Kumar et al. propose an algorithm to learn in batch reinforcement learning (RL), a setting where an agent learns purely form a fixed batch of data, $B$, without any interactions with the environments. The data in the batch is collected according to a batch policy $\pi_b$. Whereas most previous methods (like BCQ) constrain the learned policy to stay close to the behavior policy, Kumar et al. propose bootstrapping error accumulation reduction (BEAR), which constrains the newly learned policy to place some probability mass on every non negligible action. 
The difference is illustrated in the picture from the BEAR blog post:
https://i.imgur.com/zUw7XNt.png
The behavior policy is in both images the dotted red line, the left image shows the policy matching where the algorithm is constrained to the purple choices, while the right image shows the support matching. 

**Theoretical Contribution:**
The paper analysis formally how the use of out-of-distribution actions to compute the target in the Bellman equation influences the back-propagated error.

Firstly a distribution constrained backup operator is defined as
$T^{\Pi}Q(s,a) = \mathbb{E}[R(s,a) + \gamma \max_{\pi \in \Pi} \mathbb{E}_{P(s' \vert s,a)} V(s')]$ and $V(s) = \max_{\pi \in \Pi} \mathbb{E}_{\pi}[Q(s,a)]$
which considers only policies $\pi \in \Pi$. 

It is possible that the optimal policy $\pi^*$ is not contained in the policy set $\Pi$, thus there is a suboptimallity constant $\alpha (\Pi) = \max_{s,a} \vert \mathcal{T}^{\Pi}Q^{*}(s,a) - \mathcal{T}Q^{*}(s,a) ]\vert $ which captures how far $\pi^{*}$ is from $\Pi$. 

Letting $P^{\pi_i}$ be the transition-matrix when following policy $\pi_i$, $\rho_0$ the state marginal distribution of the training data in the batch and $\pi_1, \dots, \pi_k \in \Pi $. The error analysis relies upon a concentrability assumption $\rho_0 P^{\pi_1} \dots P^{\pi_k} \leq c(k)\mu(s)$, with $\mu(s)$ the state marginal. Note that $c(k)$ might be infinite if the support of $\Pi$ is not contained in the state marginal of the batch. Using the coefficients $c(k)$ a concentrability coefficient is defined as:
$C(\Pi) = (1-\gamma)^2\sum_{k=1}^{\infty}k \gamma^{k-1}c(k).$
The concentrability takes values between 1 und $\infty$, where 1 corresponds to the case that the batch data were collected by $\pi$ and $\Pi = \{\pi\}$ and $\infty$ to cases where $\Pi$ has support outside of $\pi$. 

Combining this Kumar et a. get a bound of the Bellman error for distribution constrained value iteration with the constrained Bellman operator $T^{\Pi}$:

$\lim_{k \rightarrow \infty} \mathbb{E}_{\rho_0}[\vert V^{\pi_k}(s)- V^{*}(s)]  \leq  \frac{\gamma}{(1-\gamma^2)} [C(\Pi) \mathbb{E}_{\mu}[\max_{\pi \in \Pi}\mathbb{E}_{\pi}[\delta(s,a)] + \frac{1-\gamma}{\gamma}\alpha(\Pi)  ] ]$, where $\delta(s,a)$ is the Bellman error. 

This presents the inherent batch RL trade-off between keeping policies close to the behavior policy of the batch (captured by $C(\Pi)$ and keeping $\Pi$ sufficiently large (captured by $\alpha(\Pi)$). 

It is finally proposed to use support sets to construct $\Pi$, that is $\Pi_{\epsilon} = \{\pi \vert \pi(a \vert s)=0 \text{ whenever } \beta(a \vert s) < \epsilon \}$. This amounts to the set of all policies that place probability on all non-negligible actions of the behavior policy. For this particular choice of $\Pi = \Pi_{\epsilon}$ the concentrability coefficient can be bounded. 

**Algorithm**:
The algorithm has an actor critic style, where the Q-value to update the policy is taken to be the minimum over the ensemble. The support constraint to place at least some probability mass on every non negligible action from the batch is enforced via sampled MMD. The proposed algorithm is a member of the policy regularized algorithms as the policy is updated to optimize:
$\pi_{\Phi} = \max_{\pi} \mathbb{E}_{s \sim B} \mathbb{E}_{a \sim \pi(\cdot \vert s)} [min_{j = 1 \dots, k} Q_j(s,a)] s.t. \mathbb{E}_{s \sim B}[MMD(D(s), \pi(\cdot \vert s))] \leq \epsilon$

The Bellman target to update the Q-functions is computed as the convex combination of minimum and maximum of the ensemble. 

**Experiments**
The experiments use the Mujoco environments Halfcheetah, Walker, Hopper and Ant. Three scenarios of batch collection, always consisting of 1Mio. samples,  are considered:
- completely random behavior policy
- partially trained behavior policy
- optimal policy as behavior policy

The experiments confirm that BEAR outperforms other off-policy methods like BCQ or KL-control. The ablations show further that the choice of MMD is crucial as it is sometimes on par and sometimes substantially better than choosing KL-divergence. 

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!

The authors propose a unified setting to evaluate the performance of batch reinforcement learning algorithms. The proposed benchmark is discrete and based on the popular Atari Domain. The authors review and benchmark several current batch RL algorithms against a newly introduced version of BCQ (Batch Constrained Deep Q Learning) for discrete environments. 

https://i.imgur.com/zrCZ173.png

Note in line 5 that the policy chooses actions with a restricted argmax operation, eliminating actions that have not enough support in the batch. 

One of the key difficulties in batch-RL is the divergence of value estimates. In this paper the authors use Double DQN, which means actions are selected with a value net $Q_{\theta}$ and the policy evaluation is done with a target network $Q_{\theta'}$ (line 6). 

