This paper presents a recurrent neural network architecture in which some of the recurrent weights dynamically change during the forward pass, using a hebbian-like rule. They correspond to the matrices $A(t)$ in the figure below:

![Fast weights RNN figure](http://i.imgur.com/DCznSf4.png)

These weights $A(t)$ are referred to as *fast weights*. Comparatively, the recurrent weights $W$ are referred to as slow weights, since they are only changing due to normal training and are otherwise kept constant at test time.

More specifically, the proposed fast weights RNN compute a series of hidden states $h(t)$ over time steps $t$, but, unlike regular RNNs, the transition from $h(t)$ to $h(t+1)$ consists of multiple ($S$) recurrent layers $h_1(t+1), \dots, h_{S-1}(t+1), h_S(t+1)$, defined as follows:

$$h_{s+1}(t+1) = f(W h(t) + C x(t) + A(t) h_s(t+1))$$

where $f$ is an element-wise non-linearity such as the ReLU activation. The next hidden state $h(t+1)$ is simply defined as the last "inner loop" hidden state $h_S(t+1)$, before moving to the next time step. 

As for the fast weights $A(t)$, they too change between time steps, using the hebbian-like rule:

$$A(t+1) = \lambda A(t) + \eta h(t) h(t)^T$$

where $\lambda$ acts as a decay rate (to partially forget some of what's in the past)  and $\eta$ as the fast weight's "learning rate" (not to be confused with the learning rate used during backprop). Thus, the role played by the fast weights is to rapidly adjust to the recent hidden states and remember the recent past.

In fact, the authors show an explicit relation between these fast weights and memory-augmented architectures that have recently been popular. Indeed, by recursively applying and expending the equation for the fast weights, one obtains

$$A(t) = \eta \sum_{\tau = 1}^{\tau = t-1}\lambda^{t-\tau-1} h(\tau) h(\tau)^T$$

*(note the difference with Equation 3 of the paper... I think there was a typo)* which implies that when computing the $A(t) h_s(t+1)$ term in the expression to go from $h_s(t+1)$ to $h_{s+1}(t+1)$, this term actually corresponds to

$$A(t) h_s(t+1) = \eta \sum_{\tau =1}^{\tau = t-1} \lambda^{t-\tau-1} h(\tau) (h(\tau)^T h_s(t+1))$$

i.e. $A(t) h_s(t+1)$ is a weighted sum of all previous hidden states $h(\tau)$, with each hidden states weighted by an "attention weight" $h(\tau)^T h_s(t+1)$. The difference with many recent memory-augmented architectures is thus that the attention weights aren't computed using a softmax non-linearity.

Experimentally, they find it beneficial to use [layer normalization](https://arxiv.org/abs/1607.06450). Good values for $\eta$ and $\lambda$ seem to be 0.5 and 0.9 respectively. I'm not 100% sure, but I also understand that using $S=1$, i.e. using the fast weights only once per time steps, was usually found to be optimal. Also see Figure 3 for the architecture used on the image classification datasets, which is slightly more involved.

The authors present a series 4 experiments, comparing with regular RNNs (IRNNs, which are RNNs with ReLU units and whose recurrent weights are initialized to a scaled identity matrix) and LSTMs (as well as an associative LSTM for a synthetic associative retrieval task and ConvNets for the two image datasets). Generally, experiments illustrate that the fast weights RNN tends to train faster (in number of updates) and better than the other recurrent architectures. Surprisingly, the fast weights RNN can even be competitive with a ConvNet on the two image classification benchmarks, where the RNN traverses glimpses from the image using a fixed policy.

**My two cents**

This is a very thought provoking paper which, based on the comparison with LSTMs, suggests that fast weights RNNs might be a very good alternative. I'd be quite curious to see what would happen if one was to replace LSTMs with them in the myriad of papers using LSTMs (e.g. all the Seq2Seq work). Intuitively, LSTMs seem to be able to do more than just attending to the recent past. But, for a given task, if one was to observe that fast weights RNNs are competitive to LSTMs, it would suggests that the LSTM isn't doing something that much more complex. So it would be interesting to determine what are the tasks where the extra capacity of an LSTM is actually valuable and exploitable. Hopefully the authors will release some code, to facilitate this exploration. 

The discussion at the end of Section 3 on how exploiting the "memory augmented" view of fast weights is useful to allow the use of minibatches is interesting. However, it also suggests that computations in the fast weights RNN scales quadratically with the sequence size (since in this view, the RNN technically must attend to all previous hidden states, since the beginning of the sequence). This is something to keep in mind, if one was to consider applying this to very long sequences (i.e. much longer than the hidden state dimensionality).

