In Spite of Gains Google’s Neural Networks Won’t Scale

As I have mentioned before, computing power is more than just paying an electric bill. When a computer or a cell phone fails it requires a technician. If your phone dies you have to take it to a technician. If your Mac fails to the Mac store. Modern digital devices are extremely complex systems. Some GPS systems working on your phone require software taking into consideration Einstein’s Theory of Relativity. If you have not understood it yet,  complex systems require lots of time and lots of money. Although, the progress is laudable over at Google Translate ultimately good translators will still have jobs. Machine translation will be limited to daily banter and discussion.

Last week Google announced on its research blog that Google Translate is going over to ‘neural machine translation’. It’s amazing what the engineers have accomplished: accuracy has increased by several orders of magnitude. Computational cost is also lower. However, most impressive is the discovery of ‘interlingua’ a semantical language that exists between unknown languages. It was all accomplished using neural networks. Those are software techniques fundamental to artificial intelligence. In other words, the machines no longer need to have huge corpora of translated texts like in statistical or phrase-based translation. Instead, the software discovers this interlingua between languages. The interlingua makes it possible to translate between language pairs the computer has never seen.

Leave a Reply

Your email address will not be published.