Sebastian Ruder recently wrote an article on The Gradient and asserted that the oracle of natural language processing is emerging. While I am not sure such confident statement is overstated, I do look forward to the moment that we will download pre-trained embedded language models and transfer to our use cases, just like we are using pre-trained word-embedding models such as Word2Vec and FastText.

I do not think one can really draw a parallelism between computer vision and natural language processing. Computer vision is challenging, but natural language processing is even more difficult because the tasks regarding linguistics are not limited to object or meaning recognition, but also human psychology, cultures, and linguistic diversities. The objectives are far from being identical.

However, the transferrable use of embedded language models is definitely a big step forward. Ruder quoted three articles, which I would summarize below in a few words.

• Embeddings from Language Models (ELMo, arXiv:1802.05365): based on the successful bidirectional LSTM language models, the authors developed a deep contextualized embedded models by collapses all layers in the neural network architecture.
• Universal Language Model Fine-Tuning for Text Classification (ULMFiT, arXiv:1801.06146): the authors proposed a type of architectures that learn representations for specific tasks, which involve three steps in training: a) LM pre-training: learning through unlabeled corpus with abundant data; b) LM fine-tuning: learning through labeled corpus; and c) classifier fine-tuning: transferred training for specific classification tasks.
• OpenAI Transformer (article still in progress): the author proposed a simple generative language model with the three similar steps in ULMFit: a) unsupervised pre-training: training a language model that maximizes the likelihood of a sequence of tokens within a context window; b) supervised fine-tuning: a supervised classification training that maximizes the likelihood using the Bayesian approach; c) task-specific input transformations: training the classifiers on a specific task.

These three articles are intricately related to each other. Without abundant data and good hardware, it is almost impossible to produce the language models. As Ruder suggested, we will probably have a pre-trained model up to the second step of the ULMFit and OpenAI Transformer papers, but we train our own specific model for our use. We have been doing this for word-embedding models, and this approach has been common in computer vision too.

Google launches her AutoML project last year, in an effort to automate the process of seeking the most appropriate neural net designs for a particular classification problem. Designing neural networks have been time consuming, despite the use of TensorFlow / Keras or other deep learning architecture nowadays. Therefore, the Google Brain team devised the Neural Architecture Search (NAS) using a recurrent neural network to perform reinforcement learning. (See their blog entry.) It is used to find the neural networks for image classifiers. (See their blog entry.)

Apparently, with a state-of-the-art hardware, it is of Google’s advantage to perform such an experiment on the CIFAR-10 dataset using 450 GPUs for 3-4 days. But this makes the work inaccessible for small companies or personal computers.

Then it comes an improvement to NAS: the Efficient Neural Architecture Search via Parameter Sharing (ENAS), which is a much more efficient method to search for a neural networks, by narrowing down the search in a subgraph. It reduces the need of GPUs.

While I do not think it is a threat to machine learning engineers, it is a great algorithm to note. It looks to me a brute-force algorithm, but it needs scientists and engineers to gain insights. Still, I believe development of the theory behind neural networks is much needed.

Recently, Geoffrey Hinton and his colleagues made the article about capsules available. He has been known to heavily criticize the use of pooling and back propagation.

“A capsule is a group of neurons whose activity vector represents the instantiation parameters of a specific type of entity such as an object or object part.” The nodes of inputs and outputs are vectors, instead of scalars as in neural networks. A cheat sheet comparing the traditional neurons and capsules is as follow:

Based on the capsule, the authors suggested a new type of layer called CapsNet.

Huadong Liao implemented CapsNet with TensorFlow according to the paper. (Refer to his repository.)

The theory and the interpretability of deep neural networks have always been called into questions. In the recent few years, there have been several ideas uncovering the theory of neural networks.

# Renormalization Group (RG)

Mehta and Schwab analytically connected renormalization group (RG) with one particular type of deep learning networks, the restricted Boltzmann machines (RBM). (See their paper and a previous post.) RBM is similar to Heisenberg model in statistical physics. This weakness of this work is that it can only explain only one type of deep learning algorithms.

