Introduction
Text in documents is easy to search, but there's a lot of information in other formats. Voice recognition turns audio – and video soundtracks – into text you can index and search. But what about the video itself, or other images?
Searching for images on the web would be a lot more accurate if instead of just looking for text on the page or in the caption that suggests a picture is relevant, the search engine could actually recognise what was in the picture. Thanks to machine learning techniques using neural networks and deep learning, that's becoming more achievable.
Caption competition
When a team of Microsoft and Facebook researchers created a massive data dump of over 300,000 images with 2.5 million objects labelled by people (called Common Objects in Context), they said all those objects are things a four-year-old child could recognise. So a team of Microsoft researchers working on machine learning decided to see how well their systems could do with the same images – not just recognising them, but breaking them up into different objects, putting a name to each object and writing a caption to describe the whole image.
To measure the results, they asked one set of people to write their own captions and another set to compare the two and say which they preferred.
"That's what the true measure of quality is," explains distinguished scientist John Platt from Microsoft Research. "How good do people think these captions are? 23% of the time they thought ours were at least as good as what people wrote for the caption. That means a quarter of the time that machine has reached as good a level as the human."
Part of the problem was the visual recogniser. Sometimes it would mistake a cat for a dog, or think that long hair was a cat, or decide that there was a football in a photograph of people gesticulating at a sculpture. This is just what a small team was able to build in four months over the summer, and it's the first time they had a labelled a set of images this large to train and test against.
"We can do a better job," Platt says confidently.
Machine strengths
Machine learning already does much better on simple images that only have one thing in the frame. "The systems are getting to be as good as an untrained human," Platt claims. That's testing against a set of pictures called ImageNet, which are labelled to show how they fit into 22,000 different categories.
"That includes some very fine distinctions an untrained human wouldn't know," he explains. "Like Pembroke Welsh corgis and Cardigan Welsh corgis – one of which has a longer tail. A person can look at a series of corgis and learn to tell the difference, but a priori they wouldn't know. If there are objects you're familiar with you can recognise them very easily but if I show you 22,000 strange objects you might get them all mixed up." Humans are wrong about 5% of the time with the ImageNet tests and machine learning systems are down to about 6%.
That means machine learning systems could do better at recognising things like dog breeds or poisonous plants than ordinary people. Another recognition system called Project Adam, that MSR head Peter Lee showed off earlier this year, tries to do that from your phone.
Project Adam
Project Adam was looking at whether you can make image recognition faster by distributing the system across multiple computers rather than running it on a single fast computer (so it can run in the cloud and work with your phone). However, it was trained on images with just one thing in them.
"They ask 'what object is in this image?'" explains Platt. "We broke the image into boxes and we were evaluating different sub-pieces of the image, detecting common words. What are the objects in the scene? Those are the nouns. What are they doing? Those are verbs like flying or looking.
"Then there are the relationships like next to and on top of, and the attributes of the objects, adjectives like red or purple or beautiful. The natural next step after whole image recognition is to put together multiple objects in a scene and try to come up with a coherent explanation. It's very interesting that you can look in the image and detect verbs and adjectives."
Powerful search
Making images useful
There are plenty of ways in which having your images automatically captioned and labelled will be useful, especially if you're a keen photographer trying to stay on top of your image library or a news site looking for the right photograph.
"Indexing your photos by who's in them is a very natural way to way to think about organising photos," Platt points out. With more powerful labelling, you can search for objects in images (a picture of a cat) or actions (a picture of a cat drinking) or the relation between different objects in an image. "If I remember that I had a picture of a boy and a horse, I'd like to be able to index that – both the objects of the boy and the horse, and the relation between them – and put them in an index so I can go and search for them later."
If you're putting together a catalogue of products, having an automatically generated caption might be useful, but Platt doesn't see much demand for something that specific. There is a lot of interest from different product teams at Microsoft, he says, but instead of creating captions for you he expects that "the pieces will be used in various products; behind the scenes, these bits will be running."
Search relevance
Dealing with videos will mean making the recognition faster, and working out how to spot what's interesting (because not every frame will be). But what's important here is not just the speed, but the way the kind of understanding that underlies captioning complex images could transform search.
The deep learning neural networks and machine learning systems this image recognition uses are the same technologies that have revolutionised speech recognition and translation in the last few years (powering Microsoft's upcoming Skype Translator). "Every time you talk to the Bing search engine on your phone you're talking to a deep network," says Platt. Microsoft's video search system, MAVIS, uses a deep network.
The next step is to do more than recognise, and actually understand what things mean.
"Even for text there's a fair amount of work and that's where there's a lot of interesting value, if we can truly understand text as opposed to just doing keyword search. Just doing keyword search gets you a long way, that's how all of our search engines work today. But imagine if you had a system that could truly understand what your documents were about and truly be an assistant to you."
The goal, he says, is to "try to truly understand the semantics of objects like video or speech or image or text, as opposed to the surface forms like just the words or just the colours."
No comments:
Post a Comment