The authors analyse in the very well written paper the relation between Fisher $F(\theta) = \sum_n \mathbb{E}_{p_{\theta}(y \vert x)}[\nabla_{\theta} \log(p_{\theta}(y \vert x_n))\nabla_{\theta} \log(p_{\theta}(y \vert x_n))^T] $ and empirical Fisher $\bar{F}(\theta) = \sum_n [\nabla_{\theta} \log(p_{\theta}(y_n \vert x_n))\nabla_{\theta} \log(p_{\theta}(y_n \vert x_n))^T] $, which has recently seen a surge in interest. . The definitions differ in that $y_n$ is a training label instead of a sample of the model $p_{\theta}(y \vert x_n)$, thus even so the name suggests otherwise $\bar{F}$ is not a empirical, for example Monte Carlo, estimate of the Fisher. The authors rebuff common arguments used to justify the use of the empirical fisher by an amendment to the generalized Gauss-Newton, give conditions when the empirical Fisher does indeed approach the Fisher and give an argument why the empirical fisher might work in practice nonetheless.   

The Fisher, capturing the curvature of the parameter space, provides information about the geometry of the parameters pace, the empirical Fisher might however fail so capture the curvature as the striking plot from the paper shows:
https://i.imgur.com/c5iCqXW.png

The authors rebuff the two major justifications for the use of empirical Fisher:
1. "the empirical Fisher matches the construction of a generalized Gauss-Newton"
 * for the log-likelihood $log(p(y \vert f) = \log \exp(-\frac{1}{2}(y-f)^2))$ the generalized Gauss-Newton intuition that small residuals $f(x_n, \theta) - y_n$ lead to a good approximation of the Hessian is not satisfied. Whereas the Fisher approaches the Hessian, the empirical Fisher approaches 0
2. "the empirical Fisher converges to the true Fisher when the model is a good fit for the data"
 * the authors sharpen the argument to "the empirical Fisher converges at the minimum to the Fisher as the number of samples grows", which is unlikely to be satisfied in practice. 

The authors provide an alternative perspective on why the empirical Fisher might be successful, namely to adapt the gradient to the gradient noise in stochastic optimization. The empirical Fisher coincides with the second moment of the stochastic gradient estimate and encodes as such covariance information about the gradient noise. This allows to reduce the effects of gradient noise by scaling back the updates in high variance aka noise directions. 

When humans classify images, we tend to use high-level information about the shape and position of the object. However, when convolutional neural networks classify images,, they tend to use low-level, or textural, information more than high-level shape information. This paper tries to understand what factors lead to higher shape bias or texture bias. 

To investigate this, the authors look at three datasets with disagreeing shape and texture labels. The first is GST, or Geirhos Style Transfer. In this dataset, style transfer is used to render the content of one class in the style of another (for example, a cat shape in the texture of an elephant). In the Navon dataset, a large-scale letter is rendered by tiling smaller letters. And, in the ImageNet-C dataset, a given class is rendered with a particular kind of distortion; here the distortion is considered to be the "texture label". In the rest of the paper, "shape bias" refers to the extent to which a model trained on normal images will predict the shape label rather than the texture label associated with a GST image. The other datasets are used in experiments where a model explicitly tries to learn either shape or texture. 

https://i.imgur.com/aw1MThL.png

To start off, the authors try to understand whether CNNs are inherently more capable of learning texture information rather than shape information. To do this, they train models on either the shape or the textural label on each of the three aforementioned datasets. On GST and Navon, shape labels can be learned faster and more efficiently than texture ones. On ImageNet-C (i.e. distorted ImageNet), it seems to be easier to learn texture than texture, but recall here that texture corresponds to the type of noise, and I imagine that the cardinality of noise types is far smaller than that of ImageNet images, so I'm not sure how informative this comparison is. Overall, this experiment suggests that CNNs are able to learn from shape alone without low-level texture as a clue, in cases where the two sources of information disagree 

The paper moves on to try to understand what factors about a normal ImageNet model give it higher or lower shape bias - that is, a higher or lower likelihood of classifying a GST image according to its shape rather than texture. Predictably, data augmentations have an effect here. When data is augmented with aggressive random cropping, this increases texture bias relative to shape bias, presumably because when large chunks of an object are cropped away, its overall shape becomes a less useful feature. Center cropping is better for shape bias, probably because objects are likely to be at the center of the image, so center cropping has less of a chance of distorting them. On the other hand, more "naturalistic" augmentations like adding Gaussian noise or distorting colors lead to a higher shape bias in the resulting networks, up to 60% with all the modifications. However, the authors also find that pushing the shape bias up has the result of dropping final test accuracy. 

https://i.imgur.com/Lb6RMJy.png

Interestingly, while the techniques that increase shape bias seem to also harm performance, the authors also find that higher-performing models tend to have higher shape bias (though with texture bias still outweighing shape) suggesting that stronger models learn how to use shape more effectively, but also that handicapping models' ability to use texture in order to incentivize them to use shape tends to hurt performance overall. 

Overall, my take from this paper is that texture-level data is actually statistically informative and useful for classification - even in terms of generalization - even if is too high-resolution to be useful as a visual feature for humans. CNNs don't seem inherently incapable of learning from shape, but removing their ability to rely on texture seems to lead to a notable drop in accuracy, suggesting there was real signal there that we're losing out on.

