Domain translation - for example, mapping from a summer to a winter scene, or from a photorealistic image to an object segmentation map - is often performed by GANs through something called cycle consistency loss. This model works by having, for each domain, a generator to map domain A into domain B, and a discriminator to differentiate between real images from domain B, and those that were constructed through the cross-domain generator. With a given image in domain A, training happens by using the A→B generator to map it into domain B, and then then B→ A generator to map it back the original domain. These generators are then trained using two losses: one based on the B-domain discriminator, to push the generated image to look like it belongs from that domain, and another based on the L2 loss between the original domain A image, and the image you get on the other end when you translate it into B and back again. 

This paper addresses an effect (identified originally in an earlier paper) where in domains with a many to one mapping between domains (for example, mapping a realistic scene into a domain segmentation map, where information is inherently lost by translating pixels to object outlines), the cycle loss incentivizes the model to operate in a strange, steganographic way, where it saves information about the that would otherwise be lost in the form of low-amplitude random noise in the translated image. This low-amplitude information can't be isolated, but can be detected in a few ways. First, we can simply examine images and notice that information that could not have been captured in the lower-information domain is being perfectly reconstructed. Second, if you add noise to the translation in the lower-information domain, in such a way as to not perceptibly change the translation to human eyes, this can cause the predicted image off of that translation to deteriorate considerably, suggesting that the model was using information that could be modified by such small additions of noise to do its reconstruction. 

https://i.imgur.com/08i1j0J.png

The authors of this paper ask whether it's possible to train models that don't perform this steganographic information-storing (which they call "self adversarial examples"). A typical approach to such a problem would be to train generators to perform translations with and without the steganographic information, but even though we can prove the existence of the information, we can't isolate it in a way that would allow us to remove it, and thus create these kinds of training pairs. The two tactics the paper uses are: 

1) Simply training the generators to be able to translate a domain-mapped image with noise as well as one without noise, in the hope that this would train it not use information that can be interfered with by the application of such noise. 

2) In addition to a L2 cycle loss, adding a discriminator to differentiate between the back-translated image and the original one. I believe the idea here is that if both of the encoders are adding in noise as a kind of secret signal, this would be a way for the discriminator to distinguish between the original and reconstructed image, and would thus be penalized.  

They find that both of these methods reduce the use of steganographic information, as determined both by sensitivity to noise (where less sensitivity of reconstruction to noise means less use of coded information) and reconstruction honesty (which constrains accuracy of reconstruction in many to one domains to be no greater than the prediction that a supervised predictor could make given the image from the compressed domain

The paper designs some basic tests to compare saliency methods. It founds that some of the most popular methods are independent of model parameters and the data, meaning they are effectively useless.

## Methods compared

The paper compare the following methods: gradient explanation, gradient x input, integrated gradients, guided backprop, guided GradCam and SmoothGrad. They provide a refresher on those methods in the appendix.

All those methods can be put in the same framework. They require a classification model and an input (typically an image). The output of the method is an *explanation map* of the shape of the input where a higher value for a feature implies greater relevance in the decision of the model. 

## Metrics of comparison

The authors argue that visual inspection of the saliency maps can be misleading. They propose to compute the Spearman rank correlation, the structural similarity index (SSMI) and the Pearson correlation of the histogram of gradients. The authors point out that those metrics capture various notions of similarity, but it is an active area of research and those metrics are imperfect.

## First test: model parameters randomization

A saliency method must be dependent of model parameters, otherwise it cannot help us understand a model. In this test, the authors randomize the model parameters, layer per layer, starting from the top.

Surprisingly, methods such as guided backprop and guided gradcam are completely insensitive to model parameters, as illustrated on this Inception v3 trained on ImageNet:

![image](https://user-images.githubusercontent.com/8659132/61403152-b10b8000-a8a2-11e9-9f6a-cf1ed6a876cc.png)

Integrated gradients looks also dubious as the bird is still visible with a mostly fully randomized model, but the quantitative metrics reveal the difference is actually big between the two models.

## Second test: data randomization

It is well-known that randomly shuffling the labels of a dataset does not prevent a neural network from getting a high accuracy on the training set, though it does prevent generalization. The model is able to learn by either memorizing the data or finding spurious patterns. As a result, saliency maps obtained from such a network should have no clearly interpretable signal.

Here is the result for a ConvNet trained on MNIST and a shuffled MNIST:

![image](https://user-images.githubusercontent.com/8659132/61406757-7efe1c00-a8aa-11e9-9826-a859a373cb4f.png)

The results are very damning for most methods. Only gradients and GradCam are very different between both models, as confirmed by the low correlation.

## Discussion

- Even though some methods do no depend on model parameters and data, they might still depend on the architecture of the models, which could be of some use in some contexts.
- Methods that multiply the input with the gradient are dominated by the input.
- Complex saliency methods are just fancy edge detectors.
- Only gradient, smooth gradient and GradCam survives the sanity checks.

# Comments

- Why is their GradCam maps so ugly? They don't look like usual GradCam maps at all.
- Their tests are simple enough that it's hard to defend a method that doesn't pass them.
- The methods that are left are not very good either. They give fuzzy maps that are difficult to interpret.
- In the case of integrated gradients (IG), I'm not convinced this is sufficient to discard the method. IG requires a "baseline input" that represents the absence of features. In the case of images, people usually just set the image to 0, which is not at all the absence of a feature. The authors also use the "set the image to 0" strategy, and I'd say their tests are damning for this strategy, not for IG in general. I'd expect an estimation of the baseline such as done in [this paper](https://arxiv.org/abs/1702.04595) would be a fairer evaluation of IG.

Code: [GitHub](https://github.com/adebayoj/sanity_checks_saliency) (not available as of 17/07/19)

The authors introduce their contribution as an alternative way to approximate the KL divergence between prior and variational posterior used in [Variational Dropout and the Local Reparameterization Trick][kingma] which allows unbounded variance on the multiplicative noise. When the noise variance parameter associated with a weight tends to infinity you can say that the weight is effectively being removed, and in their implementation this is what they do.

There are some important details differing from the [original algorithm][kingma] on per-weight variational dropout. For both methods we have the following initialization for each dense layer:

```
theta = initialize weight matrix with shape (number of input units, number of hidden units)
log_alpha = initialize zero matrix with shape (number of input units, number of hidden units)
b = biases initialized to zero with length the number of hidden units
```

Where `log_alpha` is going to parameterise the variational posterior variance.

In the original paper the algorithm was the following:

```
mean = dot(input, theta) + b # standard dense layer
# marginal variance over activations (eq. 10 in [original paper][kingma])
variance = dot(input^2, theta^2 * exp(log_alpha)) 
# sample from marginal distribution by scaling Normal 
activations = mean + sqrt(variance)*unit_normal(number of output units) 
```

The final step is a standard [reparameterization trick][shakir], but since it is a marginal distribution this is referred to as a local reparameterization trick (directly inspired by the [fast dropout paper][fast]).

The sparsifying algorithm starts with an alternative parameterisation for `log_alpha`

```
log_sigma2 = matrix filled with negative constant (default -8) with size (number of input units, number of hidden units)
log_alpha = log_sigma2 - log(theta^2)
log_alpha = log_alpha clipped between 8 and -8
```

The authors discuss this in section 4.1, the $\sigma_{ij}^2$ term corresponds to an additive noise variance on each weight with $\sigma_{ij}^2 = \alpha_{ij}\theta_{ij}^2$. Since this can then be reversed to define `log_alpha` the forward pass remains unchanged, but the variance of the gradient is reduced. It is quite a counter-intuitive trick, so much so I can't quite believe it works.

They then define a mask removing contributions to units where the noise variance has gone too high:

```
clip_mask = matrix shape of log_alpha, equals 1 if log_alpha is greater than thresh (default 3)
```

The clip mask is used to set elements of `theta` to zero, and then the forward pass is exactly the same as in the original paper.

The difference in the approximation to the KL divergence is illustrated in figure 1 of the paper; the sparsifying version tends to zero as the variance increases, which matches the true KL divergence. In the [original paper][kingma] the KL divergence would explode, forcing them to clip the variances at a certain point.

[kingma]: https://arxiv.org/abs/1506.02557
[shakir]: http://blog.shakirm.com/2015/10/machine-learning-trick-of-the-day-4-reparameterisation-tricks/
[fast]: http://proceedings.mlr.press/v28/wang13a.html

If you were to survey researchers, and ask them to name the 5 most broadly influential ideas in Machine Learning from the last 5 years, I’d bet good money that Batch Normalization would be somewhere on everyone’s lists. Before Batch Norm, training meaningfully deep neural networks was an unstable process, and one that often took a long time to converge to success. When we added Batch Norm to models, it allowed us to increase our learning rates substantially (leading to quicker training) without the risk of activations either collapsing or blowing up in values. It had this effect because it addressed one of the key difficulties of deep networks: internal covariate shift. 

To understand this, imagine the smaller problem, of a one-layer model that’s trying to classify based on a set of input features. Now, imagine that, over the course of training, the input distribution of features moved around, so that, perhaps, a value that was at the 70th percentile of the data distribution initially is now at the 30th. We have an obvious intuition that this would make the model quite hard to train, because it would learn some mapping between feature values and class at the beginning of training, but that would become invalid by the end. This is, fundamentally, the problem faced by higher layers of deep networks, since, if the distribution of activations in a lower layer changed even by a small amount, that can cause a “butterfly effect” style outcome, where the activation distributions of higher layers change more dramatically. Batch Normalization - which takes each feature “channel” a network learns, and normalizes [normalize = subtract mean, divide by variance] it by the mean and variance of that feature over spatial locations and over all the observations in a given batch - helps solve this problem because it ensures that, throughout the course of training, the distribution of inputs that a given layer sees stays roughly constant, no matter what the lower layers get up to. 

On the whole, Batch Norm has been wildly successful at stabilizing training, and is now canonized - along with the likes of ReLU and Dropout - as one of the default sensible training procedures for any given network. However, it does have its difficulties and downsides. One salient one of these comes about when you train using very small batch sizes - in the range of 2-16 examples per batch. Under these circumstance, the mean and variance calculated off of that batch are noisy and high variance (for the general reason that statistics calculated off of small sample sizes are noisy and high variance), which takes away from the stability that Batch Norm is trying to provide. 

One proposed alternative to Batch Norm, that didn’t run into this problem of small sample sizes, is Layer Normalization. This operates under the assumption that the activations of all feature “channels” within a given layer hopefully have roughly similar distributions, and, so, you an normalize all of them by taking the aggregate mean over all channels, *for a given observation*, and use that as the mean and variance you normalize by. Because there are typically many channels in a given layer, this means that you have many “samples” that go into the mean and variance. However, this assumption - that the distributions for each feature channel are roughly the same - can be an incorrect one. 

A useful model I have for thinking about the distinction between these two approaches is the idea that both are calculating approximations of an underlying abstract notion: the in-the-limit mean and variance of a single feature channel, at a given point in time. Batch Normalization is an approximation of that insofar as it only has a small sample of points to work with, and so its estimate will tend to be high variance. Layer Normalization is an approximation insofar as it makes the assumption that feature distributions are aligned across channels: if this turns out not to be the case, individual channels will have normalizations that are biased, due to being pulled towards the mean and variance calculated over an aggregate of channels that are different than them. 

Group Norm tries to find a balance point between these two approaches, one that uses multiple channels, and normalizes within a given instance (to avoid the problems of small batch size), but, instead of calculating the mean and variance over all channels, calculates them over a group of channels that represents a subset. The inspiration for this idea comes from the fact that, in old school computer vision, it was typical to have parts of your feature vector that - for example - represented a histogram of some value (say: localized contrast) over the image. Since these multiple values all corresponded to a larger shared “group” feature. If a group of features all represent a similar idea, then their distributions will be more likely to be aligned, and therefore you have less of the bias issue. 

One confusing element of this paper for me was that the motivation part of the paper strongly implied that the reason group norm is sensible is that you are able to combine statistically dependent channels into a group together. However, as far as I an tell, there’s no actually clustering or similarity analysis of channels that is done to place certain channels into certain groups; it’s just done so semi-randomly based on the index location within the feature channel vector. So, under this implementation, it seems like the benefits of group norm are less because of any explicit seeking out of dependant channels, and more that just having fewer channels in each group means that each individual channel makes up more of the weight in its group, which does something to reduce the bias effect anyway.

The upshot of the Group Norm paper, results-wise, is that Group Norm performs better than both Batch Norm and Layer Norm at very low batch sizes. This is useful if you’re training on very dense data (e.g. high res video), where it might be difficult to store more than a few observations in memory at a time. However, once you get to batch sizes of ~24, Batch Norm starts to do better, presumably since that’s a large enough sample size to reduce variance, and you get to the point where the variance of BN is preferable to the bias of GN.  

This paper continues in the tradition of curiosity-based models, which try to reward models for exploring novel parts of their environment, in the hopes this can intrinsically motivate learning. However, this paper argues that it’s insufficient to just treat novelty as an occasional bonus on top of a normal reward function, and that instead you should figure out a process that’s more specifically designed to increase novelty. Specifically: you should design a policy whose goal is to experience transitions and world-states that are high novelty. 

In this setup, like in other curiosity-based papers, “high novelty” is defined in terms of a state being unpredictable given a prior state, history, and action. However, where other papers saw novelty reward as something only applied when the agent arrived at somewhere novel, here, the authors build a model (technically, an ensemble of models) to predict the state at various future points. The ensemble is important here because it’s (quasi) bootstrapped, and thus gives us a measure of uncertainty. States where the predictions of the ensemble diverge represent places of uncertainty, and thus of high value to explore. I don’t 100% follow the analytic specification of this idea (even though the heuristic/algorithmic description makes sense). The authors frame the Utility function of a state and action as being equivalent to the Jenson Shannon Divergence (~distance between probability distributions) shown below. 

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


Here, P(S | S, a, T) is the probability of a state given prior state and action under a given model of the environment (Transition Model), and P(gamma) is the distribution over the space of possible transition models one might learn. A “model” here is one network out of the ensemble of networks that makes up our bootstrapped (trained on different sets) distribution over models. Conceptually, I think this calculation is measuring “how different is each sampled model/state distribution from all the other models in the distribution over possible models”. If the models within the distribution diverge from one another, that indicates a location of higher uncertainty. 

What’s important about this is that, by building a full transition model, the authors can calculate the expected novelty or “utility” of future transitions it might take, because it can make a best guess based on this transition model (which, while called a “prior”, is really something trained on all data up to this current iteration). My understanding is that these kinds of models function similarly to a Q(s,a) or V(s) in a pure-reward case: they estimate the “utility reward” of different states and actions, and then the policy is updated to increase that expected reward. 

I’ve recently read papers on ICM, and I was a little disappointed that this paper didn’t appear to benchmark against that, but against Bootstrapped DQN and Exploration Bonus DQN, which I know less well and can less speak to the conceptual differences from this approach. Another difficulty in actually getting a good sense of results was that the task being tested on is fairly specific, and different from RL results coming out of the world of e.g. Atari and Deep Mind Labs. All of that said, this is a cautiously interesting idea, if the results generate to beat more baselines on more environments.