Translation Battles - Google versus Edinburgh
Google have enhanced their translation technology with some sexy features. If you happen to notice that a translation is rubbish you can offer a better one. Google's system will learn.
Yup. This is machine learning. It's a clever way to learn as long as you confident you know what your machines are learning. I'm reminded of early Neural Network experiments the US military tried - machines were shown aerial photographs and taught how to spot tanks hiding among trees. It seemed to work very well. The military then tested their Neural Network on photographs they hadn't shown the computer before - and the computer did very well. As a second test the military took some more photographs - and the machines flunked out. The did very badly! As the story goes (and I've not seen the evidence; just been told the story) the computers taught themselves to recognise whether it was cloudy or not. As it happened all the "tank" photographs had been taken on a cloudy day. The test pictures of forests and woods without tanks had been taken on a clear day.
I'm sure Google won't make that mistake with their language learning.
Back in 2005 Google did very well on the National Institute of Standards and Technology (NIST) Machine Translation evaluation. So did Edinburgh University.
One of the great advantages Edinburgh offers businesses is the wide range of talented students (three Unis, more colleges and some excellent schools). An added advantage is that you can talk to Edinburgh professors who work on some of the same technologies. Some of these professors have spent time at Google too - though they're very good about honouring their NDAs with the secretive search engine. Nevertheless; I'm always acutely interested in their personal projects and where they think technology is going.
If you're ever in Edinburgh and in the right pub - I recommend buying a round of Innis and Gunn and popping over to the table chatting about backwards propagation and machine prediction.
Yup. This is machine learning. It's a clever way to learn as long as you confident you know what your machines are learning. I'm reminded of early Neural Network experiments the US military tried - machines were shown aerial photographs and taught how to spot tanks hiding among trees. It seemed to work very well. The military then tested their Neural Network on photographs they hadn't shown the computer before - and the computer did very well. As a second test the military took some more photographs - and the machines flunked out. The did very badly! As the story goes (and I've not seen the evidence; just been told the story) the computers taught themselves to recognise whether it was cloudy or not. As it happened all the "tank" photographs had been taken on a cloudy day. The test pictures of forests and woods without tanks had been taken on a clear day.
I'm sure Google won't make that mistake with their language learning.
Back in 2005 Google did very well on the National Institute of Standards and Technology (NIST) Machine Translation evaluation. So did Edinburgh University.
One of the great advantages Edinburgh offers businesses is the wide range of talented students (three Unis, more colleges and some excellent schools). An added advantage is that you can talk to Edinburgh professors who work on some of the same technologies. Some of these professors have spent time at Google too - though they're very good about honouring their NDAs with the secretive search engine. Nevertheless; I'm always acutely interested in their personal projects and where they think technology is going.
If you're ever in Edinburgh and in the right pub - I recommend buying a round of Innis and Gunn and popping over to the table chatting about backwards propagation and machine prediction.
Comments
Glen / VipeChill