Useful resources: Deep Learning papers & blog posts to get started
Papers, videos, blogs, papers and readinglists to go from scratch to Recurrent Neural Networks
The content in this post is still extremely relevant and you should go through it. Still, I’m aware it’s a couple years old, I’ll update it some day (20180517).
How to use
Here is a collection of useful resources to learn / understand / discover more about (Deep) Learning. It will be updated regularly, I’ll be glad if you have suggestions to this list.
Resources can be either research papers, explanatory website or even a wellwritten Wikipedia page. Topics are wide, I hope you’ll find what you need.
Starting with Machine Learning

http://www.r2d3.us/ A beginner’s introduction to machine learning (with decision trees), how it works and what is at stake. The website is amazingly beautiful and didactical. If you’re ehre, on this page, you probably won’t need to read through that. But still, it’s too good to miss. And worth sharing with everyone. [Translations available in French, Russian and Chinese]

https://blog.monkeylearn.com/agentleguidetomachinelearning/ Here is another blog post to understand the first principles of machine learning. It nicely starts with realworld problems Machine Learning helps address, and ends with how to test your model, introduciont overfitting, precision and recall.

https://www.coursera.org/learn/machinelearning Here is Andrew Ng’s most famous MOOC on Coursera. Again, this is a “startfromscratch” ressource, it needs little maths and is easy going with coding. The point here is to teach about machine learning in general, amongst which stand neural networks. It uses Octave (free version of Matlab) which is a downside for me vs Python but this is not important if you’re just starting and if you’re not you can always do it in Python anyway. Also, it starts quite slowly especially regarding maths. Hold on to the course untill it gets a little harder and more interesting.

https://www.quora.com/topic/MachineLearning You could spend years exploring Quora so jump into it! You’ll find interesting, funny and/or weird questions and fascinating, detailed and/or concise answers. It is a great resource to both widen and deepen your interests. It’s up to you.

Reading lists Explore these and look for the topics you’re interested in / you have questions about. Needless to say, these reading lists also lead to other reading lists ;)

https://github.com/Vict0rSch/data_science_polytechnique Learn more on a few broad machine learning topics with these assignment descriptions and slides
Starting with Deep Learning

http://neuralnetworksanddeeplearning.com/ This is a goldmine to start with Deep Learning and understand neural networks. It starts from scratch, and takes you through the backpropagation algorithm, regularization, tips to train your networks etc. up to explaining convolutional networks and introducing recurrent ones. It needs very little maths and uses a simple Python code to wallk you through the implementation of feedforward and convolutional neural nets.
This is where I started with Deep Learning and I want to thank M. Nielsen for his very informative and pedagogical work.

http://cs231n.stanford.edu/syllabus.html A great set of lessons, from a Python/Numpy introduction to Convolutional Neural Networks. It also rapidly mentions other machine learning techniques such as SVMs and kNN. You’ll find there the slides, videos and assignments (notebooks). See all the code at http://cs231n.github.io/

https://www.udacity.com/course/deeplearningud730 This is Google’s Deep Learning MOOC with Udacity. They use Tensorflow rather than Theano (of course, TensorFlow it theirs…) but they do tackle issues that do not depend on your programming framework. Moreover, Keras can use TensorFlow as a backend.
Reading papers
If you like reading papers and read a lot of papers (or at least once a month) I suggest you download and use Mendeley:
Free reference manager and PDF organizer
The point is that you will be able to store and organize the papers you read to find them later on without having to dig in the web. Moreover it is superuseful to generate Latex bibliographies.
I do not know if it is the best tool you can find, it’s just that I use it and like it.
General Deep Learning papers and books

Deep Learning review in Nature (2015) by Y. LeCun, Y. Bengio and G. Hinton
Deep learning allows computational models that are composed of multiple processing layers to learn representations of data with multiple levels of abstraction. These methods have dramatically improved the stateoftheart in speech recognition, visual object recognition, object detection and many other domains such as drug discovery and genomics. Deep learning discovers intricate structure in large data sets by using the backpropagation algorithm to indicate how a machine should change its internal parameters that are used to compute the representation in each layer from the representation in the previous layer. Deep convolutional nets have brought about breakthroughs in processing images, video, speech and audio, whereas recurrent nets have shone light on sequential data such as text and speech.

Deep Learning Book (Draft) by I. Goodfellow, Y. Bengio and A. Courville. The book is not finished and the html version is quite ugly but it still is exhaustive and precise.

Pattern Recognition and Machine Learning by C. Bishop (2006). This 700 pages book is one of the Bibles of machine learning, tackling major subjects such as Graphical Models, Kernel Methods, Linear Models and … Neural Networks. It is quite mathsoriented but very precise and useful.

