Deep rectified neural networks are over-parameterized in the sense that scaling of the weights in one layer, can be compensated for exactly in the subsequent layer. This paper introduces Path-SGD, a simple modification to the SGD update rule, whose update is invariant to such rescaling. The method is derived from the proximal form of gradient descent, whereby a constraint term is added which preserves the norm of the "product weight" formed along each path in the network (from input to output node). Path-SGD is thus principled and shown to yield faster convergence for a standard 2 layer rectifier network, across a variety of dataset (MNIST, CIFAR-10, CIFAR-100, SVHN). As the method implicitly regularizes the neural weights, this also translates to better generalization performance on half of the datasets.

At its core, Path-SGD belongs to the family of learning algorithms which aim to be invariant to model reparametrizations. This is the central tenet of Amari's natural gradient (NG) \cite{amari_natural_1998}, whose importance has resurfaced in the area of deep learning. Path-SGD can thus be cast an approximation to NG, which focuses on a particular type of rescaling between neighboring layers. The paper would greatly benefit from such a discussion in my opinion. I also believe NG to be a much more direct way to motivate Path-SGD, than the heuristics of max-norm regularization.

The paper introduces a sequential variational auto-encoder that generates complex images iteratively. The authors also introduce a new spatial attention mechanism that allows the model to focus on small subsets of the image. This new approach for image generation produces images that can’t be distinguished from the training data.

#### What is DRAW:
The deep recurrent attention writer (DRAW) model has two differences with respect to other variational auto-encoders. First, the encoder and the decoder are recurrent networks. Second, it includes an attention mechanism that restricts the input region observed by the encoder and the output region observed by the decoder.

#### What do we gain?
The resulting images are greatly improved by allowing a conditional and sequential generation. In addition, the spatial attention mechanism can be used in other contexts to solve the “Where to look?” problem.

#### What follows?
A possible extension to this model would be to use a convolutional architecture in the encoder or the decoder. Although this might be less useful since we are already restricting the input of the network.

#### Like:
* As observed in the samples generated by the model, the attention mechanism works effectively by reconstructing images in a local way.
* The attention model is fully differentiable.

#### Dislike:
* I think a better exposition of the attention mechanism would improve this paper.

Offline reinforcement learning is potentially high-value thing for the machine learning community learn to do well, because there are many applications where it'd be useful to generate a learnt policy for responding to a dynamic environment, but where it'd be too unsafe or expensive to learn in an on-policy or online way, where we continually evaluate our actions in the environment to test their value. In such settings, we'd like to be able to take a batch of existing data - collected from a human demonstrator, or from some other algorithm - and be able to learn a policy from those pre-collected transitions, without being able to query the environment further by taking arbitrary actions. 

There are two broad strategies for learning a policy from precollected transitions. One is to simply learn to mimic the action policy used by the demonstrator, predicting the action the demonstrator would take in a given state, without making use of reward data at all. This is Behavioral Cloning, and has the advantage of being somewhat more conservative (in terms of not experimenting with possibly-unsafe-or-low-reward actions the demonstrator never took), but this is also a disadvantage, because it's not possible to get higher reward than the demonstrator themselves got if you're simply copying their behavior. Another approach is to learn a Q function - estimating the value of a given action in a given state - using the reward data from the precollected transitions. This can also have some downsides, mostly in the direction of overconfidence. Q value Temporal Difference learning works by using the current reward added to the max Q value over possible next actions as the target for the current-state Q estimate. This tends to lead to overestimates, because regression to the mean effects mean that the highest value Q estimates are disproportionately likely to be noisy (possibly because they correspond to an action with little data in the demonstrator dataset). In on-policy Q learning, this is less problematic, because the agent can take the action associated with their noisily inaccurate estimate, and as a result get more data for that action, and get an estimate that is less noisy in future. But when we're in a fully offline setting, all our learning is completed before we actually start taking actions with our policy, so taking high-uncertainty actions isn't a valuable source of new information, but just risky. 