**How is the batch created?**
A partially trained DQN-agent (trained online for 10mio steps, aka 40mio frames) is used as behavioral policy to collect a batch $B$ containing 10mio transitions. The DQN agent uses either with probability 0.8 an $\epsilon=0.2$ and with probability 0.2 an $\epsilon = 0.001$. The batch RL agents are trained on this batch for 10mio steps and evaluated every 50k time steps for 10 episodes. This process of batch creation differs from the settings used in other papers in i) having only a single behavioral policy, ii) the batch size and iii) the proficiency level of the batch policy.

The experiments, performed on the arcade learning environment include DQN, REM, QR-DQN, KL-Control, BCQ, OnlineDQN and Behavioral Cloning and show that:
- for conventional RL algorithms distributional algorithms (QR-DQN) outperform the plain algorithms (DQN)
- batch RL algorithms perform better than conventional algorithms with BCQ outperforming every other algorithm in every tested game

In addition to the return the authors plot the value estimates for the Q-networks. A drop in performance corresponds in all cases to a divergence (up or down) in value estimates. 

The paper is an important contribution to the debate about what is the right setting to evaluate batch RL algorithms. It remains however to be seen if the proposed choice of i) a single behavior policy, ii) the batch size and iii) quality level of the behavior policy will be accepted as standard. Further work is in any case required to decide upon a benchmark for continuous domains.

This summary builds substantially on my summary of NERFs, so if you haven't yet read that, I recommend doing so first! 

The idea of a NERF is learn a neural network that represents a 3D scene, and from which you can, once the model is trained, sample an image of that scene from any desired angle. This involves structuring your neural network as a function that predicts the RGB color and density/opacity for a given point in 3D space (x, y, z), from a given viewing angle (theta, phi). With such a function, you can generate predictions of what images taken from certain angles would look like by sampling along a viewing ray, and integrating the combined hue and opacity into an aggregated view. This prediction can then be compared to a true image taken from that direction, and gradients passed backwards into the prediction model. 

An important assumption of this model is that the scene being photographed is static; specifically, that every point in space is always inhabited by the same part of the 3D object, regardless of what angle it's viewed from. This is a reasonable assumption for photos of inanimate objects, or of humans in highly controlled lab settings, but it is often not true for humans when you, say, ask them to take a selfie video of themselves. Even if they're trying to keep roughly still, there will be slight shifts in the location and position of their head between frames, and the authors of this paper show that this can lead to strange artifacts if you naively try to train a NERF from the images (including a particularly odd one where it hallucinates tiny copies of the image in the air surrounding the face). 

https://i.imgur.com/IUVh6uM.png

The fix proposed by this paper is to apply a learnable deformation field to each image, where the notion is to deform each view into being in one canonical position (fixed per network, since, again, one network corresponds to a single scene). This means that, along with learning the parameters of the NERF itself, you're also learning what deformation to apply to each training image to get it into this canonical position. This is done by parametrizing the deformation in a particular way, and then having that deformation be conditioned by a latent vector that's trained similar to how you'd train an embedding (one learned vector per image example). The parametrization of the deformation is honestly a little bit over my head, given my lack of grounding in 3D modeling, but my general sense is that it applies some constraints and regularization to ensure that the learned deformations are realistic, insofar as humans are mostly rigid (one patch of skin on my forehead generally doesn't move except in concordance with the rest of my forehead), but with some possibility for elasticity (skin can stretch if I, say, smile). The authors also include an annealing scheme whereby, early in training, the model focuses on learning course (large-scale) deformations, and later in training, it's allowed to also learn weights for more precise deformations. This is to hopefully match macro-scale shifts before adding the noise of precise changes. 

This addition of a learned deformation is most of the contribution of this method: with it applied, they show that they're able to learn realistic NERFs from selfies, which they term "NERFIES". They mention a few pieces of concurrent work that try to solve the same problem of non-static human subjects in different ways, but I haven't had a chance to read those, so I can't really comment on how NERFIES stacks up to alternate approaches, but it appears to be as least one empirically convincing solution to the problem it's aiming at.

This new architecture out of Deepmind applies combines information extraction and bottlenecks to a traditional Transformer base to get a model that can theoretically apply self-attention to meaningfully larger input sizes than earlier architectures allowed. 

Currently, self-attention models are quite powerful and capable, but because attention is quadratic-in-sequence-length in both time, and, often more saliently, memory, it's infeasible to use on long sequences without some modification. This paper propose what they call "cross-attention," where some smaller-dimensional latent vector attends to the input (the latent generates the queries, the input the keys and values). This lets the network pull information out of the larger-dimensional input into a smaller and fixed-by-hyperparameter, size of latent. From there, multiple self-attention layers are applied to generate a new latent, which can be fed back into the beginning of the process to query new information from the input, accounting for the "iterative" in the title of this work. 

The authors argue this approach lets them take larger inputs, and create deeper models, because the cost of each self-attention layer (going from latent-dim to latent-dim) is small and controlled. Like many other Transformer-based architectures, they use positional encodings, theirs based on Fourier features at different frequencies. 

https://i.imgur.com/Wc8rzII.png

My overall take from the results presented is that it is competitive on many of the audio and vision tasks tested, with none of the convolutional priors that even something like Vision Transformer (which does course convolution-style preprocessing before going into Transformer layers) require, though it didn't dramatically outperform the state-of-the-art on any of the tested tasks. One thing that was strange to me was that they didn't (at least in the main paper, haven't read the appendix) seem to evaluate on text, which would seem like an obvious benchmark if you're proposing a Transformer-alternate architecture.