Pocketables

Guest post: Google’s Machine Translation is moving us closer to general AI

Today’s guest post comes from Rachel Wheeler of Morningside Translations.

Edited for formatting


Google’s Machine Translation is Moving Us Closer to General AI

Artificial Intelligence (AI) is no longer the stuff of fiction. In past years, it has permeated many aspects of our lives and is now on course to become even more widely adopted in society. We’ve already seen the advent of AI chatbots and machine learning recommendation tools that have improved consumer engagement and experiences. The translation industry is one of the verticals where AI is moving at a fast pace, with natural language recognition technologies being used for a vast number of applications.

Looking back at how we got to where we are helps us to better understand where we’re headed, so let’s begin with an overview of this technology.

A Brief History of AI

Narrow Artificial Intelligence – Developed in the 1950’s, narrow AI was designed to complete rule-based tasks to the same standard as or better than humans. A contemporary example of this sort of AI is Facebook’s facial recognition software.

Machine Learning – In the 1990’s, algorithms and the beginnings of ‘big data’ were used to teach computers to identify specific subsets of data and separate them. This allows the computer to make informed decisions on expected actions.

Rather than confining the machine to a specific set of rules, the algorithms program the system to ‘learn’. Voice skill software used in assistants like Siri and Alexa are perfect examples of this technology in use.

Deep Learning – An extension of Machine Learning, deep learning is based on complex structures that are similar to a human’s brain or in other cases an animal’s visual cortex. In both instances, these artificial neural networks learn from past input to solve problems they haven’t previously been presented with. This is where Google’s Neural Machine Translation (GNMT) fits.

General AI ­– This is the predicted near future for AI, and refers to complex neural networks powering machines that can independently perform general actions. This will be done with similar characteristics and at a similar level of intelligence as humans.  

Google’s Extraordinary Breakthrough

Google Translate hit the headlines when it switched from Narrow AI technology for translation to a deep learning system based on artificial neural networks. It was officially dubbed Google Neural Machine Translation (GNMT) shortly after development.

As well as enabling much more natural sounding translations and continually learning from constant input, the system did something that none of its creators expected: On asking it to translate between two languages it had never directly attempted before, the software produced a direct translation without additional input. This makes GNMT the first instance of transfer learning in Machine Translation.

Next Steps for AI

Other systems for translation are also making great strides in the area of natural language recognition and translation. Facebook’s CNN technique is still using neural networks, but in different ways. The results have consistently been found to be nine times faster than Google’s GNMT approach.

These different approaches are all feeding into the creation of robust translation platforms that will be further developed with wide collaboration. The range of possibilities will depend on the decisions of both Google and Facebook to open up their findings to the public and make their research available through their open-source code websites and blogs.

But it is not just translation that these deep learning artificial neural networks can be used for. From medical diagnoses to fighting crime, AI has the potential to touch every area of our lives.

Hardware advances have further enabled neural networks and deep learning techniques, and computers are constantly learning from many types of input, such as temperature and touch. Google and the startup Nervana collaborated to create processing chips in 2017 that could support larger neural networks and machine learning applications. These small pieces of software are expected to be ten times faster than the CNN GPU technique for processing multiple parts of a data set simultaneously. Additionally, Intel announced in late 2017 that they were designing ‘neuromorphic chips’ called Loihi that are modelled on the human brain in order to support general AI programming.

Google’s Machine Translation breakthrough appears to have released the AI genie from the lamp and set us on a path that will make not only make General AI possible, but also imminently usable.