Deep Learning Weekly - 🤖 - Issue #51: ImageNet Challenges, Deep Learning and Neuroscience, Reinforcement Learning for Complex Goals, RAWGraphs & more

A look at Salesforce AI strategy and outlook starting with their rollout the machine learning platform “Einstein”. Salesforce predicts that AI’s impact through CRM software alone will add over $1 trillion to GDPs around the globe and create 800,000 new jobs. Naturally, they see themselves at the forefront of this development.

If you have been wondering what Ray Kurzweil is up to these days, this article holds the answer, he and his team are behind Gmail’s Smart Reply feature, which offers users three terse replies depending on the email’s content. Not quite the singularity, yet, but certainly a productive and promising application of AI.

This year, Kaggle hosts all three ImageNet Challenges for the first time.  The other two are the Object Detection Challenge and the Object Detection from Video Challenge.

An interesting rebuttal of a recent NY Times article on why AI is supposedly stuck. 

Adrian Rosebrock from PyImageSearch.com is gearing up to launch his latest book on deep learning + computer vision in September. We’ve had a chance to review a pre-release draft, and have to agree — it’s the best resource online to master deep learning and computer vision we have come across. We’ll publish an in-depth review on our blog soon.

This extensive article and accompanying notebook walks you through both the traditional reinforcement learning paradigm in machine learning as well as a new and emerging paradigm for extending reinforcement learning to allow for complex goals that vary over time.

A fun data science project description of building a music recommender system by training a CNN on spectrograms of songs from 9 different genres.

A two part series on Understanding Siamese networks, and Implementing them in Pytorch walking you through the architecture essentials, and implementation details.

This post by DeepMind makes the case for closer collaboration between AI and neuroscience arguing that drawing inspiration from neuroscience can, on the one hand, validate existing techniques if discovered that they mimic a function of the brain and on the other hand, that neuroscience can provide a rich source of inspiration for new algorithms.

Although published in 2016 we thought this talk merits inclusion in this issue since it expands on the point of the DeepMind article above giving an introduction to theoretical neuroscience and laying out the mutually beneficial relationship with deep learning. This is a dense and information packed talked but it comes highly recommended.

A comprehensive collection of Jupyter notebooks designed to teach deep learning, Apache MXNet, and the gluon interface.

Caffe2 has added impressive RNN capabilities resulting in several product teams at Facebook, including speech recognition and ads ranking, using Caffe2 to train RNN models.

TensorFire is powered by WebGL making it work on any GPU. This demo app shows off TensorFire’s ability to run the style-transfer neural network in your browser as fast as CPU TensorFlow on a desktop. The library has not been released yet, but you can sign up for updates.

Incredibly neat visualization tool to create custom vector-based visualizations on top of the d3.js

Fascinating paper showing that small and shallow feed-forward neural networks can achieve near state-of-the-art results on a range of unstructured and structured language processing tasks while being considerably cheaper in memory and computational requirements than deep recurrent models.

The authors propose a new self-organizing hierarchical softmax formulation for neural-network-based language models over large vocabularies. Instead of using a predefined hierarchical structure, their approach is capable of learning word clusters with clear syntactical and semantic meaning during the language model training process.

Really fun paper introducing a model that learns to convert simple hand drawings into graphics programs written in a subset of LATEX. The model combines techniques from deep learning and program synthesis. First a convolutional neural network proposes plausible drawing primitives that explain an image, then this set of drawing primitives is used like an execution trace for a graphics program.

Portmanteaus are a word formation phenomenon where two words are combined to form a new word. The authors propose character-level neural sequence-to-sequence (S2S) methods for the task of portmanteau generation that are end-to-end-trainable, language independent, and do not explicitly use additional phonetic information.