ENet, a fully convolutional net is aimed at making real-time inferences for segmentation.  ENet is up to 18×
faster, requires 75× less FLOPs, has 79× less parameters, and provides similar or better accuracy to existing models. SegNet has been used for benchamarking purposes.
##### Innovation
ENet has been able to achieve faster speeds because  instead of following the regular encoder-decoder structure followed by most state-of-art semantic segmentation nets, it uses a ResNet kind of approach. So while most nets learn features from a highly down-sampled convolutional map, and then upsample the learnt features, ENet tries to learn the features using underlying principles of ResNet from a convolution map that is only downsampled thrice in the encoder part. The encoder part is designed to have same functionality as convolutional architectures used for classification tasks. Also, the author mentions that multiple downsampling tends to hurt accuracy apart from huritng memory constraints by way of upsampling layera. Hence , by way of including only 2 upsampling operations & a single deconvolution, ENet aims to achieve faster semantic segmentation.

##### Architecture
The architecture of the net looks like this. 
![enet_architecture](https://i.imgur.com/rw1lVKQ.png)

where the initial block and each of the bottlenecks look like this respectively
![enet_bottlenecks](https://i.imgur.com/sveifk5.png)

In oder to reduce FLOPs further, no bias terms were included in any of the projection steps as cuDNN uses separate kernels for convolution and bias addition. 

##### Training
The encoder and decoder parts are trained sparately. The encoder part was trained to categorize downsampled regions of the input image, then the decoder was appended and trained the network to perform upsampling and pixel-wise classification

##### Results
The results are reported on  on widely used NVIDIA Titan X GPU as well as on NVIDIA TX1 embedded system module. For segmentation results on CamVid, CityScape & SUN RGB-D datasets, the inference times and model sizes were significantly lower, while the IoU or accuracy (as applicable) was nearly equivalent to SegNet for most cases

If you were to survey researchers, and ask them to name the 5 most broadly influential ideas in Machine Learning from the last 5 years, I’d bet good money that Batch Normalization would be somewhere on everyone’s lists. Before Batch Norm, training meaningfully deep neural networks was an unstable process, and one that often took a long time to converge to success. When we added Batch Norm to models, it allowed us to increase our learning rates substantially (leading to quicker training) without the risk of activations either collapsing or blowing up in values. It had this effect because it addressed one of the key difficulties of deep networks: internal covariate shift. 

To understand this, imagine the smaller problem, of a one-layer model that’s trying to classify based on a set of input features. Now, imagine that, over the course of training, the input distribution of features moved around, so that, perhaps, a value that was at the 70th percentile of the data distribution initially is now at the 30th. We have an obvious intuition that this would make the model quite hard to train, because it would learn some mapping between feature values and class at the beginning of training, but that would become invalid by the end. This is, fundamentally, the problem faced by higher layers of deep networks, since, if the distribution of activations in a lower layer changed even by a small amount, that can cause a “butterfly effect” style outcome, where the activation distributions of higher layers change more dramatically. Batch Normalization - which takes each feature “channel” a network learns, and normalizes [normalize = subtract mean, divide by variance] it by the mean and variance of that feature over spatial locations and over all the observations in a given batch - helps solve this problem because it ensures that, throughout the course of training, the distribution of inputs that a given layer sees stays roughly constant, no matter what the lower layers get up to. 

On the whole, Batch Norm has been wildly successful at stabilizing training, and is now canonized - along with the likes of ReLU and Dropout - as one of the default sensible training procedures for any given network. However, it does have its difficulties and downsides. One salient one of these comes about when you train using very small batch sizes - in the range of 2-16 examples per batch. Under these circumstance, the mean and variance calculated off of that batch are noisy and high variance (for the general reason that statistics calculated off of small sample sizes are noisy and high variance), which takes away from the stability that Batch Norm is trying to provide. 

One proposed alternative to Batch Norm, that didn’t run into this problem of small sample sizes, is Layer Normalization. This operates under the assumption that the activations of all feature “channels” within a given layer hopefully have roughly similar distributions, and, so, you an normalize all of them by taking the aggregate mean over all channels, *for a given observation*, and use that as the mean and variance you normalize by. Because there are typically many channels in a given layer, this means that you have many “samples” that go into the mean and variance. However, this assumption - that the distributions for each feature channel are roughly the same - can be an incorrect one. 

