natural language processing |

Google’s AI Language Software Now Open Sourced

May 13, 2016 By Nicholas Anderton Leave a Comment

"Google's AI Language Software SyntaxNet now open sourced"

Google’s AI Language Software SyntaxNet is Now in Open Source

Google has just put its AI language software called SyntaxNet in the open source section, a smart move that will not only benefit Google tremendously, but it will also help all the AI researchers working on this technology to push the market forward.

Only recently, Dag Kittlaus, co-founder of Siri, the infamous AI assistant, has presented a new digital assistant called Viv. Viv is an AI that promises cutting-edge technology able to understand more complex and natural aspects of language than the other AIs on the market now.

So how about it? Will the open sourced SyntaxNet news be a game changer for the not-yet-launched Kittlaus’s Viv?

SyntaxNet is known amongst natural language researchers as a “syntactic parser,” a processor that analyzes computer linguistics by using “deep neural networks” and tries to create a meaningful dialect.

The system combines punctuation, grammar logic and everything involved in speech – what is the subject, the object, a noun, a preposition, the verb and how they all relate – to try to figure out what a phrase might be about.

Fernando Pereira, Google’s Research Director, says SyntaxNet, the AI language software, has improved Google’s error rate up to 40% compared to previous software methods.

Researchers outside Google say SyntaxNet is probably the most advanced existing AI language software. Noah Smith, a University of Washington profession specialized in natural language understanding is surprised that Google is sharing so much of their tools, something they didn’t do as often in the past.

Many people are doing natural language research and with SyntaxNet’s availability, Smith says their work will be pushed “forward a little bit.”

Another tool open-sourced by Google was the TensorFlow, the software engine behind the whole Mountain View Company.

With these two incredibly huge implements in the open source environment, Google could have added brains working for the advancement of this technology and challenge even further the real AI conversation.

What Pereira says Google wants, is for people “to move beyond parsing, towards the deeper semantic reasoning that is necessary” and explore past parsing. They just gave parsing for free. But with people using language in so many different ways, this will seem like more than just an uphill walk.

Let’s see where Google’s available AI language software will lead the digital assistant technology.

Image source: Pixabay

Deep Learning Course Offered Free of Charge by Google

January 23, 2016 By Nicholas Anderton Leave a Comment

A deep learning course is being offered free of charge by Google, in an effort to make the field of artificial intelligence more accessible than ever.

The online classes, which are available on the Udacity learning platform, have been developed as a subdivision of the Machine Learning Engineer Nanodegree program, aimed at enabling students to devise intelligent systems and algorithms, that can process vast amounts of data in order to formulate conclusions and accurate predictions.

The person who has been in charge with conceiving the study material was Vincent Vanhoucke, tech lead at Google Brain, and principal scientist at the company.

He has been working in conjunction with Arpan Chakraborty, course developer at Udacity, with a PhD in computer science, awarded by North Carolina State University.

Together, they have created a highly challenging and comprehensive set of instructionals, which includes video lectures (with factual, up-to-date information and examples), as well as exercises, practical applications and projects.

The purpose of this self-paced course is to assist students in gaining more insight into the promising and ever-expanding field represented by machine learning and artificial intelligence.

The focus will be on deep learning, through which complex, vast amounts of data can be interpreted and processed, by arranging them layer by layer into hierarchies, depending on their level of abstraction.

Through the clips presented by Vincent Vanhoucke, it will be possible to find out how to use artificial neural networks, including convolutional ones, that are commonly employed for image recognition, natural language processing or video analysis.

There will also be detailed information regarding long short-term memory neural networks, which have proven reliable and efficient in speech recognition, and have also yielded impressive results when it comes to intelligent word recognition.

The class assignments and various tasks and projects will also introduce students to TensorFlow, an open-source machine learning library developed by Google.

The level of difficulty for this newly launched deep learning course is considered to lie between medium and advanced, given the fact that the material is mostly addressed to data scientists or engineers who have already completed prior training in machine learning, and who have become acquainted with supervised learning strategies and tasks.

