This is a very techincal paper and I only covered items that interested me
* Model
  * Encoder
    * 8 layers LSTM 
    * bi-directional only first encoder layer
    * top 4 layers add input to output (residual network)
  * Decoder
    * same as encoder except all layers are just forward direction
  * encoder state is not passed as a start point to Decoder state
  * Attention
    * energy computed using NN with one hidden layer as appose to dot product or the usual practice of no hidden layer and $\tanh$ activation at the output layer
    * computed from output of 1st decoder layer
    * pre-feed to all layers
* Training has two steps: ML and RL
  * ML (cross-entropy) training:
    * common wisdom, initialize all trainable parameters uniformly between [-0.04, 0.04]
    * clipping=5, batch=128
    * Adam (lr=2e-4) 60K steps followed by SGD (lr=.5 which is probably a typo!) 1.2M steps + 4x(lr/=2 200K steps)
    * 12 async machines, each machine with 8 GPUs (K80) on which the model is spread X 6days
    * [dropout](http://www.shortscience.org/paper?bibtexKey=journals/corr/ZarembaSV14) 0.2-0.3 (higher for smaller datasets)
  * RL - [Reinforcement Learning](http://www.shortscience.org/paper?bibtexKey=journals/corr/RanzatoCAZ15) 
    * sequence score, $\text{GLEU} = r = \min(\text{precision}, \text{recall})$ computed on n-grams of size 1-4
    * mixed loss $\alpha \text{ML} + \text{RL}, \alpha =0.25$
    * mean $r$ computed from $m=15$ samples
    * SGD, 400K steps, 3 days, no drouput
* Prediction (i.e. Decoder)
  * beam search (3 beams)
  * A normalized score is computed to every beam that ended (died)
    * did not normalize beam score by $\text{beam_length}^\alpha , \alpha \in [0.6-0.7]$
    * normalized with similar formula in which 5 is add to length and a coverage factor is added, which is the sum-log of attention weight of every input word (i.e. after summing over all output words)
    * Do a second pruning using normalized scores

Authors investigated why humans play some video games better than machines. That is the case for games that do not have continuous rewards (e.g., scores). They experimented with a game -- inspired by _Montezuma's Revenge_ -- in which the player has to climb stairs, collect keys and jump over enemies. RL algorithms can only know if they succeed if they finish the game, as there is no rewards during the gameplay, so they tend to do much worse than humans in these games.

To compare between humans and machines, they set up RL algorithms and recruite players from Amazon Mechanical Turk. Humans did much better than machines for the original game setup. However, authors wanted to check the impact of semantics and prior knowledge on humans performance. They set up scenarios with different levels of reduced semantic information, as shown in Figure 2.

https://i.imgur.com/e0Dq1WO.png

This is what the game originally looked like:

https://rach0012.github.io/humanRL_website/main.gif

And this is the version with lesser semantic clues:

https://rach0012.github.io/humanRL_website/random2.gif

You can try yourself in the [paper's website](https://rach0012.github.io/humanRL_website/).

Not surprisingly, humans took much more time to complete the game in scenarios with less semantic information, indicating that humans strongly rely on prior knowledge to play video games.

The authors argue that this prior knowledge should also be somehow included into RL algorithms in order to move their efficiency towards the human level.

## Additional reading

[Why humans learn faster than AI—for now](https://www.technologyreview.com/s/610434/why-humans-learn-faster-than-ai-for-now/).

[OpenReview submission](https://openreview.net/forum?id=Hk91SGWR-)

This paper focuses on taking advances from the (mostly heuristic, practical) world of data augmentation for supervised learning, and applying those to the unsupervised setting, as a way of inducing better performance in a semi-supervised environment (with many unlabeled points, and few labeled ones) 

Data augmentation has been a mostly behind-the-scenes implementation detail in modern deep learning: minor modifications like shifting a dataset by a few pixels, rotating it slightly, or flipping it horizontally, to generate additional pseudoexamples for the model to train on. The core idea motivating such approaches is that the tactics of data augmentation are modifications that should not change the class label in a platonic, ground truth sense, which allows us to use them as training examples of the same class label as the original image from which the perturbations were made. In addition to just giving then network generically more data, this approach communicates to the network the specific kinds of modifications that can be made to an image and have it still be of the same class. 

The Unsupervised Data Augmentation (UDA) tactic from this paper notices two things: 
(1) Within the sphere of supervised learning, there has been dataset-specific innovation in generating augmented data that will be particularly helpful to a given dataset. Inlanguage modeling, an example of this is having a sentence go into another language and back again through two well-trained translation networks, and use the resulting sentence as another example of the same class. For ImageNet, there’s an approach called AutoAugment that uses reinforcement learning on a validation set to learn a policy of image operations (rotate, shear, change color) in order to increase validation accuracy. [I am a little confused about this as an approach, since I worry about essentially overfitting to the validation set. That said, I don’t have time to delve specifically into the AutoAugment paper today, so I’ll just leave this here as a caveat] 

(2) Within semi-supervised learning, there’s a growing tendency to use consistency loss as a way of making use of unlabeled data. The basic idea of consistency loss is that, even if you don’t know the class of a given datapoint, if you modify it in some small way, you can be confident that the model’s prediction should be consistent between the point and its perturbation, even if you don’t have knowledge of what the actual ground truth is. Often, systems like this have been designed using simple Gaussian noise on top of original unlabeled images. The key proposal of this paper is to substitute this more simplified perturbation procedure with the augmentation approaches being iterated on in supervised learning, since the goal of both - to modify inputs so as to capture ways in which they might differ but still be notionally of the same class - is nearly identical 

On top of this core idea, the UDA paper also proposes an additional clever training tactic: in cases where you have many unlabeled examples and few labeled ones, you may need a large model to capture the information from the unlabeled examples, but this may result in overfitting on the labeled ones. To avoid this, they use an approach called “Training Signal Annealing,” where at each point in training they remove from the loss calculation any examples that the model is particularly confident about: where the prediction of the true class is above some threshold eta. As training goes on, the network is gradually allowed to see more of the training signals. In this kind of a framework, the model can’t overfit as easily, because once it starts getting the right answer on supervised examples, they drop out of the loss calculation. 