TLDR; The authors propose a web navigation task where an agent must find a target page containing a search query (typically a few sentences) by navigating a web graph with restrictions on memory, path length and number of exlorable nodes. Tey train Feedforward and Recurrent Neural Networks and evaluate their performance against that of human volunteers.


#### Key Points


- Datasets: Wiki-[NUM_ALLOWED_HOPS]: WikiNav-4 (6k train), WikiNav-8 (890k train), WikiNav-16 (12M train). Authors evaluate variosu query lengths for all data sets.
- Vector representation of pages: BoW of pre-trained word2vec embeddings. 
- State-dependent action space: All possible outgoing links on the current page. At each step, the agent can peek at the neighboring nodes and see their full content.
- Training, a single correct path is fed to the agent. Beam search to make predictions.
- NeuAgent-FF uses a single tanh layer. NeuAgent-Rec uses LSTM.
- Human performance typically worse than that of Neural agents


#### Notes/Questions

- Is it reasonable to allow the agents to "peek" at neighboring pages? Humans can make decisions based on the hyperlink context. In practice, peaking at each page may not be feasible if there are many links on the page.
- I'm not sure if I buy the fact that this task requires Natural Language Understanding. Agents are just matching query word vectors against pages, which is no indication of NLU. An indication of NLU would be if the query was posed in a question format, which is typically short. But here, the authors use several sentences as queries and longer queries lead to better results, suggesting that the agents don't actually have any understanding of language. They just match text.
- Authors say that NeuAgent-Rec performed consistently better for high hop length, but I don't see that in the data.
- The training method seems a bit strange to me because the agent is fed only one correct path, but in reality there are a large number of correct paths and target pages. It may be more sensible to train the agent with all possible target pages and paths to answer a query.

This paper synthesizes a high-quality video of Barack Obama given the audio. Practically, it only synthesizes the region around the mouth, while the rest of the elements (i.e. pixels) come from a video in a database. 
The overall pipeline is the following:
- Given a video, an audio and a mouth shape are extracted. Audio is represented as MFCC coefficients; mouth shape - 18 lip markers;
- Train audio to mouth shape mapping with time-delayed unidirectional LSTM.
- Synthesize mouth texture: retrieve a number of video frames in a database where a mouth shape is similar to the output of LSTM; synthesize median texture by applying weighted median on mouth shapes from retrieved video frames; manually select teeth target frame (selection criteria are purely subjected) and enhance teeth median texture with selected teeth target frame.
- Re-timing to avoid situations where Obama is not speaking but his head is moving which looks very unnatural. 
- Final composition into the target video involves jaw correction to make it more natural.
![Algorithm flow](http://www.kurzweilai.net/images/Obama-lip-Sync-Graphic.jpg)

The results look ridiculously natural. Authors suggest that one of the applications of this paper is speech summarization, where you summarize a speech not only with selected parts as text and audio but also synthesize a video for it. Personally, this work inspires me to work on a method that is able to generate natural sign language interpreter that takes sound/text as input and produces sign language moves. 

The method is a multi-task learning model performing person detection, keypoint detection, person segmentation, and pose estimation. It is a bottom-up approach as it first localizes identity-free semantics and then group them into instances.
https://i.imgur.com/kRs9687.png

Model structure:
 - **Backbone**. A feature extractor is presented by ResNet-(50 or 101) with one [Feature Pyramid Network](https://arxiv.org/pdf/1612.03144.pdf) (FPN) for keypoint branch and one for person detection branch. FPN enhances extracted features through multi-level representation.
 -  **Keypoint detection** detects keypoints as well as produces a pixel-level segmentation mask.
https://i.imgur.com/XFAi3ga.png
FPN features $K_i$ are processed with multiple $3\times3$ convolutions followed by concatenation and final $1\times1$  convolution to obtain predictions for each keypoint, as well as segmentation mask (see Figure for details). This results in #keypoints_in_dataset_per_person + 1 output layers. Additionally, intermediate supervision (i.e. loss) is applied at the FPN outputs. $L_2$ loss between predictions and Gaussian peaks at the keypoint locations is used. Similarly, $L_2$ loss is applied for segmentation predictions and corresponding ground truth masks.
 - **Person detection** is essentially a [RetinaNet](https://arxiv.org/pdf/1708.02002.pdf), a one-stage object detector, modified to only handle *person* class.
 - **Pose estimation**. Given initial keypoint predictions, Pose Estimation Network (PRN) selects a single keypoint for each class. 
https://i.imgur.com/k8wNP5p.png
During inference, PRN takes cropped outputs from keypoint detection branch defined by the predicted bounding boxes from the person detection branch, resizes it to a fixed size, and forwards it through a multilayer perceptron with residual connection. During the training, the same process is performed, except the cropped keypoints come from the ground truth annotation defined by a labeled bounding box.

This model is not an end-to-end trainable model. While keypoint and person detection branches can, in theory, be trained simultaneously, PRN network requires separate training.

**Personal note**. Interestingly, PRN training with ground truth inputs (i.e. "perfect" inputs) only reaches 89.4 mAP validation score which is surprisingly quite far from the max possible score. This presumably means that even if preceding networks or branches perform god-like, the PRN might become a bottleneck in the performance. Therefore, more efforts should be directed to PRN itself. Moreover, modifying the network to support end-to-end training might help in boosting the performance.

Open-source implementations used to make sure the paper apprehension is correct: [link1](https://github.com/LiMeng95/MultiPoseNet.pytorch), [link2](https://github.com/IcewineChen/pytorch-MultiPoseNet).