Also, I don't quite get the argument that the "memory augmented" view of fast weights is more amenable to mini-batch training. I understand that having an explicit weight matrix $A(t)$ for each minibatch sequence complicates things. However, in the memory augmented view, we also have a "memory matrix" that is different for each sequence, and yet we can handle that fine. The problem I can imagine is that storing a *sequence of arbitrary weight matrices* for each sequence might be storage demanding (and thus perhaps make it impossible to store a forward/backward pass for more than one sequence at a time), while the implicit memory matrix only requires appending a new row at each time step. Perhaps the argument to be made here is more that there's already mini-batch compatible code out there for dealing with the use of a memory matrix of stored previous memory states.

This work strikes some (partial) resemblance to other recent work, which may serve as food for thought here. The use of possibly multiple computation layers between time steps reminds me of [Adaptive Computation Time (ACT) RNN]( http://www.shortscience.org/paper?bibtexKey=journals/corr/Graves16). Also, expressing a backpropable architecture that involves updates to weights (here, hebbian-like updates) reminds me of recent work that does backprop through the updates of a gradient descent procedure (for instance as in [this work]( http://www.shortscience.org/paper?bibtexKey=conf/icml/MaclaurinDA15)). 

Finally, while I was familiar with the notion of fast weights from the work on [Using Fast Weights to Improve Persistent Contrastive Divergence](http://people.ee.duke.edu/~lcarin/FastGibbsMixing.pdf), I didn't realize that this concept dated as far back as the late 80s. So, for young researchers out there looking for inspiration for research ideas, this paper confirms that looking at the older neural network literature for inspiration is probably a very good strategy :-)

To sum up, this is really nice work, and I'm looking forward to the NIPS 2016 oral presentation of it!

Moosavi-Dezfooli et al. propose universal adversarial perturbations – perturbations that are image-agnostic. Specifically, they extend the framework for crafting adversarial examples, i.e. by iteratively solving

$\arg\min_r \|r \|_2$ s.t. $f(x + r) \neq f(x)$.

Here, $r$ denotes the adversarial perturbation, $x$ a training sample and $f$ the neural network. Instead of solving this problem for a specific $x$, the authors propose to solve the problem over the full training set, i.e. in each iteration, a different sample $x$ is chosen, one step in the direction of the gradient is taken and the perturbation is updated accordingly. In experiments, they show that these universal perturbations are indeed able to fool networks an several images; in addition, these perturbations are – sometimes – transferable to other networks.

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

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.

This paper presents a way to differentiate through discrete random variables by replacing them with continuous random variables. Say you have a discrete [categorical variable][cat] and you're sampling it with the [Gumbel trick][gumbel] like this ($G_k$ is a Gumbel distributed variable and $\boldsymbol{\alpha}/\sum_k \alpha_k$ are our categorical probabilities):

$$
z = \text{one_hot} \left( \underset{k}{\text{arg max}} [ G_k + \log \alpha_k ] \right)
$$

This paper replaces the one hot and argmax with a softmax, and they add a $\lambda$ variable to control the "temperature". As $\lambda$ tends to zero it will equal the above equation.

$$
z = \text{softmax} \left( \frac{  G_k + \log \alpha_k }{\lambda} \right)
$$

I made [some notes][nb] on how this process works, if you'd like more intuition.

Comparison to [Gumbel-softmax][gs]
--------------------------------------------

These papers are proposed precisely the same distribution with notation changes ([noted there][gs]). Both papers also reference each other and the differences. Although, they mention differences in the variatonal objectives to the Gumbel-softmax. This paper also compares to [VIMCO][], which is probably a harder benchmark to compare against (multi-sample versus single sample).

The results in both papers compare to SOTA score function based estimators and both report high scoring results (often the best). There are some details about implementations to consider though, such as scheduling and exactly how to define the variational objective.

[cat]: https://en.wikipedia.org/wiki/Categorical_distribution
[gumbel]: https://hips.seas.harvard.edu/blog/2013/04/06/the-gumbel-max-trick-for-discrete-distributions/
[gs]: http://www.shortscience.org/paper?bibtexKey=journals/corr/JangGP16
[nb]: https://gist.github.com/gngdb/ef1999ce3a8e0c5cc2ed35f488e19748
[vimco]: https://arxiv.org/abs/1602.06725