Practical Recommendations for GradientBased Training of Deep Architectures by Y. Bengio (2012).
Learning algorithms related to artificial neural networks and in particular for Deep Learning may seem to involve many bells and whistles, called hyperparameters. This chapter is meant as a practical guide with recommendations for some of the most commonly used hyperparameters, in particular in the context of learning algorithms based on backpropagated gradient and gradientbased optimization. It also discusses how to deal with the fact that more interesting results can be obtained when allowing one to adjust many hyperparameters. Overall, it describes elements of the practice used to successfully and efficiently train and debug largescale and often deep multilayer neural networks. It closes with open questions about the training difficulties observed with deeper architectures.

Random Search for HyperParameter Optimization in Journal of Machine Learning Research by J. Bergstra and Y. Bengio (2012)
Grid search and manual search are the most widely used strategies for hyperparameter optimization. This paper shows empirically and theoretically that randomly chosen trials are more efficient for hyperparameter optimization than trials on a grid. Empirical evidence comes from a comparison with a large previous study that used grid search and manual search to configure neural networks and deep belief networks. Compared with neural networks configured by a pure grid search, we find that random search over the same domain is able to find models that are as good or better within a small fraction of the computation time. Granting random search the same computational budget, random search finds better models by effectively searching a larger, less promising con figuration space. Compared with deep belief networks configured by a thoughtful combination of manual search and grid search, purely random search over the same 32dimensional configuration space found statistically equal performance on four of seven data sets, and superior performance on one of seven. […]

Delving Deep into Rectifiers: Surpassing HumanLevel Performance on ImageNet Classification by K. He, X. Zhang, S. Ren and J. Sun (2015)
Rectified activation units (rectifiers) are essential for stateoftheart neural networks. In this work, we study rectifier neural networks for image classification from two aspects. First, we propose a Parametric Rectified Linear Unit (PReLU) that generalizes the traditional rectified unit. PReLU improves model fitting with nearly zero extra computational cost and little overfitting risk. Second, we derive a robust initialization method that particularly considers the rectifier nonlinearities. This method enables us to train extremely deep rectified models directly from scratch and to investigate deeper or wider network architectures. Based on our PReLU networks (PReLUnets), we achieve 4.94% top5 test error on the ImageNet 2012 classification dataset. This is a 26% relative improvement over the ILSVRC 2014 winner (GoogLeNet, 6.66% [29]). To our knowledge, our result is the first to surpass humanlevel performance (5.1%, [22]) on this visual recognition challenge.

Dropout : A Simple Way to Prevent Neural Networks from Overfitting in Journal of Machine Learning Research by N. Srivastava, G. Hinton, A. Krizhevsky, I Sutskever and R. Salakhutdinov (2014)
Deep neural nets with a large number of parameters are very powerful machine learning systems. However, overfitting is a serious problem in such networks. Large networks are also slow to use, making it difficult to deal with overfitting by combining the predictions of many different large neural nets at test time. Dropout is a technique for addressing this problem. The key idea is to randomly drop units (along with their connections) from the neural network during training. This prevents units from coadapting too much. During training, dropout samples from an exponential number of different “thinned” networks. At test time, it is easy to approximate the effect of averaging the predictions of all these thinned networks by simply using a single unthinned network that has smaller weights. This significantly reduces overfitting and gives major improvements over other regularization methods. We show that dropout improves the performance of neural networks on supervised learning tasks in vision, speech recognition, document classification and computational biology, obtaining stateoftheart results on many benchmark data sets.

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift by S. Ioffe and C. Szegedy (2015)
Training Deep Neural Networks is complicated by the fact that the distribution of each layer’s inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training minibatch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a stateoftheart image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batchnormalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top5 validation error (and 4.8% test error), exceeding the accuracy of human raters.
On Recurrent Neural Networks

Check out A. Karpathy’s famous blog post on The Unreasonable Effectiveness of Recurrent Neural Networks

Nando de Freitas’s 50min video from his Oxford class is a really good introduction to RNNs and LSTMs, from the vanishing gradient problem to the LSTM Torch code : Deep Learning Lecture 12: Recurrent Neural Nets and LSTMs

Also to get a clearer understanding of how LSTMs work, see this nice blog post by Chris Olah : Understanding LSTM Networks

Another great post by Chris Olah, on how RNNs understand and represent data in the task of Natural Language Processing. Also explains the use of Word Embeddings: Deep Learning, NLP, and Representations

A Critical Review of Recurrent Neural Networks for Sequence Learning by Z. Lipton, J. Berkowitz and C. Elkan (2015)
[…]Recurrent neural networks (RNNs) are connectionist models that capture the dynamics of sequences via cycles in the network of nodes. Unlike standard feedforward neural networks, recurrent networks retain a state that can represent information from an arbitrarily long context window. Although recurrent neural networks have traditionally been difficult to train, and often contain millions of parameters, recent advances in network architectures, optimization techniques, and parallel computation have enabled successful largescale learning with them. In recent years, systems based on long shortterm memory (LSTM) and bidirectional (BRNN) architectures have demonstrated groundbreaking performance on tasks as varied as image captioning, language translation, and handwriting recognition. In this survey, we review and synthesize the research that over the past three decades first yielded and then made practical these powerful learning models. When appropriate, we reconcile conflicting notation and nomenclature. Our goal is to provide a selfcontained explication of the state of the art together with a historical perspective and references to primary research.

