One of the cool by-products of our training process is that the words arrange themselves in the high-dimensional space in such a way that similar words are near each other in the high dimensional space, and also in such a way that linear operations on these word vectors often produce interesting results. For example, capital cities all tend to lie in roughly the same direction from their corresponding country (see image). This often makes it possible to solve simple analogies by performing arithmetic on the vectors and looking for the nearest words in the embedding space. For example:
E("biggest") - E("big") + E("small") ~= E("smallest")
E("Italy") - E("Rome") + E("Paris") ~= E("France")
E("scientist") - E("Einstein") + E("Picasso") ~= E("painter")
The image in the blog post has a nice mapping of some of these high dimensional vectors down to a two-dimensional representation using principal components analysis where you can really see this effect for a few countries and capital cities.
Today, we're open sourcing some code to train models like this on your own textual data (see the blog post for details), and also releasing some pre-trained embeddings:
Thinking about what you want to do in the New Year? Instead of making resolutions, why not try a few 30 day challenges? They're over faster, you learn neat things from them, and they can be a ton of fun. Think of them like experiments to figure out how to be happier or healthier or learn new skills.
- Indian Institute of Technology GuwahatiComputer Science, 2012 - presentPattern Recognition and Machine Learning, Algorithmic Game Theory
A Comparison of Web Servers for Python Based Web Applications | DigitalO...
Introduction In this article, we will talk about three main things: Python, Web Servers, and most importantly the bits and bobs in-between
Google App Engine Blog: 10 things you (probably) didn't know about App E...
The official Google App Engine blog. The latest news on Google App Engine and the App Engine community.
Om Bhadram Karnebhih Shrnuyama Devah - in sanskrit with meaning - mantra...
ॐ भद्रं कर्णेभिः शृणुयाम देवाः । भद्रं पश्येमाक्षभिर्यजत्राः । स्थिरैरङ्गैस्तुष्टुवागँसस्तनूभिः । व्यशेम देवहितं यदायूः । Om Bhadram Karnneb
Google App Engine: Hello World using Push-to-Deploy
Ever wished you could deploy to Google App Engine with the same standard tools you use to version your code? Now you can. With the Push-to-D
Blind signal separation - Wikipedia, the free encyclopedia
Blind signal separation, also known as blind source separation, is the separation of a set of source signals from a set of mixed signals, wi