This paper proposes a neural architecture that allows to backpropagate gradients though a procedure that can go through a variable and adaptive number of iterations. These "iterations" for instance could be the number of times computations are passed through the same recurrent layer (connected to the same input) before producing an output, which is the case considered in this paper. 

This is essentially achieved by pooling the recurrent states and respective outputs computed by each iteration. The pooling mechanism is essentially the same as that used in the really cool Neural Stack architecture of Edward Grefenstette, Karl Moritz Hermann, Mustafa Suleyman and Phil Blunsom \cite{conf/nips/GrefenstetteHSB15}. It relies on the introduction of halting units, which are sigmoidal units computed at each iteration and which gives a soft weight on whether the computation should stop at the current iteration.

Crucially, the paper introduces a new ponder cost $P(x)$, which is a regularization cost that penalizes what is meant to be a smooth upper bound on the number of iterations $N(t)$ (more on that below).

The paper presents experiment on RNNs applied on sequences where, at each time step t (not to be confused with what I'm calling computation iterations, which are indexed by n) in the sequence the RNN can produce a variable number $N(t)$ of intermediate states and outputs. These are the states and outputs that are pooled, to produce a single recurrent state and output for the time step t. During each of the $N(t)$ iterations at time step t, the intermediate states are connected to the same time-step-t input. After the $N(t)$ iterations, the RNN pools the $N(t)$ intermediate states and outputs, and then moves to the next time step $t+1$. To mark the transitions between time steps, an extra binary input is appended, which is 1 only for the first intermediate computation iteration. 

Results are presented on a variety of synthetic problems and a character prediction problem.

## General stuff about face recognition

Face recognition has 4 main tasks:

* **Face detection**: Given an image, draw a rectangle around every face
* **Face alignment**: Transform a face to be in a canonical pose
* **Face representation**: Find a representation of a face which is suitable for follow-up tasks (small size, computationally cheap to compare, invariant to irrelevant changes)
* **Face verification**: Images of two faces are given. Decide if it is the same person or not.

The face verification task is sometimes (more simply) a face classification task (given a face, decide which of a fixed set of people it is).

Datasets being used are:

* **LFW** (Labeled Faces in the Wild): 97.35% accuracy; 13 323 web photos of 5 749 celebrities
* **YTF** (YouTube Faces): 3425 YouTube videos of 1 595 subjects
* **SFC** (Social Face Classification): 4.4 million labeled faces from 4030 people, each 800 to 1200 faces
* **USF** (Human-ID database): 3D scans of faces

## Ideas in this paper

This paper deals with face alignment and face representation.

**Face Alignment**

They made an average face with the USF dataset. Then, for each new face, they apply the following procedure:

* Find 6 points in a face (2 eyes, 1 nose tip, 2 corners of the lip, 1 middle point of the bottom lip)
* Crop according to those
* Find 67 points in the face / apply them to a normalized 3D model of a face
* Transform (=align) face to a normalized position

**Representation**

Train a neural network on 152x152 images of faces to classify 4030 celebrities. Remove the softmax output layer and use the output of the second-last layer as the transformed representation.

The network is:

* C1 (convolution): 32 filters of size $11 \times 11 \times 3$ (RGB-channels) (returns $142\times 142$ "images")
* M2 (max pooling): $3 \times 3$, stride of 2  (returns $71\times 71$ "images")
* C3 (convolution): 16 filters of size $9 \times 9 \times 16$ (returns $63\times 63$ "images")
* L4 (locally connected): $16\times9\times9\times16$ (returns $55\times 55$ "images")
* L5 (locally connected): $16\times7\times7\times16$ (returns $25\times 25$ "images")
* L6 (locally connected): $16\times5\times5\times16$ (returns $21\times 21$ "images")
* F7 (fully connected): ReLU, 4096 units
* F8 (fully connected): softmax layer with 4030 output neurons

The training was done with:

* Stochastic Gradient Descent (SGD)
* Momentum of 0.9
* Performance scheduling (LR starting at 0.01, ending at 0.0001)
* Weight initialization: $w \sim \mathcal{N}(\mu=0, \sigma=0.01)$, $b = 0.5$
* ~15 epochs ($\approx$ 3 days) of training


## Evaluation results

* **Quality**:
  * 97.35% accuracy (or mean accuracy?) with an Ensemble of DNNs for LFW
  * 91.4% accuracy with a single network on YTF
* **Speed**: DeepFace runs in 0.33 seconds per image (I'm not sure which size). This includes image decoding, face detection and alignment, **the** feed forward network (why only one? wasn't this the best performing Ensemble?) and final classification output

## See also

* Andrew Ng: [C4W4L03 Siamese Network](https://www.youtube.com/watch?v=6jfw8MuKwpI)

This paper suggests a novel explanation for why dropout training is helpful: because it corresponds to an adaptive data augmentation method. Indeed, the authors point out that, when sampling a mask of the hidden units in a network (effectively setting the corresponding units to 0), the same effect would have been obtained by feeding as input an example tailored to yield activations of 0 for these units and otherwise the same activation for all other units. Since this "ghost" example will have to be different from the original example, and since each different mask would correspond to a different "ghost" example, then effectively mask sampling is similar to data augmentation.

