A Review of the Neural History of NLP

This article was published about a month ago. But we found it relevant, readable and illuminating. It gives you an idea about language modeling from pre-deep-learning eras: such as smoothing by Kneser and Ney to our current days, where large scale pre-train model is a norm. Also notable is its non-neural section which mention BLEU, LDA, OneNote and many significant NLP innovations which has nothing to do with deep learning. This should be refreshing for many of us who believe deep learning is the only solution for all our problems.

ruder.io

Nathaniel Popper on Blockchain in AI

Blockchain and AI are perhaps the two trendiest technologies now. So no wonder someone try to thread them together. One project we mentioned is OpenMined led by Andrew Trask which combine model training technology with cryptography.

But the companies you may hear more is perhaps SingularityNet, by Dr. Ben Goertzel, which attempts to use blockchain to link AI services together. The author of this piece, Nathaniel Popper, has covered blockchain for years. He penned the book Digital Gold which is a very interesting historical accounts of the rise of blockchain. So we think you might be interested in his view of how blockchain affects AI.

nytimes.com

Object Detection Using dlib by Adrian Rosebrock

Here is another article we quote from our great teacher, Adrian Rosebrock. This time he teach you how to write a fast object tracker using dlib. Other than the tutorial on dlib, perhaps the more interesting part is his presentation of dlib’s correlation tracking. We really enjoy it so don’t want to spoil it for you. Go check it out!

pyimagesearch.com

Curiosity vs Procrastination

Here is an interesting post from Google on how to deal with the sparse reward problem in RL, i.e. how to use curiosity to drive exploration of an RL algorithm. There are two interesting high-lights:

1. First, when Google authors try to create rewards based on inability of prediction, or just surprise. They found that agents could fall into a state which is close to human’s procrastination. That lead to the agent to continue to explore infinitely. So for example, when agent saw a (virtual) TV in a 3-D maze, he will get stuck infinitely because the sensory experience is always a surprise.
2. So the researchers figure out you can just store your recent experience in a memory bank, and try to reward exploration by comparing if the new experience is dissimilar to the old one. Sounds interesting. And that comparison is done by deep neural network.

We don’t pretend we fully understand the technical details, but Google’s demo is fairly intuitive, and we are fascinated.

googleblog.com