However, this insight gives rise to subsequent work, with the use of density matrix renormalization group (DMRG), entanglement renormalization (in quantum information), and tensor networks, a new supervised learning algorithm was invented. (See their paper and a previous post.)

# Neural Networks as Polynomial Approximation

Lin and Tegmark were not satisfied with the RG intuition, and pointed out a special case that RG does not explain. However, they argue that neural networks are good approximation to several polynomial and asymptotic behaviors of the physical universe, making neural networks work so well in predictive analytics. (See their paper, Lin’s reply on Quora, and a previous post.)

# Information Bottleneck (IB)

Tishby and his colleagues have been promoting information bottleneck as a backing theory of deep learning. (See previous post.) In recent papers such as arXiv:1612.00410, on top of his information bottleneck, they devised an algorithm using variation inference.

# Generalization

Recently, Kawaguchi, Kaelbling, and Bengio suggested that “deep model classes have an exponential advantage to represent certain natural target functions when compared to shallow model classes.” (See their paper and a previous post.) They provided their proof using generalization theory. With this, they introduced a new family of regularization methods.

# Geometric View on Generative Adversarial Networks (GAN)

Recently, Lei, Su, Cui, Yau, and Gu tried to offer a geometric view of generative adversarial networks (GAN), and provided a simpler method of training the discriminator and generator with a large class of transportation problems. However, I am still yet to understand their work, and their experimental results were done on low-dimensional feature spaces. (See their paper.) Their work is very mathematical.

In their paper, Kawaguchi, Kaelbling, and Bengio explored the theory of why generalization in deep learning is so good. Based on their theoretical insights, they proposed a new regularization method, called Directly Approximately Regularizing Complexity (DARC), in addition to commonly used Lp-regularization and dropout methods.

This paper explains why deep learning can generalize well, despite large capacity and possible algorithmic instability, nonrobustness, and sharp minima, effectively addressing an open problem in the literature. Based on our theoretical insight, this paper also proposes a family of new regularization methods. Its simplest member was empirically shown to improve base models and achieve state-of-the-art performance on MNIST and CIFAR-10 benchmarks. Moreover, this paper presents both data-dependent and data-independent generalization guarantees with improved convergence rates. Our results suggest several new open areas of research.

If you have been taking Andrew Ng’s deeplearning.ai course on Coursera, you must have learned in Course 1 about the graph operations, and the method of back propagation using derivatives in terms of graph. In fact, it is the basis of TensorFlow, a Python package commonly used in deep learning. Because it is based on the graph model of computation, we can see it as a “programming language.”

Google published a paper about the big picture of computational model in TensorFlow:

TensorFlow is a powerful, programmable system for machine learning. This paper aims to provide the basics of a conceptual framework for understanding the behavior of TensorFlow models during training and inference: it describes an operational semantics, of the kind common in the literature on programming languages. More broadly, the paper suggests that a programming-language perspective is fruitful in designing and in explaining systems such as TensorFlow.

Beware that this model is not limited to deep learning.

There are many embeddings algorithm for representations. Sammon embedding is the oldest one, and we have Word2Vec, GloVe, FastText etc. for word-embedding algorithms. Embeddings are useful for dimensionality reduction.

Traditionally, quantum many-body states are represented by Fock states, which is useful when the excitations of quasi-particles are the concern. But to capture the quantum entanglement between many solitons or particles in a statistical systems, it is important not to lose the topological correlation between the states. It has been known that restricted Boltzmann machines (RBM) have been used to represent such states, but it has its limitation, which Xun Gao and Lu-Ming Duan have stated in their article published in Nature Communications:

There exist states, which can be generated by a constant-depth quantum circuit or expressed as PEPS (projected entangled pair states) or ground states of gapped Hamiltonians, but cannot be efficiently represented by any RBM unless the polynomial hierarchy collapses in the computational complexity theory.

PEPS is a generalization of matrix product states (MPS) to higher dimensions. (See this.)