While in practice finding a ghost example that replicates exactly the same dropout hidden activations might not be possible, the authors show that finding an "approximate" ghost example that minimizes a distance between the target dropout activation and the deterministic activation of the ghost example works well. Indeed, they show that training a deep neural net on additional data generated by this procedure yields results that are at least as good as regular dropout on MNIST and CIFAR-10 (actually, the deterministic neural net still uses regular dropout at the input layer, however they do show that the additional ghost examples are necessary to match the neural net trained with dropout at all layers).

Then the authors use that interpretation to justify a variation of dropout where the dropout rate isn't fixed, but itself is randomly sampled in some range for each example. Indeed, if we think of dropout at a fixed rate as a specific class of ghost data being added, varying the dropout rate corresponds to enriching even more the ghost data pool. The experiments show that this can help, though not by much.

Finally, the authors propose an explanation of a property of dropout: that it tends to generate hidden representations that are sparser. Again, the authors rely on their interpretation of dropout as data augmentation. The explanation goes as follows. Training on the ghost data distribution might imply that the classification problem has become significantly harder. Specifically, it is quite possible that the addition of new ghost examples generates new isolated class clusters in input space that the model most now learn to  discriminate. And they hypothesize that the generation of such additional clusters would encourage sparsity. To test this hypothesis, the authors synthetically simulate this scenario, by sampling data on a circle, which is clustered in small arcs each assigned to one of 10 possible classes in cycling order. Decreasing the arc length thus increases the number of arcs, i.e. class clusters. They show that training deep networks on datasets with increasing number of class clusters does yield representations that are increasingly sparser. This thus suggests that dropout might indeed be equivalent to modifying the input distribution by adding such isolated class-specific clusters in input space. 

One assumption behind this analysis is that the sparsity patterns (i.e. the set of non-zero dimensions) play an important role in classification and incorporate most of the discriminative class information. This assumption is also confirmed in experiments, where converting the ReLU activation function by a binary activation (that is 1 if the pre-activation is positive and 0 otherwise) after training still yields a network with good performance (though slightly worse).


#### My two cents

This is a really original and thought provoking paper. One interpretation I make of these results is that the inductive bias corresponding to using a deep neural network with ReLU activations is more valuable than one might have thought, and that the usefulness of deep neural networks goes beyond just being black boxes that can learn data-dependent representations. Otherwise, it's not clear to me why the ghost data implicitly generated by the architecture would be useful at all. This also suggests an experiment where such ghost samples would be fed to  another type of classifier, such as an SVM, to test whether the data augmentation is useful in itself and reflects meaningful structure in the data, as opposed to being somehow useful only for neural nets.

I note that the results are mostly specific to architectures based on ReLU activations (not that this is a problem, but one should keep this in mind).

I'd really like to see what the ghost samples look like. Do they correspond to interpretable images? The authors also mention that exploring how the samples change with training would be interesting to investigate, and I agree.

Finally, I think there might be a typo in Figure 1. While the labels of a) and b) states that the arc length is smaller for a) than b), the plot clearly show otherwise.

I really enjoyed this paper - in addition to being a clean, fundamentally empirical work, it was also clearly written, and had some pretty delightful moments of quotable zen, which I’ll reference at the end. The paper’s goal is to figure out how far curiosity-driven learning alone can take reinforcement learning systems, without the presence of an external reward signal. “Intrinsic” reward learning is when you construct a reward out of internal, inherent features of the environment, rather than using an explicit reward function. In some ways, intrinsic learning in RL can be thought of as analogous to unsupervised learning in classification problems, since reward functions are not inherent to most useful environments, and (when outside of game environments that inherently provide rewards), frequently need to be hand-designed. Curiosity-driven learning is a subset of intrinsic learning, which uses as a reward signal the difference between a prediction made by the dynamics model (predicting next state, given action) and the true observed next state. Situations where the this prediction area are high generate high reward for the agent, which incentivizes it to reach those states, which allows the dynamics model to then make ever-better predictions about them. 

Two key questions this paper raises are: 


1) Does this approach even work when used on it’s own? Curiosity had previously most often been used as a supplement to extrinsic rewards, and the authors wanted to know how far it could go separately. 


2) What is the best feature to do this “surprisal difference” calculation in? Predicting raw pixels is a high-dimensional and noisy process, so naively we might want something with fewer, more informationally-dense dimensions, but it’s not obvious which methods that satisfy these criteria will work the best, so the paper empirically tried them. 