A useful model I have for thinking about the distinction between these two approaches is the idea that both are calculating approximations of an underlying abstract notion: the in-the-limit mean and variance of a single feature channel, at a given point in time. Batch Normalization is an approximation of that insofar as it only has a small sample of points to work with, and so its estimate will tend to be high variance. Layer Normalization is an approximation insofar as it makes the assumption that feature distributions are aligned across channels: if this turns out not to be the case, individual channels will have normalizations that are biased, due to being pulled towards the mean and variance calculated over an aggregate of channels that are different than them. 

Group Norm tries to find a balance point between these two approaches, one that uses multiple channels, and normalizes within a given instance (to avoid the problems of small batch size), but, instead of calculating the mean and variance over all channels, calculates them over a group of channels that represents a subset. The inspiration for this idea comes from the fact that, in old school computer vision, it was typical to have parts of your feature vector that - for example - represented a histogram of some value (say: localized contrast) over the image. Since these multiple values all corresponded to a larger shared “group” feature. If a group of features all represent a similar idea, then their distributions will be more likely to be aligned, and therefore you have less of the bias issue. 

One confusing element of this paper for me was that the motivation part of the paper strongly implied that the reason group norm is sensible is that you are able to combine statistically dependent channels into a group together. However, as far as I an tell, there’s no actually clustering or similarity analysis of channels that is done to place certain channels into certain groups; it’s just done so semi-randomly based on the index location within the feature channel vector. So, under this implementation, it seems like the benefits of group norm are less because of any explicit seeking out of dependant channels, and more that just having fewer channels in each group means that each individual channel makes up more of the weight in its group, which does something to reduce the bias effect anyway.

The upshot of the Group Norm paper, results-wise, is that Group Norm performs better than both Batch Norm and Layer Norm at very low batch sizes. This is useful if you’re training on very dense data (e.g. high res video), where it might be difficult to store more than a few observations in memory at a time. However, once you get to batch sizes of ~24, Batch Norm starts to do better, presumably since that’s a large enough sample size to reduce variance, and you get to the point where the variance of BN is preferable to the bias of GN.  

This paper continues in the tradition of curiosity-based models, which try to reward models for exploring novel parts of their environment, in the hopes this can intrinsically motivate learning. However, this paper argues that it’s insufficient to just treat novelty as an occasional bonus on top of a normal reward function, and that instead you should figure out a process that’s more specifically designed to increase novelty. Specifically: you should design a policy whose goal is to experience transitions and world-states that are high novelty. 

In this setup, like in other curiosity-based papers, “high novelty” is defined in terms of a state being unpredictable given a prior state, history, and action. However, where other papers saw novelty reward as something only applied when the agent arrived at somewhere novel, here, the authors build a model (technically, an ensemble of models) to predict the state at various future points. The ensemble is important here because it’s (quasi) bootstrapped, and thus gives us a measure of uncertainty. States where the predictions of the ensemble diverge represent places of uncertainty, and thus of high value to explore. I don’t 100% follow the analytic specification of this idea (even though the heuristic/algorithmic description makes sense). The authors frame the Utility function of a state and action as being equivalent to the Jenson Shannon Divergence (~distance between probability distributions) shown below. 

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


Here, P(S | S, a, T) is the probability of a state given prior state and action under a given model of the environment (Transition Model), and P(gamma) is the distribution over the space of possible transition models one might learn. A “model” here is one network out of the ensemble of networks that makes up our bootstrapped (trained on different sets) distribution over models. Conceptually, I think this calculation is measuring “how different is each sampled model/state distribution from all the other models in the distribution over possible models”. If the models within the distribution diverge from one another, that indicates a location of higher uncertainty. 

What’s important about this is that, by building a full transition model, the authors can calculate the expected novelty or “utility” of future transitions it might take, because it can make a best guess based on this transition model (which, while called a “prior”, is really something trained on all data up to this current iteration). My understanding is that these kinds of models function similarly to a Q(s,a) or V(s) in a pure-reward case: they estimate the “utility reward” of different states and actions, and then the policy is updated to increase that expected reward. 

