0.2$, and the parameter $q$ is varied between $0$ and $1$.

![](http://www.inference.vc/content/images/2016/06/fr9ra2qB2t9stl8sVnqIAAH1GqAAAhM3YsWOVkpKiTz75RD6fL9D-0UcfhbEqAEBfRYe7AABA5DIMQ-vXr9eLL76op556So8--qjOnDmjsrIyGYYR7vIAAL3ESAUAIKxmzpyp9957T52dndqyZYuOHTum4uJi7v4EAIMIIxUAgLCbPn26pk-fHtj2er1hrAYA0FeMVAAAAAAICaECAAAAQEgIFQCAm5JhGFysDQCDhOHnSjgAAAAAIWCkA.png)

You can see that all divergences have a unique global minimum around $p=q=0.2$. However, their behaviour at the tails changes as the minibatch-size increases. This change in behaviour is due to saturation: JS divergence is upper bounded by $1$, which corresponds to 1 bit of information. If I continued increasing the minibatch-size (which would blow up the memory footprint of my super-naive script), eventually the divergence would reach $1$ almost everywhere except for a dip down to $0$ around $p=q=0.2$.

Below are the same divergences normalised to be roughly the same scale.

![](http://www.inference.vc/content/images/2016/06/Rle8idcTS3gAAAABJRU5ErkJggg--.png)

The problem of GANs that minibatch discrimination was meant to fix is that it favours low-entropy solutions. In this plot, this would correspond to the $q<0.1$ regime. You can argue that as the batch-size increases, the relative penalty for low-entropy approximations $q<0.1$ do indeed decrease when compared to completely wrong solutions $q>0.5$. However, the effect is pretty subtle.

Bonus track: adversarial preference loss

In this context, I also revisited the adversarial preference loss. Here, the discriminator receives two inputs $x_1$ and $x_2$ (one synthetic, one real) and it has to decide which one was real.

This algorithm, too, can be related to the minibatch discrimination approach, as it minimises the following divergence:

$$ d(P,Q) = d(P\times Q\|Q\times P), 
$$

where $P\times Q(x_1,x_2) = P(x_1)Q(x_2)$. Again, if $d$ is the $KL$ divergence, the training objective boils down to the same thing as the original GAN. However, if $d$ is the JS divergence, we will end up minimising something weird, $\operatorname{JS}[Q\times P\| P\times Q]$

### Keywords
Adversarial example ,  Perturbations 

------

### Summary

##### Introduction 

* Explain two properties of  neural network that cause it to misclassify images and cause  difficulty to get solid understanding of network.
1. Theoretical understanding  of the individual high level unit of a network and a combination of these units or layers.
2. Understanding the continuity of input - output mapping space and the stability of the output wrt. the input.

* Performing a few experiments on different networks and architectures
1. MNIST dataset - Autoencoder , Fully Connected net
2. ImageNet - “AlexNet”
3. 10M youtube images - “QuocNet”  

##### Understanding individual units of the Network 

* Previous work used individual images to maximize the activation value of each feature unit.
Similar experiment was done by the authors on the MNIST data set.

* The interpretation of the results are as following ;
1. Random direction vector (V)  gives rise to similarly interpretable semantic properties.
2. Each feature unit is able to generate invariance on a particular subset of input   distribution.
https://i.imgur.com/SeyXJeV.png

##### Blind spots in the neural network

* Output layers are highly non-linear and are able to give a nonlinear generalization over the input space.
* It is possible for the output layers to give non-significant probabilities to regions of the input space that contain no training examples in their vicinity. Ie. It is possible to obtain probability of the different viewpoints of the object without training.
* Deep learning kernel methods can't be assumed to have smooth decision boundaries. 
* Using optimization techniques, small changes to the image can lead to very large deviations in the output
* __“Adversarial examples”__ represent pockets or holes in the input-space which are difficult to find simply moving around the input images.


##### Experimental Results
* Adversarial examples that are indistinguishable from the actual image can be created for all networks.
1. Cross model generalization : Adversarial images created for one network can affect the other networks also.
2. Cross training generalization 

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

##### Conclusion 
* Neural network have a counter intuitive properties wrt. the working of the individual units and discontinuities.
* Occurance of the adversarial examples and its properties. 

-----
### Notes 

* Feeding adversarial examples during the model training can improve the generalization of the model.
* The adversarial examples on the higher layers are more effective than those of input and lower layers.
* Adversarial examples affect models trained with different hyper parameters.
* According to the the test conducted , autoencoders are more resilient to the adversarial examples.
* Deep learning networks which are trained from purely supervised training are unstable to a few particular types of perturbations. Small addition of perturbations to the input leads to large perturbations at the output of the last layers.

### Open research questions

[1] Comparing the effects of adversarial examples on lower layers to that of the higher layers.     
[2] Dependence of the adversarial attacks on training data set of the model.     
[3] Why the adversarial examples generalize across different hyperparameters or training sets.     
[4] How often do adversarial example occur?      

Zahavy et al. introduce the concept of ensemble robustness and show that it can be used as indicator for generalization performance. In particular, the main idea is to lift he concept of robustness against adversarial examples to ensemble of networks – as trained, e.g. through Dropout or Bayes-by-Backprop. Letting $Z$ denote the sample set, a learning algorithm is $(K, \epsilon)$ robust if $Z$ can be divided into $K$ disjoint sets $C_1,\ldots,C_K$ such that for every training set $s_1,\ldots,s_n  \in Z$ it holds:

$\forall i, \forall z \in Z, \forall k = 1,\ldots, K$: if $s,z \in C_k$, then $l(f,s_i) – l(f,z)| \leq \epsilon(s_1,\ldots,s_n)$

where $f$ is the model produced by the learning algorithm, $l$ measures the loss and $\epsilon:Z^n \mapsto \mathbb{R}$. For ensembles (explicit or implicit) this definition is extended by considering the maximum generalization loss under the expectation of a randomized learning algorithm:

$\forall i, \forall k = 1,\ldots,K$: if $s \in C_k$, then $\mathbb{E}_f \max_{z \in C_k} |l(f,s_i) – l(f,z)| \leq \epsilon(s_1,\ldots,s_n)$

Here, the randomized learning algorithm computes a distribution over models given a training set.

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