At DISCO, we want to be the legal technology company providing solutions that save legal professionals’ time throughout the life of a matter, from collection to discovery to litigation. That’s why we created an artificial intelligence platform, DISCO AI – our answer to predictive coding that layers machine learning technology on top of our revolutionary ediscovery interface. DISCO AI is built on the Deep Learning technology that is transforming academic and commercial machine learning efforts in fields from self-driving cars to cancer diagnostics.
Understanding how Deep Learning works
Machine learning is about recognizing patterns in data. The key insight to Deep Learning is that serious learning tasks depend on patterns that are deep (complex) and not shallow. It is perhaps easiest to convey the distinction by talking about images. Imagine writing a program to detect faces. A shallow approach would attempt to find the whole face in the content directly. Such an approach wouldn’t be able to account well for new types of faces, for example, someone wearing an eyepatch. By contrast, Deep Learning systems learn to recognize noses, mouths, ears, hair, and eyebrows, and they learn to assemble these parts together to form a face.
Put otherwise, a deep learning system discovers secondary features in order to support its primary judgements. The patterns it finds are complex and have many parts. Patterns can be matched when most of the parts are present, and the system isn’t thrown off the track just because of some small deviation from its prior experience. That is to say, context matters, and Deep Learning does a better job of finding relevant context.
The same ideas apply to text as well as images.
How is DISCO AI different than other machine learning tools in our industry?
Most of our competitors use shallow machine learning techniques. These methods are usually based solely on the words contained in a document, the so-called Bag-of-Words model. These systems are looking for a particular combination of words that do or don’t appear. Specific technologies include support vector machines (SVMs) and latent semantic indexing (LSI).DISCO AI doesn’t use the Bag-of-Words simplification. Instead, we convert words into meaning using Google’s Word2Vec technology. To understand Word2Vec, imagine trying to write every word in the English language on a huge sheet of paper so that words that mean similar things end up close to each other. That’s basically what Word2Vec does, except that it’s a multi-dimensional sheet of paper and not just 2-D. Each word can then be replaced by its location on the paper, and since words with similar meanings are close to each other, these locations let DISCO AI see meanings and not just words.
These meanings are then fed into a Convolutional Neural Network (CNN), which is one of the most popular forms of Deep Learning. CNNs are built out of many layers of pattern recognizers stacked on top of each other. Convolutional is a fancy way of saying that the machine looks at small parts of a document first rather than trying to account for the whole thing. Each successive layer combines information from these small parts to fill in the bigger picture and assemble complex patterns of meaning, just as was described for images above.
With just the document text, we’ve proven internally that DISCO’s CNNs make better predictions than competing technologies such as SVMs. But DISCO AI doesn’t stop with the document text. It also includes information about document metadata, such as the type of document, the custodian, the date, and more. Because of the modular structure of deep neural networks, we can plug all kinds of background information into our prediction engine that are hidden from our competitors’ shallow methods.
So DISCO AI analyzes your documents deeply, searching for complex patterns of meaning that our competitors can’t identify. We believe that translates into saving you time and your clients’ money, all while making the world a better place.
Want to see DISCO AI in action?
View our webinar, DISCO AI: Deep Learning for Legal Technology