This paper considers the problem of structured output prediction, in the specific case where the output is a sequence and we represent the sequence as a (conditional) directed graphical model that generates from the first token to the last. The paper starts from the observation that training such models by maximum likelihood (ML) does not reflect well how the model is actually used at test time. Indeed, ML training implies that the model is effectively trained to predict each token conditioned on the previous tokens *from the ground truth* sequence (this is known as "teacher forcing"). Yet, when making a prediction for a new input, the model will actually generate a sequence by generating tokens one after another and conditioning on *its own predicted tokens* instead. 

So the authors propose a different training procedure, where at training time each *conditioning* ground truth token is sometimes replaced by the model's previous prediction. The choice of replacing the ground truth by the model's prediction is made by "flipping a coin" with some probability, independently for each token. Importantly, the authors propose to start with a high probability of using the ground truth (i.e. start close to ML) and anneal that probability closer to 0, according to some schedule (thus the name Schedule Sampling). 

Experiments on 3 tasks (image caption generation, constituency parsing and speech recognition) based on neural networks with LSTM units, demonstrate that this approach indeed improves over ML training in terms of the various performance metrics appropriate for each problem, and yields better sequence prediction models. 



#### My two cents

Big fan of this paper. It both identifies an important flaw in how sequential prediction models are currently trained and, most importantly, suggests a solution that is simple yet effective. I also believe that this approach played a non-negligible role in Google's winner system for image caption generation, in the Microsoft COCO competition. 

My alternative interpretation of why Scheduled Sampling helps is that ML training does not inform the model about the relative quality of the errors it can make. In terms of ML, it is as bad to put high probability on an output sequence that has just 1 token that's wrong, than it is to put the same amount of probability on a sequence that has all tokens wrong. Yet, say for image caption generation, outputting a sentence that is one word away from the ground truth is clearly preferable from making a mistake on a words (something that is also reflected in the performance metrics, such as BLEU). 

By training the model to be robust to its own mistakes, Scheduled Sampling ensures that errors won't accumulate and makes predictions that are entirely off much less likely.

An alternative to Scheduled Sampling is DAgger (Dataset Aggregation: \cite{journals/jmlr/RossGB11}), which briefly put alternates between training the model and adding to the training set examples that mix model predictions and the ground truth. However, Scheduled Sampling has the advantage that there is no need to explicitly create and store that increasingly large dataset of sampled examples, something that isn't appealing for online learning or learning on large datasets.

I'm also very curious and interested by one of the direction of future work mentioned in the conclusion: figuring out a way to backprop through the stochastic predictions made by the model. Indeed, as the authors point out, the current algorithm ignores the fact that, by sometimes taking as input its previous prediction, this induces an additional relationship between the model's parameters and its ultimate prediction, a relationship that isn't taken into account during training. To take it into account, you'd need to somehow backpropagate through the stochastic process that generated the previous token prediction. While the work on variational autoencoders has shown that we can backprop through gaussian samples, backpropagating through the sampling of a discrete multinomial distribution is essentially an open problem. I do believe that there is work that tried to tackle propagating through stochastic binary units however, so perhaps that's a start. Anyways, if the authors could make progress on that specific issue, it could be quite useful not just in the context of Schedule Sampling, but possibly in the context of training networks with discrete stochastic units in general!

This paper proposes a method to send messages between cell phones over Bluetooth by using the device name field. This allows devices to communicate directly with each other without pairing. 

Miyato et al. propose distributional smoothing (or virtual adversarial training) as defense against adversarial examples. However, I think that both terms do not give a good intuition of what is actually done. Essentially, a regularization term is introduced. Letting $p(y|x,\theta)$ be the learned model, the regularizer is expressed as

$\text{KL}(p(y|x,\theta)|p(y|x+r,\theta)$

where $r$ is the perturbation that maximizes the Kullback-Leibler divergence above, i.e.

$r = \arg\max_r \{\text{KL}(p(y|x,\theta)|p(y|x+r,\theta) | \|r\|_2 \leq \epsilon\}$

with hyper-parameter $\epsilon$. Essentially, the regularizer is supposed to “simulate” adversarial training – thus, the method is also called virtual adversarial training.

The discussed implementation, however, is somewhat cumbersome. In particular, $r$ cannot be computed using first-order methods as the gradient of $\text{KL}$ is $0$ for $r = 0$. So a second-order method is used – for which the Hessian needs to be approximated and the corresponding eigenvectors need to be computed. For me it is unclear why $r$ cannot be initialized randomly to solve this issue … Then, the derivative of the regularizer needs to be computed during training. Here, the authors make several simplifications (such as fixing $\theta$ in the first part of the Kullback-Leibler divergence and ignoring the derivative of $r$ w.r.t. $\theta$).

Overall, however, I like the idea of “virtual” adversarial training as it avoids the need of explicitly using attacks during training to craft adversarial examples. Then, the trained model is often robust against the chosen attacks, but new adversarial examples can be found easily through novel attacks.

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

### Main Idea:
It basically tunes the hyper-parameters of the neural network architecture using reinforcement learning. The reward signal is taken as evaluation on the validation set. The method is policy gradient as the cost function is non-differentiable. 

