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.

Spatial Pyramid Pooling (SPP) is a technique which allows Convolutional Neural Networks (CNNs) to use input images of any size, not only $224\text{px} \times 224\text{px}$ as most architectures do. (However, there is a lower bound for the size of the input image).

## Idea

* Convolutional layers operate on any size, but fully connected layers need fixed-size inputs
* Solution:
  * Add a new SPP layer on top of the last convolutional layer, before the fully connected layer
  * Use an approach similar to bag of words (BoW), but maintain the spatial information. The BoW approach is used for text classification, where the order of the words is discarded and only the number of occurences is kept.
  * The SPP layer operates on each feature map independently.
  * The output of the SPP layer is of dimension $k \cdot M$, where $k$ is the number of feature maps the SPP layer got as input and $M$ is the number of bins.

Example: We could use spatial pyramid pooling with 21 bins:

* 1 bin which is the max of the complete feature map
* 4 bins which divide the image into 4 regions of equal size (depending on the input size) and rectangular shape. Each bin gets the max of its region.
* 16 bins which divide the image into 4 regions of equal size (depending on the input size) and rectangular shape. Each bin gets the max of its region.

## Evaluation

* Pascal VOC 2007, Caltech101: state-of-the-art, without finetuning
* ImageNet 2012: Boosts accuracy for various CNN architectures
* ImageNet Large Scale Visual Recognition Challenge (ILSVRC) 2014: Rank #2


## Code

The paper claims that the code is [here](http://research.microsoft.com/en-us/um/people/kahe/), but this seems not to be the case any more.

People have tried to implement it with Tensorflow ([1](http://stackoverflow.com/q/40913794/562769), [2](https://github.com/fchollet/keras/issues/2080), [3](https://github.com/tensorflow/tensorflow/issues/6011)), but by now no public working implementation is available.


## Related papers

* [Atrous Convolution](https://arxiv.org/abs/1606.00915)

## Terms

* Semantic Segmentation: Traditional segmentation divides the image in visually similar patches. Semantic segmentation on the other hand divides the image in semantically meaningful patches. This usually means to classify each pixel (e.g.: This pixel belongs to a cat, that pixel belongs to a dog, the other pixel is background).


## Main ideas

* Complete neural networks which were trained for image classification can be used as a convolution. Those networks can be trained on Image Net (e.g. VGG, AlexNet, GoogLeNet)
* Use upsampling to (1) reduce training and prediction time (2) improve consistency of output. (See [What are deconvolutional layers?](http://datascience.stackexchange.com/a/12110/8820) for an explanation.)


## How FCNs work

1. Train a neural network for image classification which is trained on input images of a fixed size ($d \times w \times h$)
2. Interpret the network as a single convolutional filter for each output neuron (so $k$ output neurons means you have $k$ filters) over the complete image area on which the original network was trained.
3. Run the network as a CNN over an image of any size (but at least $d \times w \times h$) with a stride $s \in \mathbb{N}_{\geq 1}$
4. If $s > 1$, then you need an upsampling layer (deconvolutional layer) to convert the coarse output into a dense output.

## Nice properties

* FCNs take images of arbitrary size and produce an image of the same output size.
* Computationally efficient

## See also:

https://www.quora.com/What-are-the-benefits-of-converting-a-fully-connected-layer-in-a-deep-neural-network-to-an-equivalent-convolutional-layer

> They allow you to treat the convolutional neural network as one giant filter. You can then spatially apply the neural net as a convolution to images larger than the original training image size, getting a spatially dense output.
>
> Let's say you train a neural net (with some loss function) with a convolutional layer (3 x 3, stride of 2), pooling layer (3 x 3, stride of 2), and a fully connected layer with 10 units, using 25 x 25 images. Note that the receptive field size of each max pooling unit is 7 x 7, so the pooling output is 5 x 5. You can convert the fully connected layer to to a set of  10 5 x 5 convolutional filters (unit strides). If you do that, the entire net can be treated as a filter with receptive field size 35 x 35 and stride of 4. You can then take that net and apply it to a 50 x 50 image, and you'd get a 3 x 3 x 10 spatially dense output.

So the hypervector is just a big vector created from a network:

`"We concatenate features from some or all of the feature
maps in the network into one long vector for every location
which we call the hypercolumn at that location. As an
example, using pool2 (256 channels), conv4 (384 channels)
and fc7 (4096 channels) from the architecture of [28] would
lead to a 4736 dimensional vector."`

So how exactly do we construct the vector? 

![](https://i.imgur.com/hDvHRwT.png)

Each activation map results in a single element of the resulting hypervector. The corresponding pixel location in each activation map is used as if the activation maps were all scaled to the size of the original image.

The paper shows the below formula for the calculation. Here $\mathbf{f}_i$ is the value of the pixel in the scaled space and each $\mathbf{F}_{k}$ are points in the activation map. $\alpha_{ik}$ scales the known values to produce the midway points.

$$\mathbf{f}_i  = \sum_k \alpha_{ik} \mathbf{F}_{k}$$

Then the fully connected layers are simply appended to complete the vector. 

So this gives us a representation for each pixel but is it a good one? The later layers will have the input pixel in their receptive field. After the first few layers it is expected that the spatial constraint is not strong.

This paper proposes to train a neural network generative model by optimizing an importance sampling (IS) weighted estimate of the log probability under the model. The authors show that the case of an estimate based on a single sample actually corresponds to the learning objective of variational autoencoders (VAE). Importantly, they exploit this connection by showing that, similarly to VAE, a gradient can be passed through the approximate posterior (the IS proposal) samples, thus yielding an importance weighted autoencoder (IWAE). The authors also show that, by using more samples, this objective, which is a lower bound of the actual log-likelihood, becomes an increasingly tighter approximation to the log-likelihood. In other words, the IWAE is expected to better optimize the real log-likelihood of the neural network, compared to VAE.

The experiments presented show that the model achieves competitive performance on a version of the binarized MNIST benchmark and on the Omniglot dataset.

#### My two cents

This is a really neat contribution! While simple (both conceptually and algorithmically), it really seems to be an important step forward for the VAE framework. I really like the theoretical result showing that IWAE provides a better approximation to the real log-likelihood, it's quite neat and provides an excellent motivation for the method.

The results on binarized MNIST are certainly impressive. Unfortunately, it appears that the training setup isn't actually comparable to the majority of published results on this dataset. Indeed, it seems that they didn't use the stochastic but *fixed* binarization of the inputs that other publications on this benchmark have used (since my paper on NADE with Iain Murray, we've made available that fixed training set for everyone to use, along with fixed validation and test sets as well). I believe instead they've re-sampled the binarization for each minibatch, effectively creating a setup with a somewhat larger training set than usual. It's unfortunate that this is the case, since it makes this result effectively impossible to compare directly with previous work.

I'm being picky on this issue only because I'm super interested in this problem (that is of generative modeling with neural networks) and this little issue is pretty much the only thing that stops this paper from being a slam dunk. Hopefully the authors (or perhaps someone interested in reimplementing IWAE) can clarify this question eventually.

Otherwise, it seems quite clear to me that IWAE is an improvement over VAE. The experiments of section 5.2, showing that fine-tuning a VAE model with IWAE training improves performance, while fine-tuning a IWAE model using VAE actually makes things worse, is further demonstration that IWAE is indeed a good idea.