I’ve recently read papers on ICM, and I was a little disappointed that this paper didn’t appear to benchmark against that, but against Bootstrapped DQN and Exploration Bonus DQN, which I know less well and can less speak to the conceptual differences from this approach. Another difficulty in actually getting a good sense of results was that the task being tested on is fairly specific, and different from RL results coming out of the world of e.g. Atari and Deep Mind Labs. All of that said, this is a cautiously interesting idea, if the results generate to beat more baselines on more environments. 

This paper explores the use of so-called Monte Carlo objectives for training directed generative models with latent variables. Monte Carlo objectives take the form of the logarithm of a Monte Carlo estimate (i.e. an average over samples) of the marginal probability $P(x)$. One important motivation for using Monte Carlo objectives is that they can be shown (see the Importance Weighted Variational Autoencoder paper \cite{journals/corr/BurdaGS15} and my notes on it) to correspond to bounds on the true likelihood of the model, and one can tighten the bound simply by drawing more samples in the Monte Carlo objective. Currently, the most successful application of Monte Carlo objectives is based on an importance sampling estimate, which involves training a proposal distribution $Q(h|x)$ in addition to the model $P(x,h)$.

This paper considers the problem of training with gradient descent on such objectives, in the context of a model to which the reparametrization trick cannot be used (e.g. for discrete latent variables). They analyze the sources of variance in the estimation of the gradients (see Equation 5) and propose a very simple approach to reducing the variance of a sampling-based estimator of these gradients. 

First, they argue that gradients with respect to the $P(x,h)$ parameters are less susceptible to problems due to high variance gradients.

Second, and most importantly, they derive a multi-sample estimate of the gradient that is meant to reduce the variance of gradients on the proposal distribution parameters $Q(h|x)$. The end result is the gradient estimate of Equations 10-11. It is based on the observation that the first term of the gradient of Equation 5 doesn't distinguish between the contribution of each sampled latent hi. The key contribution is this: they notice that one can incorporate a variance reducing baseline for each sample hi, corresponding to the Monte Carlo estimate of the log-likelihood when removing hi from the estimate (see Equation 10). The authors show that this is a proper baseline, in that using it doesn't introduce a bias in the estimation for the gradients.

Experiments show that this approach yields better performance than training based on Reweighted Wake Sleep \cite{journals/corr/BornscheinB14} or the use of NVIL baselines \cite{conf/icml/MnihG14}, when training sigmoid belief networks as generative models or as structured output prediction (image completion) models on binarized MNIST.

The paper discusses and empirically investigates by empirical testing the effect of "catastrophic forgetting" (**CF**), i.e. the inability of a model to perform a task it was previously trained to perform if retrained to perform a second task. 

An illuminating example is what happens in ML systems with convex objectives: regardless of the initialization (i.e. of what was learnt by doing the first task), the training of the second task will always end in the global minimum, thus totally "forgetting" the first one. 

Neuroscientific evidence (and common sense) suggest that the outcome of the experiment is deeply influenced by the similarity of the tasks involved. Namely, if (i) the two tasks are *functionally identical but input is presented in a different format* or if (ii)  *tasks are similar* and the third case for (iii) *dissimilar tasks*. 

Relevant examples may be provided respectively by (i) performing the same image classification task starting from two different image representations as RGB or HSL, (ii) performing image classification tasks with semantically similar as classifying two similar animals and (iii) performing a text classification followed by image classification. 

The problem is investigated by an empirical study covering two methods of training ("SGD" and "dropout") combined with 4 activations functions (logistic sigmoid, RELU, LWTA, Maxout). A random search is carried out on these parameters. 

From a practitioner's point of view, it is interesting to note that dropout has been set to 0.5 in hidden units and 0.2 in the visible one since this is a reasonably well-known parameter. 

## Why the paper is important
It is apparently the first to provide a systematic empirical analysis of CF. Establishes a framework and baselines to face the problem.

## Key conclusions, takeaways and modelling remarks
* dropout helps in preventing CF
* dropout seems to increase the optimal model size with respect to the model without dropout 
* choice of activation function has a less consistent effect than dropout\no dropout choice
* dissimilar task experiment provides a  notable exception of then dissimilar task experiment
* the previous hypothesis that LWTA activation is particularly resistant to CF is rejected  (even if it performs best in the new task in the dissimilar task pair the behaviour is inconsistent)
* choice of activation function should always be cross-validated
* If computational resources are insufficient for cross-validation the combination dropout + maxout activation function is recommended.