The approach suggested by this DeepMind paper - Critic Regularized Regression, or CRR - is essentially a synthesis of these two possible approaches. The method learns a Q function as normal, using temporal difference methods. The distinction in this method comes from how to get a policy, given a learned Q function. Rather than simply taking the action your Q estimate says is highest-value at a particular point, CRR optimizes a policy according to the formula shown below. The f() function is a stand-in for various potential functions, all of which are monotonic with respect to the Q function, meaning they increase when the Q function does. 

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

This basically amounts to a form of a behavioral cloning loss (with the part that maximizes the probability under your policy of the actions sampled from the demonstrator dataset), but weighted or, as the paper terms it, filtered, by the learned Q function. The higher the estimated q value for a transition, the more weight is placed on that transition from the demo dataset having high probability under your policy. Rather than trying to mimic all of the actions of the demonstrator, the policy preferentially tries to mimic the demonstrator actions that it estimates were particularly high-quality. Different f() functions lead to different kinds of filtration. The `binary`version is an indicator function for the Advantage of an action (the Q value for that action at that state minus some reference value for the state, describing how much better the action is than other alternatives at that state) being greater than zero. Another, `exp`, uses exponential weightings which do a more "soft" upweighting or downweighting of transitions based on advantage, rather than the sharp binary of whether an actions advantage is above 1. 

The authors demonstrate that, on multiple environments from three different environment suites, CRR outperforms other off-policy baselines - either more pure behavioral cloning, or more pure RL - and in many cases does so quite dramatically. They find that the sharper binary weighting scheme does better on simpler tasks, since the trade-off of fewer but higher-quality samples to learn from works there. However, on more complex tasks, the policy benefits from the exp weighting, which still uses and learns from more samples (albeit at lower weights), which introduces some potential mimicking of lower-quality transitions, but at the trade of a larger effective dataset size to learn from.

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.

Arguably, the central achievement of the deep learning era is multi-layer neural networks' ability to learn useful intermediate feature representations using a supervised learning signal. In a supervised task, it's easy to define what makes a feature representation useful: the fact that's easier for a subsequent layer to use to make the final class prediction. When we want to learn features in an unsupervised way, things get a bit trickier. There's the obvious problem of what kinds of problem structures and architectures work to extract representations at all. But there's also a deeper problem: when we ask for a good feature representation, outside of the context of any given task, what are we asking for? Are there some inherent aspects of a representation that can be analyzed without ground truth labels to tell you whether the representations you've learned are good are not? 

The notion of "disentangled" features is one answer to that question: it suggests that a representation is good when the underlying "factors of variation" (things that are independently variable in the underlying generative process of the data) are captured in independent dimensions of the feature representation. That is, if your representation is a ten-dimensional vector, and it just so happens that there are ten independent factors along which datapoints differ (color, shape, rotation, etc), you'd ideally want each dimension to correspond to each factor. 

This criteria has an elegance to it, and it's previously been shown useful in predicting when the representations learned by a model will be useful in predicting the values of the factors of variation. This paper goes one step further, and tests the value representations for solving a visual reasoning task that involves the factors of variation, but doesn't just involve predicting them. In particular, the authors use learned representations to solve a task patterned on a human IQ test, where some factors stay fixed across a row in a grid, and some vary, and the model needs to generate the image that "fits the pattern". 

https://i.imgur.com/O1aZzcN.png

To test the value of disentanglement, they looked at a few canonical metrics of disentanglement, including scores that represent "how many factors are captured in each dimension" and "how many dimensions is a factor spread across". They measured the correlation of these metrics with task performance, and compared that with the correlation between simple autoencoder reconstruction error and performance.

They found that at early stages of training on top of the representations, the disentanglement metrics were more predictive of performance than reconstruction accuracy. This distinction went away as the model learning on top of the representations had more time to train. It makes reasonable sense that you'd mostly see value for disentangled features in a low-data regime, since after long enough the fine-tuning network can learn its own features regardless. But, this paper does appear to contribute to evidence that disentangled features are predictive of task performance, at least when that task directly involves manipulation of specific, known, underlying factors of variation.