Go back 20 years and most of us would have struggled to comprehend that computers would soon be able to recognise objects in images. Some of us are still struggling with this idea.
Deep learning involves ‘training’ a computational system to ‘understand’ natural language, so inferring complex meaning rather than just understanding the surface meaning. The computer is then quizzed on the information and goes back to learn from its mistakes.
These multiple layers help systems analyse and make decisions about data more independently, “Such as whether or not an email is spam, to use a simple example”, explains Rob Speer, chief science officer at Luminoso, a Massachusetts-based text analysis and artificial intelligence company. “For many companies, a major reason to turn to deep learning over machine learning is that there are fewer steps of human intervention required to train the system before it can work with data.” This significantly cuts staff effort and reduces the burden on the poor workies.
Luminoso was founded at MIT Media Labs in 2010 after a decade of research. Its flagship products, Analytics and Compass, offer accurate, unbiased, real-time understanding of what consumers are saying. It does this using its ConceptNet system, which works on the premise that meanings are “vectors”, representable in a multi-dimensional space, with those vectors near to each other representing similar meanings. These insights are then used to increase business performance and build better customer experiences.
The trouble is that the vast majority of most companies' data is unstructured nowadays and, says Speer, “Deep learning has particular advantages over machine learning when working with this unstructured text data.”
When it’s using natural language, a system can learn to organise text by what it means, not just by the set of words it contains. This means it can take in companies' customer feedback, for example, and instead of just counting up the number of customers who are happy or unhappy, it can reveal that they’re (for example) livid about the lack of toilet paper or the waiter’s bad attitude.
Deep learning advances mean it’s now easier to discover customer intent in real-time and use that knowledge to accurately route support calls. It’s also easier to determine things like employee engagement and satisfaction.
A good deep-learning process, says Speer, “can even reveal ‘unknown unknowns’, so patterns that you weren't looking for, answers to the questions you haven't even asked yet.”
However, he warns, it's important to keep in mind that the quality of data still matters, not just the quantity: “Deep learning hasn't yet abolished the law of ‘garbage in, garbage out’.”
The AI and natural language processing industry is changing so fast it’s hard to keep up. But, Speer says, “We know that good data can be hard to find. We think that what's important to the future of deep learning is learning techniques that work well on a relatively small amount of data, that don't require gigabytes or terabytes of input before they start making sense. This is why we provide a natural-language learning process that starts out knowing a lot about what words mean, instead of trying to learn it all from your data alone.”