In terms of empirical results, the authors find that in UDA they’re able to improve on many semi-supervised benchmarks with quite small numbers of labeled examples. At one point, they use as a baseline a BERT model that was fine-tuned in an unsupervised way prior to its semi-supervised training, and show that their augmentation method can add value even on top of the value that the unsupervised pre-training adds. 

Reinforcement Learning is often broadly separated into two categories of approaches: model-free and model-based. In the former category, networks simply take observations and input and produce predicted best-actions (or predicted values of available actions) as output. In order to perform well, the model obviously needs to gain an understanding of how its actions influence the world, but it doesn't explicitly make predictions about what the state of the world will be after an action is taken. In model-based approaches, the agent explicitly builds a dynamics model, or a model in which it takes in (past state, action) and predicts next state. In theory, learning such a model can lead to both interpretability (because you can "see" what the model thinks the world is like) and robustness to different reward functions (because you're learning about the world in a way not explicitly tied up with the reward). 

This paper proposes an interesting melding of these two paradigms, where an agent learns a model of the world as part of an end-to-end policy learning. This works through something the authors call "observational dropout": the internal model predicts the next state of the world given the prior one and the action, and then with some probability, the state of the world that both the policy and the next iteration of the dynamics model sees is replaced with the model's prediction. This incentivizes the network to learn an effective dynamics model, because the farther the predictions of the model are from the true state of the world, the worse the performance of the learned policy will be on the iterations where the only observation it can see is the predicted one. So, this architecture is model-free in the sense that the gradient used to train the system is based on applying policy gradients to the reward, but model-based in the sense that it does have an internal world representation. 

https://i.imgur.com/H0TNfTh.png

The authors find that, at a simple task, Swing Up Cartpole, very low probabilities of seeing the true world (and thus very high probabilities of the policy only seeing the dynamics model output) lead to world models good enough that a policy trained only on trajectories sampled from that model can perform relatively well. This suggests that at higher probabilities of the true world, there was less value in the dynamics model being accurate, and consequently less training signal for it. (Of course, policies that often could only see the predicted world performed worse during their original training iteration compared to policies that could see the real world more frequently).

 On a more complex task of CarRacing, the authors looked at how well a policy trained using the representations of the world model as input could perform, to examine whether it was learning useful things about the world. 
https://i.imgur.com/v9etll0.png
They found an interesting trade-off, where at high probabilities (like before) the dynamics model had little incentive to be good, but at low probabilities it didn't have enough contact with the real dynamics of the world to learn a sensible policy.

We want to find two matrices $W$ and $H$ such that $V = WH$. Often a goal is to determine underlying patterns in the relationships between the concepts represented by each row and column. $W$ is some $m$ by $n$ matrix and we want the inner dimension of the factorization to be $r$. So 

$$\underbrace{V}_{m \times n} = \underbrace{W}_{m \times r} \underbrace{H}_{r \times n}$$

Let's consider an example matrix where of three customers (as rows) are associated with three movies (the columns) by a rating value.

$$
V = \left[\begin{array}{c c c}
5 & 4 & 1  \\\\
4 & 5 & 1 \\\\
2 & 1 & 5
\end{array}\right]
$$


We can decompose this into two matrices with $r = 1$. First lets do this without any non-negative constraint using an SVD reshaping matrices based on removing eigenvalues:


$$
W = \left[\begin{array}{c c c}
-0.656 \\\
 -0.652 \\\
 -0.379
\end{array}\right],
H = \left[\begin{array}{c c c}
-6.48 & -6.26 & -3.20\\\\
\end{array}\right]
$$

We can also decompose this into two matrices with $r = 1$ subject to the constraint that $w_{ij} \ge 0$ and  $h_{ij} \ge 0$. (Note: this is only possible when $v_{ij} \ge 0$):

$$
W = \left[\begin{array}{c c c}
0.388 \\\\
0.386 \\\\
0.224
\end{array}\right],
H = \left[\begin{array}{c c c}
11.22 & 10.57 & 5.41  \\\\
\end{array}\right]
$$

Both of these $r=1$ factorizations reconstruct matrix $V$ with the same error. 

$$
V \approx WH = \left[\begin{array}{c c c}
4.36 & 4.11 & 2.10 \\\
4.33 & 4.08 & 2.09 \\\
2.52 & 2.37 & 1.21 \\\
\end{array}\right]
$$


If they both yield the same reconstruction error then why is a non-negativity constraint useful? We can see above that it is easy to observe patterns in both factorizations such as similar customers and similar movies. `TODO: motivate why NMF is better`



#### Paper Contribution 

This paper discusses two approaches for iteratively creating a non-negative $W$ and $H$ based on random initial matrices. The paper discusses a multiplicative update rule where the elements of $W$ and $H$ are iteratively transformed by scaling each value such that error is not increased. 

The multiplicative approach is discussed in contrast to an additive gradient decent based approach where small corrections are iteratively applied. The multiplicative approach can be reduced to this by setting the learning rate ($\eta$) to a ratio that represents the magnitude of the element in $H$ to the scaling factor of $W$ on $H$.



### Still a draft