Hi guys this is my first post here so apologies if something is amiss. I have a keen interest in AI and Robotics and I am currently pursuing an MS in the field. I would like to share some interesting stuff I come across every now and then if other people show an interest in it. Here's the first one: A team of scientists has recently demonstrated 93% accurate lipreading AI. Though it must be pointed out that the input data is a bit controlled (only a single fixed facial orientation + simple sentences) so we might not be seeing it applied in real life just yet.
It's a research paper by deepmind describing their methodology but without the source. I'm sure someone would use tensorflow or theano to replicate their method and opensource the code and models on github.
Yeah deepmind has been trying to catch up to certain open source projects for a while. It's surprising to me that they used such a limited data set tbh. Sometime Google works in mysterious ways lol.
To be honest only one of the people working on this is from deepmind so I wouldn't call it a "deepmind project". Just that it's more marketable that way :D
Just wanted to update that another team at oxford just published a paper on the same topic and they seem to have much better results. They showcased their lip reading demo working on clips taken from tv news segments with normal sentence structure.
One of those technologies you hope people have the wits to not connect their home automation to it... imagine that slapping someone in the face suffices to lock them out of their own home lol...
That's a much more impressive example! I've recently dabbled in some ML-related topics and I was shocked by how bad some of the common data sets are. They were surprisingly old and just so much 'easier' (highly constrained, perfect conditions, etc) than "in the wild" data.