An Empirical Exploration of Recurrent Network Architectures in Proceedings of the 32nd International Conference on Machine Learning by R. Jozefowicz, W. Zaremba and I. Sutskever (2015)
The Recurrent Neural Network (RNN) is an extremely powerful sequence model that is often difficult to train. The Long ShortTerm Memory (LSTM) is a specific RNN architecture whose design makes it much easier to train. While wildly successful in practice, the LSTM’s architecture appears to be adhoc so it is not clear if it is optimal, and the significance of its individual components is unclear. In this work, we aim to determine whether the LSTM architecture is optimal or whether much better architectures exist. We conducted a thorough architecture search where we evaluated over ten thousand different RNN architectures, and identified an architecture that outperforms both the LSTM and the recentlyintroduced Gated Recurrent Unit (GRU) on some but not all tasks. We found that adding a bias of 1 to the LSTM’s forget gate closes the gap between the LSTM and the GRU

Recurrent Neural Network Regularization (Under review as a conference paper at ICLR 2015) by W. Zaremba, I. Sutskever and O. Vinyals (2015)
We present a simple regularization technique for Recurrent Neural Networks (RNNs) with Long ShortTerm Memory (LSTM) units. Dropout, the most successful technique for regularizing neural networks, does not work well with RNNs and LSTMs. In this paper, we show how to correctly apply dropout to LSTMs, and show that it substantially reduces overfitting on a variety of tasks. These tasks include language modeling, speech recognition, image caption generation, and machine translation.

Sequence to sequence learning with neural networks (NIPS) by I. Sutskever, O. Vinyals and Q. Le (2014). Check out I. Sutskever’s presentation video on this paper and his website for more.
Deep Neural Networks (DNNs) are powerful models that have achieved excel lent performance on difficult learning tasks. Although DNNs work well whenever large labeled training sets are available, they cannot be used to map sequences to sequences. In this paper, we present a general endtoend approach to sequence learning that makes minimal assumptions on the sequence structure. Our method uses amultilayered Long ShortTermMemory (LSTM) tomap the input sequence to a vector of a fixed dimensionality, and then another deep LSTM to decode the target sequence from the vector. Our main result is that on an English to French translation task from the WMT14 dataset, the translations produced by the LSTM achieve a BLEU score of 34.8 on the entire test set, where the LSTM’s BLEU score was penalized on outofvocabulary words. Additionally, the LSTM did not have difficulty on long sentences. For comparison, a phrasebased SMT system achieves a BLEU score of 33.3 on the same dataset. When we used the LSTM to rerank the 1000 hypotheses produced by the aforementioned SMT system, its BLEU score increases to 36.5, which is close to the previous state of the art. The LSTM also learned sensible phrase and sentence representations that are sensitive to word order and are relatively invariant to the active and the passive voice. Fi nally, we found that reversing the order of the words in all source sentences (but not target sentences) improved the LSTM’s performancemarkedly, because doing so introduced many short term dependencies between the source and the target sentence which made the optimization problem easier.
Other

http://cs.stanford.edu/people/karpathy/convnetjs/demo/mnist.html run convolutional neural networks in your browser using JavaScript (by @Karpathy)

http://timdettmers.com/ Is Tim Demetters’s blog with really interesting posts, mainly about using the GPU with neural networks:
 Which GPU(s) to Get for Deep Learning: My Experience and Advice for Using GPUs in Deep Learning
 How to Parallelize Deep Learning on GPUs Part 1/2: Data Parallelism & Part 2/2: Model Parallelism
 The Brain vs Deep Learning Part I: Computational Complexity — Or Why the Singularity Is Nowhere Near
 See also his introduction to Deep Learning on Nvidia’s website Part 1 & Part 2
Reading lists
The following are insanely good, exhaustive and pertinent reading/resources lists. I suggest you browse them because my point here is not to compete with them.

https://github.com/ujjwalkarn/MachineLearningTutorials A general list on a lot of machine learning fields.

https://github.com/ChristosChristofidis/awesomedeeplearning Deep Learningfocused list of resources, going from researches to datasets and frameworks.

http://deeplearning.net/readinglist/ Research and Deep Learingoriented reading list.

http://www.wildml.com Nice review of deep learning, great glossary worth reading [Glossary] (http://www.wildml.com/deeplearningglossary/)