However, Gao and Duan were able to prove that deep Boltzmann machine (DBM) can bridge the loophole of RBM, as stated in their article:

Any quantum state of n qubits generated by a quantum circuit of depth T can be represented exactly by a sparse DBM with O(nT) neurons.

(diagram adapted from Gao and Duan’s article)

On 07/28/2017, shorttext published its release 0.4.1, with a few important updates. To install it, type the following in the OS X / Linux command line:

>>> pip install -U shorttext

The documentation in PythonHosted.org has been abandoned. It has been migrated to readthedocs.org. (URL: http://shorttext.readthedocs.io/ or http:// shorttext.rtfd.io)

## Exploiting the Word-Embedding Layer

This update is mainly due to an important update in gensim, motivated by earlier shorttext‘s effort in integrating scikit-learn and keras. And gensim also provides a keras layer, on the same footing as other neural networks, activation function, or dropout layers, for Word2Vec models. Because shorttext has been making use of keras layers for categorization, such advance in gensim in fact makes it a natural step to add an embedding layer of all neural networks provided in shorttext. How to do it? (See shorttext tutorial for “Deep Neural Networks with Word Embedding.”)

import shorttext
trainclassdict = shorttext.data.subjectkeywords() &nbsp; # load an example data set


To train a model, you can do it the old way, or do it the new way with additional gensim function:

kmodel = shorttext.classifiers.frameworks.CNNWordEmbed(wvmodel=wvmodel, nb_labels=len(trainclassdict.keys()), vecsize=100, with_gensim=True) &nbsp; # keras model, setting with_gensim=True
classifier = shorttext.classifiers.VarNNEmbeddedVecClassifier(wvmodel, with_gensim=True, vecsize=100) &nbsp; # instantiate the classifier, setting with_gensim=True
classifier.train(trainclassdict, kmodel)


The parameters with_gensim in both CNNWordEmbed and VarNNEmbeddedVecClassifier are set to be False by default, because of backward compatibility. However, setting it to be True will enable it to use the new gensim Word2Vec layer.

These change in gensim and shorttext are the works mainly contributed by Chinmaya Pancholi, a very bright student at Indian Institute of Technology, Kharagpur, and a GSoC (Google Summer of Code) student in 2017. He revolutionized gensim by integrating scikit-learn and keras into gensim. He also used what he did in gensim to improve the pipelines of shorttext. He provided valuable technical suggestions. You can read his GSoC proposal, and his blog posts in RaRe Technologies, Inc. Chinmaya has been diligently mentored by Ivan Menshikh and Lev Konstantinovskiy of RaRe Technologies.

## Maxent Classifier

Another important update is the adding of maximum entropy (maxent) classifier. (See the corresponding tutorial on “Maximum Entropy (MaxEnt) Classifier.”) I will devote a separate entry on the theory, but it is very easy to use it,

import shorttext
from shorttext.classifiers import MaxEntClassifier

classifier = MaxEntClassifier()


Use the NIHReports dataset as the example:

classdict = shorttext.data.nihreports()
classifier.train(classdict, nb_epochs=1000)


The classification is just like other classifiers provided by shorttext:

classifier.score('cancer immunology') # NCI tops the score
classifier.score('children health') # NIAID tops the score
classifier.score('Alzheimer disease and aging') # NIAID tops the score


A preprint on arXiv recently caught a lot of attentions. While deep learning is successful in various types of neural networks, it had not been so for feed-forward neural networks. The authors of this paper proposed normalizing the network with a new activation function, called “selu” (scaled exponential linear units):

$\text{selu}(x) =\lambda \left\{ \begin{array}{cc} x & \text{if } x>0 \\ \alpha e^x - \alpha & \text{if } x \leq 0 \end{array} \right.$.

which is an improvement to the existing “elu” function.

Despite this achievement, what caught the eyeballs is not the activation function, but the 93-page appendix of mathematical proof:

And this is one of the pages in the appendix:

Some scholars teased at it on Twitter too: