Recently, DeepMind released a new paper showing strong performance on board game tasks using a mechanism similar to the Value Prediction Network one in this paper, which inspired me to go back and get a grounding in this earlier work. 

A goal of this paper is to design a model-based RL approach that can scale to complex environment spaces, but can still be used to run simulations and do explicit planning. Traditional, model-based RL has worked by learning a dynamics model of the environment - predicting the next observation state given the current one and an action, and then using that model of the world to learn values and plan with. In addition to the advantages of explicit planning, a hope is that model-based systems generalize better to new environments, because they predict one-step changes in local dynamics in a way that can be more easily separated from long-term dynamics or reward patterns. 

However, a downside of MBRL is that it can be hard to train, especially when your observation space is high-dimensional, and learning a straight model of your environment will lead to you learning details that aren't actually unimportant for planning or creating policies. 

The synthesis proposed by this paper is the Value Prediction Network. Rather than predicting observed state at the next step, it learns a transition model in latent space, and then learns to predict next-step reward and future value from that latent space vector. Because it learns to encode latent-space state from observations, and also learns a transition model from one latent state to another, the model can be used for planning, by simulating multiple transitions between latent state. However, unlike a normal dynamics model, whose training signal comes from a loss against observational prediction, the signal for training both latent → reward/value/discount predictions, and latent → latent transitions comes from using this pipeline to predict reward values. This means that if an aspect of the environment isn't useful for predicting reward, it won't generally be encoded into latent state, meaning you don't waste model capacity predicting irrelevant detail. 

https://i.imgur.com/4bJylms.png

Once this model exists, it can be used for generating a policy through a tree-search planning approach: simulating future trajectories and aggregating the predicted reward along those trajectories, and then taking the highest-value one. 

The authors find that their model is able to do better than both model-free and model-based methods on the tasks they tested on. In particular, they find that it has many of the benefits of a model that predicts full observations, but that the Value Prediction Network learns more quickly, and is more robust to stochastic environments where there's an inherent ceiling on how well a next-step observation prediction can work. 

My main question coming into this paper is: how is this different from simply a value estimator like those used in DQN or A2C, and my impression is that the difference comes from this model's ability to do explicit state simulation in latent space, and then predict a value off of the *latent* state, whereas a value network predicts value from observational state.

# Deep Convolutional Generative Adversarial Nets

## Introduction

* The paper presents Deep Convolutional Generative Adversarial Nets (DCGAN) - a topologically constrained variant of conditional GAN.
* [Link to the paper](https://arxiv.org/abs/1511.06434)

## Benefits

* Stable to train
* Very useful to learn unsupervised image representations.

## Model

* GANs difficult to scale using CNNs.
* Paper proposes following changes to GANs:
    * Replace any pooling layers with strided convolutions (for discriminator) and fractional strided convolutions (for generators).
    * Remove fully connected hidden layers.
    * Use batch normalisation in both generator (all layers except output layer) and discriminator (all layers except input layer).
    * Use LeakyReLU in all layers of the discriminator.
    * Use ReLU activation in all layers of the generator (except output layer which uses Tanh).

## Datasets
    
* Large-Scale Scene Understanding.
* Imagenet-1K.
* Faces dataset.

## Hyperparameters

* Minibatch SGD with minibatch size of 128.
* Weights initialized with 0 centered Normal distribution with standard deviation = 0.02
* Adam    Optimizer
* Slope of leak = 0.2 for LeakyReLU.
* Learning rate = 0.0002, β1 = 0.5

## Observations

* Large-Scale Scene Understanding data
    * Demonstrates that model scales with more data and higher resolution generation.
    * Even though it is unlikely that model would have memorized images (due to low learning rate of minibatch SGD).
* Classifying CIFAR-10 dataset
    * Features
        * Train in Imagenet-1K and test on CIFAR-10.
        * Max pool discriminator's convolutional features (from all layers) to get 4x4 spatial grids.
        * Flatten and concatenate to get a 28672-dimensional vector.
        * Linear L2-SVM classifier trained over the feature vector.
    * 82.8% accuracy, outperforms K-means (80.6%)
* Street View House Number Classifier
    * Similar pipeline as CIFAR-10
    * 22.48% test error.
* The paper contains many examples of images generated by final and intermediate layers of the network.
* Images in the latent space do not show sharp transitions indicating that network did not memorize images.
* DCGAN can learn an interesting hierarchy of features.
* Networks seems to have some success in disentangling image representation from object representation.
* Vector arithmetic can be performed on the Z vectors corresponding to the face samples to get results like `smiling woman - normal woman + normal man = smiling man` visually.

TLDR; The authors adopt Generative Adversarial Networks (GANs) to RNNs and train a discriminator to distinguish between sequences generated using teacher forcing (feeding ground truth inputs to the RNN) and scheduled sampling (feeding generated outputs as the next inputs). The inputs to the discriminator are both the predictions and the hidden states of the generative RNN. The generator is trained to fool the discriminator, forcing the dynamics of teacher forcing and scheduled sampling to become more similar. This procedure acts as regularizer, and results in better sample quality and generalization, particularly for long sequences. The authors evaluate their framework on Language Model (PTB), Pixel Generation (Sequential MNIST), Handwriting Generation, and Musisc Synthesis.

### Key Points

- Problem: During inference, errors in an RNN easily compound because the conditioning context may diverge from what is seen during training when the ground-truth labels are fed as inputs (teacher forcing).
- Goal of professor forcing: Make the generative (free-run) behavior and the teacher-forced behavior match as closely as possible.
- Discriminator Details
  - Input is a behavior sequence `B(x, y, theta)` from the generative RNN that contains the hidden states and outputs.
  - The training objective is to correctly classify whether or not a behavior sequence is generated using teacher forcing vs. scheduled sampling.
- Generator
  - Standard RNN with MLE training objective and an additional term to fool the discrimator: Change the free-running behavior as to match the teacher-forced behavior while keeping the latter constant.
  - Second optional another term: Change the teacher-forced behavior to match the free-running behavior.
  - Like GAN, backprop from discriminator into generator.
- Architectures
  - Generator is a standard GRU Recurrent Neural Network with softmax
  - Behavior function `B(x, y, theta)` outputs pre-tanh activation of GRU states and tje softmax output
  - Discriminator: Bidirectional GRU with 3-layer MLP on top
  - Training trick: To prevent "bad gradients" the authors backprop from the discriminator into the generator only if the classification accuracy is between 75% and 99%.
  - Trained used Adam optimizer
- Experiments
  - PTB Chracter-Level Modeling: Reduction in test NLL, profesor forcing seem to act as a regularizier. 1.48 BPC
  - Sequential MNIST: Second-best NLL (79.58) after PixelCNN
  - Handwriting generation: Professor forcing is better at generating longer sequences than seen during training as per human eval.
  - Music Synthesis: Human eval significantly better for Professor forcing
  - Negative Results on word-level modeling: Professor forcing doesn't have any effect. Perhaps because long-term dependencies are more pronounced in character-level modeling.
  - The authors show using t-SNE that the hidden state distributions actually become more similar when using professor forcing

### Thoughts

- Props to the authors for a very clear and well-written paper. This is rarer than it should be :)
- It's an intersting idea to also match the states of the RNN instead of just the outputs. Intuitively, matching the outputs should implicitly match the state distribution. I wonder if the authors tried this and it didn't work as expected.
- Note from [Ethan Caballero](https://github.com/ethancaballero) about why they chose to match hidden states: It's significantly harder to use GANs on sampled (argmax) output tokens because they are discrete as (as opposed to continuous like the hidden states and their respective softmaxes). They would have had to estimate discrete outputs with policy gradients like in [seqGAN](https://github.com/dennybritz/deeplearning-papernotes/blob/master/notes/seq-gan.md) which is [harder to get to converge](https://www.quora.com/Do-you-have-any-ideas-on-how-to-get-GANs-to-work-with-text), which is why they probably just stuck with the hidden states which already contain info about the discrete sampled outputs (the index of the highest probability in the the distribution) anyway. Professor Forcing method is unique in that one has access to the continuous probability distribution of each token at each timestep of the two sequence generation modes trying to be pushed closer together. Conversely, when applying GANs to pushing real samples and generated samples closer together as is traditionally done in models like seqGAN, one only has access to the next dicrete token (not continuous probability distributions of next token) at each timestep, which prevents straight-forward differentiation (used in professor forcing) from being applied and forces one to use policy gradient estimation. However, there's a chance one might be able to use straight-forward differentiation to train seqGANs in the traditional sampling case if one swaps out each discrete sampled token with its continuous distributional word embedding (from pretrained word2vec, GloVe, etc.), but no one has tried it yet TTBOMK.
- I would've liked to see a comparison of  the two regularization terms in the generator. The experiments don't make it clear if both or only of them them is used.
- I'm guessing that this architecture is quite challenging to train. Woul've liked to see a bit more detail about when/how they trade off the training of discriminator and generator.
- Translation is another obvious task to apply this too. I'm interested whether or not this works for seq2seq.

In this article, the authors provide a framework for training two translation models with large accessible monolingual corpus.

In traditional methods, machine translation models always require large parallel corpus to train a good quality model, which is expensive to acquire. However, the massive monolingual data is not fully utilized. The monolingual corpus are typically used in pretraining the NMT decoder rnn and augmenting initial parallel corpus through self-generated translations.

The authors embed machine translation task into a reinforcement learning framework, in which two agents act as two different native speakers respectively and know little about each other and then they learn to translate by trying to communicate with each other. 

**The two speakers**, `A` and `B`, obviously know well about their corresponding language respectively, this situation is easily simulated by two well-trained language models for `A` and `B`. Then, speaker `A` tries to tell a sentence $x$ to `B` by translating it into $y$ in `B`'s language. Since they don't know each other, `B` is uncertain about what `A` truly means by saying $y$. However, `B` is capable of evaluate the degree of sensibility of $y$  from his own understanding. Next, `B` informs `A` his sensibility evaluation score and tries to recover what `A` truly means in `A`'s language, i.e. $x'$. And similarly, `A` can also evaluate the degree of sensibility of $x'$  from his own understanding.

In general, the very original idea that `A` tried to convey, is passed through a noisy channel to `B`, and then back to `A` through another noisy channel. The former noisy channel is a `A-B` translation model and the latter a `B-A` translation model in the framework.

Think about how the first American learnt Chinese in history and I think it is intuitively similar to the principle in this work.

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.