This paper presents a variational approach to the maximisation of mutual information in the context of a reinforcement learning agent. Mutual information in this context can provide a learning signal to the agent that is "intrinsically motivated", because it relies solely on the agent's state/beliefs and does not require from the ("outside") user an explicit definition of rewards.

Specifically, the learning objective, for a current state s, is the mutual information between the sequence of K actions a proposed by an exploration distribution $w(a|s)$ and the final state s' of the agent after performing these actions. To understand what the properties of this objective, it is useful to consider the form of this mutual information as a difference of conditional entropies:

$$I(a,s'|s) = H(a|s) - H(a|s',s)$$

Where $I(.|.)$ is the (conditional) mutual information and $H(.|.)$ is the (conditional) entropy. This objective thus asks that the agent find an exploration distribution that explores as much as possible (i.e. has high $H(a|s)$ entropy) but is such that these actions have predictable consequences (i.e. lead to predictable state s' so that $H(a|s',s)$ is low). So one could think of the agent as trying to learn to have control of as much of the environment as possible, thus this objective has also been coined as "empowerment".

The main contribution of this work is to show how to train, on a large scale (i.e. larger state space and action space) with this objective, using neural networks. They build on a variational lower bound on the mutual information and then derive from it a stochastic variational training algorithm for it. The procedure has 3 components: the exploration distribution $w(a|s)$, the environment $p(s'|s,a)$ (can be thought as an encoder, but which isn't modeled and is only interacted with/sampled from) and the planning model $p(a|s',s)$ (which is modeled and can be thought of as a decoder). The main technical contribution is in how to update the exploration distribution (see section 4.2.2 for the technical details).

This approach exploits neural networks of various forms. Neural autoregressive generative models are also used as models for the exploration distribution as well as the decoder or planning distribution. Interestingly, the framework allows to also learn the state representation s as a function of some "raw" representation x of states. For raw states corresponding to images (e.g. the pixels of the screen image in a game), CNNs are used.

This paper presents a linear algebraic trick for computing both the value and the gradient update for a loss function that compares a very high-dimensional target with a (dense) output prediction. Most of the paper exposes the specific case of the squared error loss, though it can also be applied to some other losses such as the so-called spherical softmax. One use case could be for training autoencoders with the squared error on very high-dimensional but sparse inputs. While a naive (i.e. what most people currently do) implementation would scale in $O(Dd)$ where $D$ is the input dimensionality and d the hidden layer dimensionality, they show that their trick allows to scale in $O(d^2)$. 

Their experiments show that they can achieve speedup factors of over 500 on the CPU, and over 1500 on the GPU. 

#### My two cents

This is a really neat, and frankly really surprising, mathematical contribution. I did not suspect getting rid of the dependence on D in the complexity would actually be achievable, even for the "simpler" case of the squared error. 

The jury is still out as to whether we can leverage the full power of this trick in practice. Indeed, the squared error over sparse targets isn't the most natural choice in most situations. The authors did try to use this trick in the context of a version of the neural network language model that uses the squared error instead of the negative log-softmax (or at least I think that's what was done... I couldn't confirm this with 100% confidence). They showed that good measures of word similarity (Simlex-999) could be achieved in this way, though using the hierarchical softmax actually achieves better performance in about the same time. 

But as far as I'm concerned, that doesn't make the trick less impressive. It's still a neat piece of new knowledge to have about reconstruction errors. Also, the authors mention that it would be possible to adapt the trick to the so-called (negative log) spherical softmax, which is like the softmax but where the numerator is the square of the pre-activation, instead of the exponential. I hope someone tries this out in the future, as perhaps it could be key to making this trick a real game changer!

Zhang et al. propose CROWN, a method for certifying adversarial robustness based on bounding activations functions using linear functions. Informally, the main result can be stated as follows: if the activation functions used in a deep neural network can be bounded above and below by linear functions (the activation function may also be segmented first), the network output can also be bounded by linear functions. These linear functions can be computed explicitly, as stated in the paper. Then, given an input example $x$ and a set of allowed perturbations, usually constrained to a $L_p$ norm, these bounds can be used to obtain a lower bound on the robustness of networks.

Also find this summary at [davidstutz.de](https://davidstutz.de/category/reading/).

Deeper networks should never have a higher **training** error than smaller ones. In the worst case, the layers should "simply" learn identities. It seems as this is not so easy with conventional networks, as they get much worse with more layers. So the idea is to add identity functions which skip some layers. The network only has to learn the **residuals**. 

Advantages:

* Learning the identity becomes learning 0 which is simpler
* Loss in information flow in the forward pass is not a problem anymore
    * No vanishing / exploding gradient
* Identities don't have parameters to be learned

## Evaluation

The learning rate starts at 0.1 and is divided by 10 when the error plateaus. Weight decay of 0.0001 ($10^{-4}$), momentum of 0.9. They use mini-batches of size 128.

* ImageNet ILSVRC 2015: 3.57% (ensemble)
* CIFAR-10: 6.43%
* MS COCO: 59.0% mAp@0.5 (ensemble)
* PASCAL VOC 2007: 85.6% mAp@0.5
* PASCAL VOC 2012: 83.8% mAp@0.5

## See also

* [DenseNets](http://www.shortscience.org/paper?bibtexKey=journals/corr/1608.06993)

## 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.