### Method:
 #### i.   Actions:
1. There is controller RNN which predicts some hyper-parameters of the layer conditioned on the previous predictions. This prediction is just a one-hot vector based on the previous hyper-parameter chosen. At the start for the first prediction - this vector is just all zeros.
2. Once the network is generated completely, it is trained for a fixed number of epochs, the reward signal is calculated based on the evaluation on the validation set.

#### ii.  Training:
1. Nothing fancy in the reinforcement learning approach simple policy gradients.
2. Baseline is added to reduce the variance.
3.  It takes 2-3 weeks to train it over 800 GPUs!!

#### iii. Results:
1. Use it to generate CNNs and LSTM cells. Close to state-of-art results with generated architectures.

### Possible new directions:
1. Use better techniques RL techniques like TRPO, PPO etc.
2. Right now, they generate fixed length architecture. Their reason is for variable length architectures, it is difficult to determine how much time each architecture is trained. Smaller networks are easier to train. Thus, somehow determine training time as function of the learning capacity of the network.


Code:
They haven't released the code yet. I tried to simulate it in torch.(https://github.com/abhigenie92/nn_search)

Last year, a machine translation paper came out, with an unfortunately un-memorable name (the Transformer network) and a dramatic proposal for sequence modeling that eschewed both Recurrent NNN and Convolutional NN structures, and, instead, used self-attention as its mechanism for “remembering” or aggregating information from across an input. Earlier this month, the same authors released an extension of that earlier paper, called Image Transformer, that applies the same attention-only approach to image generation, and also achieved state of the art performance there. 

The recent paper offers a framing of attention that I find valuable and compelling, and that I’ll try to explicate here. They describe attention as being a middle ground between the approaches of CNNs and RNNs, and one that, to use an over-abused cliche, gets the best of both worlds.  CNNs are explicitly local: each convolutional filter only gathers information from the cells that fall in specific locations along some predefined grid. And, because convolutional filters have a unique parameter for every relative location in the grid they’re applied to, increasing the size of any given filter’s receptive field would engender an exponential increase in parameters: to go from a 3x3 grid to a 4x4 one, you go from 9 parameters to 16.  Convolutional networks typically increase their receptive field through the mechanism of adding additional layers, but there is still this fundamental limitation that for a given number of layers, CNNs will be fairly constrained in their receptive field. 

On the other side of the receptive field balance, we have RNNs. RNNs have an effectively unlimited receptive field, because they just apply one operation again and again: take in a new input, and decide to incorporate that information into the hidden state. This gives us the theoretical ability to access things from the distant past, because they’re stored somewhere in the hidden state. However, each element is only seen once and needs to be stored in the hidden state in a way that sort of “averages over” all of the ways it’s useful for various points in the decoding/translation process. (My mental image basically views RNN hidden state as packing for a long trip in a small suitcase: you have to be very clever about what you decide to pack, averaging over all the possible situations you might need to be prepared for. You can’t go back and pull different things into your suitcase as a function of the situation you face; you had to have chosen to add them at the time you encountered them). All in all, RNNs are tricky both because they have difficulty storing information efficiently over long time frames, and also because they can be monstrously slow to train, since you have to run through the full sequence to built up hidden state, and can’t chop it into localized bits the way you can with CNNs. 

So, between CNN - with its locally-specific hidden state - and RNN - with its large receptive field but difficulty in information storage - the self-attention approach interposes itself. Attention works off of three main objects: a query, and a set of keys, each one is attached to a value. In general, all of these objects take the form of vectors. For a given query, you calculate its similarity with each key, and then normalize those into a distribution (a set of weights, all of which sum to 1) that is used as the weights in calculating a weighted average of the values. As a motivating example, think of a model that is “unrolling” or decoding a translated sentence. In order to translate a sentence properly, the model needs to “remember” not only the conceptual content of the sentence, but what it has already generated. So, at each given point in the unrolling, the model can “query” the past and get a weighted distribution over what’s relevant to it in its current context. In the original Transformer, and also in the new one, the models use “multi-headed attention”, which I think is best compared to convolution filters: in the same way that you learn different convolution filters, each with different parameters, to pick up on different features, you learn different “heads” of the attention apparatus for the same purpose. 

To go back to our CNN - Attention - RNN schematic from earlier: Attention makes it a lot easier to query a large receptive field, since you don’t need an additional set of learned parameters for each location you expand to; you just use the same query weights and key weights you use for every other key and query. And, it allows you to contextually extract information from the past, depending on the needs you have right now. That said, it’s still the case that it becomes infeasible to make the length of the past you calculate your attention distribution over excessively long, but that cost is in terms of computation, not additional parameters, and thus is a question of training time, rather than essential model complexity, the way additional parameters is. 

Jumping all the way back up the stack, to the actual most recent image paper, this question of how best to limit the receptive field is one of the more salient questions, since it still is the case that conducting attention over every prior pixel would be a very large number of calculations. The Image Transformer paper solves this in a slightly hacky way: by basically subdividing the image into chunks, and having each chunk operate over the same fixed memory region (rather than scrolling the memory region with each pixel shift) to take better advantage of the speed of batched big matrix multiplies. 

Overall, this paper showed an advantage for the Image Transformer approach relevative to PixelCNN autoregressive generation models, and cited the ability for a larger receptive field during generation - without explosion in number of parameters - as the most salient reason why.