Furthermore, in order for students to actually fare well and fully comprehend all the material from the video presentations, another one of the prerequisites for this course consists in having had at least 2 years’ worth of experience in computer programming (so as to be familiar especially with the Python programming language).

In addition, it is necessary for would-be pupils to have used source code management systems such as Git and Github before, given the fact that assignments will be stored in such repositories.

Moreover, those who sign up for this course should be knowledgeable regarding linear algebra (especially matrices and vectors), calculus (partial derivatives of multivariable functions, differentiation and integration), as well as statistics (standard deviations, variance, means, modes and medians).

Those who possess all the above-mentioned skills and qualifications should be aware that they will have to devote plenty of time to this course, in order to complete it.

Generally, it’s possible to go through all the study material and finish the entire set of exercises and projects in approximately 3 months, provided that one allots at least 6 hours per week to this new endeavor.

Image Source: Flickr

Microsoft’s Mimicker Alarm Helps You Become an Early Riser

January 22, 2016 By Mary Duncan Leave a Comment

Microsoft’s Mimicker Alarm helps you become an early riser, by forcing you to make funny faces or say tongue twisters, unless you want it to ring again and again and again.

The newly developed app, available solely for Android devices for now (versions 4.1 Jelly Bean or higher), was launched on Wednesday, January 20, by the Microsoft Corporation.

The free software assists users who have trouble waking up in the morning to complete this apparently simple, yet grueling task.

Moreover, it also ensures that people actually remain wide awake and start off their daily routine, instead of getting up just briefly, in order to dismiss the alarm and then doze off once again.

This is achieved by having the app user play a “mimic” game. Basically, the program asks the would-be early riser to complete a given task in less than 30 seconds.

If the individual manages to successfully do that, the alarm stops ringing; otherwise, it resumes as many times as necessary, making it impossible for the user to oversleep and be late for school or work.

There are 3 sets of games that players can choose from, and the app can be customized in order to select one of them or a combination of them so as to serve as a “mimic challenge”.

For example, one such quest requires the user to take a selfie, illustrating a specific emotion. The software verifies if the that feeling was exhibited vividly and convincingly enough thanks to its Emotion API (application program interface).

This is based on Microsoft’s Project Oxford, through which tools powered by artificial intelligence are developed, focusing on 4 major fields: facial recognition, speech to text conversion, photo editing and analysis and natural language processing.

Based on machine learning, the Mimicker Alarm is capable of examining the user’s face, and determine if it’s indeed enacting the required sentiment.

As explained by Allison Light, program manager at Project Oxford, if the portrayal is too weak and unpersuasive, then the app will ask the individual to try harder, in order to prove that he or she is indeed bright-eyed and bushy-tailed.

Another game requires the user to say a tongue twister, and success in completing this task is assesses based on the software’s Speech API, which examines language patterns and ensures that the incredibly hard to articulate phrase was spoken not just correctly, but also fast enough.

The last trial consists of identifying and photographing an object in the house bearing a particular color. So as to verify if the goal was achieved, the app relies on its Computer Vision API, which distinguishes hues in a photograph.

Even more, upon successfully completing the daily Mimic challenge, users can share their results on social media, in order to let friends and family know about their efforts of turning from night owls into early birds.

While the games may seem childish and simple at first, in the early hours of morning they may prove more demanding or stimulating than expected.

Also, they show just how advanced machine learning has become nowadays, allowing electronic devices to analyze visual content, speech, body language and micro-expressions.

Those wishing to test Microsoft’s Mimicker Alarm will be relieved to know that despite the fact that the app requires them to make recordings or upload personal photos, the data they transmit to the Project Oxford platform will only be used in order to make the machine learning software more reliable, without serving in any way to identify them or compromise their privacy.

The lifestyle app that promises to make early rising more entertaining and effortless than ever is completely free of charge, and can be downloaded at the Google Play Store. Also, developers wishing to create their own program inspired by the Mimicker Alarm can access the open-source technology on GitHub.

Image Source: Thurrott