The answer to (1) seems to be: yes, at least in the video games tested. Impressively, when you track against extrinsic reward (which, again, these games have, but we’re just ignoring in a curiosity-only setting), the agents manage to increase it despite not optimizing against it directly. There were some Atari games where this effect was stronger than others, but overall performance was stronger than might have been naively expected. One note the authors made, worth keeping in mind, is that it’s unclear how much of this is an artifact of the constraints and incentives surrounding game design, which might reflect back a preference for gradually-increasing novelty because humans find it pleasant. 

https://i.imgur.com/zhl39vo.png

As for (2), another interesting result of this paper is that random features performed consistently well as a feature space to do these prediction/reality comparisons in. Random features here is really just as simple as “design a convolutional net that compresses down to some dimension, randomly initialize it, and then use those randomly initialized weights to run forward passes of the network to get your lower-dimensional state”. This has the strong disadvantage of (presumably) not capturing any meaningful information about the state, but also has the advantage of being stable: the other techniques tried, like pulling out the center of a VAE bottleneck, changed over time as they were being trained on new states, so they were informative, but non-stationary. 

My two favorite quotable moments from this paper were: 
1) When the authors noted that they had removed the “done” signal associated with an agent “dying,” because it is itself a sort of intrinsic reward. However, “in practice, we do find that the agent avoids dying in the games since that brings it back to the beginning of the game, an area it has already seen many times and where it can predict the dynamics well.”. Short and sweet: “Avoiding death, because it’s really boring” 

https://i.imgur.com/SOfML8d.png

2) When they noted that an easy way to hack the motivation structure of a curiosity-driven agent was through a “noisy tv”, which, every time you pressed the button, jumped to a random channel. As expected, when they put this distraction inside a maze, the agent spent more time jacking up reward through that avenue, rather than exploring. Any resemblance to one’s Facebook feed is entirely coincidental. 

I should say from the outset: I have a lot of fondness for this paper. It goes upstream of a lot of research-community incentives: It’s not methodologically flashy, it’s not about beating the State of the Art with a bigger, better model (though, those papers certainly also have their place). The goal of this paper was, instead, to dive into a test set used to evaluate performance of models, and try to understand to what extent it’s really providing a rigorous test of what we want out of model behavior. Test sets are the often-invisible foundation upon which ML research is based, but like real-world foundations, if there are weaknesses, the research edifice built on top can suffer. 

Specifically, this paper discusses the Winograd Schema, a clever test set used to test what the NLP community calls “common sense reasoning”. An example Winograd Schema sentence is: 
The delivery truck zoomed by the school bus because it was going so fast.
A model is given this task, and asked to predict which token the underlined “it” refers to. These cases are specifically chosen because of their syntactic ambiguity - nothing structural about the order of the sentence requires “it” to refer to the delivery truck here. However, the underlying meaning of the sentence is only coherent under that parsing. This is what is meant by “common-sense” reasoning: the ability to understand meaning of a sentence in a way deeper than that allowed by simple syntactic parsing and word co-occurrence statistics.

Taking the existing Winograd examples (and, when I said tiny, there are literally 273 of them) the authors of this paper surface some concerns about ways these examples might not be as difficult or representative of “common sense” abilities as we might like. 

- First off, there is the basic, previously mentioned fact that there are so few examples that it’s possible to perform well simply by random chance, especially over combinatorially large hyperparameter optimization spaces. This isn’t so much an indictment of the set itself as it is indicative of the work involved in creating it. 
- One of the two distinct problems the paper raises is that of “associativity”. This refers to situations where simple co-occurance counts between the description and the correct entity can lead the model to the correct term, without actually having to parse the sentence. An example here is:
“I’m sure that my map will show this building; it is very famous.” 
Treasure maps aside, “famous buildings” are much more generally common than “famous maps”, and so being able to associate “it” with a building in this case doesn’t actually require the model to understand what’s going on in this specific sentence. The authors test this by creating a threshold for co-occurance, and, using that threshold, call about 40% of the examples “associative”
- The second problem is that of predictable structure - the fact that the “hinge” adjective is so often the last word in the sentence, making it possible that the model is brittle, and just attending to that, rather than the sentence as a whole 

The authors perform a few tests - examining results on associative vs non-associative examples, and examining results if you switch the ordering (in cases like “Emma did not pass the ball to Janie although she saw that she was open,” where it’s syntactically possible), to ensure the model is not just anchoring on the identity of the correct entity, regardless of its place in the sentence. Overall, they found evidence that some of the state of the art language models perform well on the Winograd Schema as a whole, but do less well (and in some cases even less well than the baselines they otherwise outperform) on these more rigorous examples. Unfortunately, these tests don’t lead us automatically to a better solution - design of examples like this is still tricky and hard to scale - but does provide valuable caution and food for thought.