Machine learning is moving out of the backend shadows and into mainstream applications that ordinary consumers will use on a daily basis. These applications will recognize the intent behind written and spoken language and quickly classify the content of images and video. It would be a leap to say that these systems "understand" the data they process, but a recent academic paper on deep learning claims just that.
Most important for non-technical users, machine learning has gained traction in natural language-intensive forms of communication, like email and social media. These messages can now be classified in terms of things like how the user feels about something, what the user is interested in and what they are specifically asking for. Machine learning is behind the ongoing improvement of Google search queries and Facebook's active filtering of its users' newsfeeds, but we don't perceive this technology as the core of those applications.
Two new business-oriented services place machine learning in the foreground to help make customer communication more efficient. San Francisco-based startup Parakweet has just launched an automated email assistant called InboxVudu. A bit south, near Palo Alto, a startup called Guesswork is building a platform on top of Google's Prediction API that improve the quality of lead generation and other CRM functions. Both companies are combining machine learning with natural language processing (NLP) technologies to understand user intent in specific contexts. In the first part of this story, I will discuss InboxVudu and then take on Guesswork in part 2.
One of the biggest challenges in terms of understanding machine learning and artificial intelligence (AI) is the natural tendency we have as humans to leap to conclusions about the capacities of these technologies. Deep learning, for instance, now can discover features, like the faces of cats or the sides of automobiles, in large collections of images without human intervention. This ability is impressive, but the world is not made of homogenous sets of data. Nor does classifying cat faces imply that our AI can now ingest the entire internet and learn Sanskrit.
The rise of deep learning has followed the availability of massive datasets in the hands of companies like Google and Facebook that possess sufficient scale and structure from which cat faces can emerge. Most companies have to deal with a lot of what AI innovator Lars Hard of Expertmaker refers to as "little data"; the kind of domain-specific knowledge that does not scale reliably. Think about what you need to know to answer your email effectively. You have colleagues and customers that you deal with, and there is a lot of data in your mind that describes those relationships. You may feel compelled to respond to some people right away, but others are less urgent. You might be working on a time-sensitive project with someone, so they temporarily move up the priority list. These kinds of relationships are very hard to model and keep updated.
In contrast, InboxVudu substitutes an easier problem: helping you keep track of all the things these various people are asking you for through email. Some messages are FYI and do not require a response. This application prioritizes the ones that do. InboxVudu rides on top of your existing email account (currently only Gmail or Google Apps accounts) and delivers a digest of all unanswered messages that require a specific response at the end of each work day. This format makes it easy to catch up at the end of the day or first thing the following morning on any urgent requests.
Kiam Choo, CTO and co-founder of Parakweet, explained to me that the company has built its own algorithms for "sentence-based intent recognition." InboxVudu processes the messages in your Gmail account which it parses, sentence by sentence, looking for "asks." An "ask" is text feature that indicates the asking of a question or a the need to fulfill a request. Choo studied AI at University of Toronto with neural nets pioneer Geoff Hinton (who now masterminds deep learning at Google) so building on Google's own machine intelligence is not a leap.
Because there is a vast amount of text in the world that contains "asks," a statistical machine learning approach works well. But merely identifying these text features is not that useful if there are a preponderance of false positives. Choo tells me that InboxVudu also uses a machine learning technique called "support vector machines" (SVMs) to weed out marketing emails. It also uses rule-based algorithms to remove the requests of "muted" senders as well as keep track of which messages no longer require a response. Finally, users can mark mistaken "asks" as false positives to improve the training of the system. In machine learning systems, human input is used to correct errors through a technique called back propagation. As a user, your feedback is a critical component of the learning of the system, so don't treat these invitations for data collection as optional!
The digest shows you just the header information from the selected messages along with a snippet of the message text with the "ask" bolded. Although you can respond to individual messages from the email digest itself, the easier workflow is to go to the linked web page that has more javascript-enabled functionality. This page operates like a real-time checklist and allows you the satisfaction of seeing the messages disappear as you resolve them.
What is particularly useful about InboxVudu is that it requires virtually no setup to start using. Since it works on top of Gmail it picks up your credentials through OAuth if you are logged into your Google account (which you probably are if you are using Gmail or Google Apps). Sign up takes 10-seconds, and you get your first digest the same day.
The low barrier to entry makes this an easy supplemental product to try. (Parakweet has provided Forbes.com readers with priority access to the beta with this link.) Because it rides on top of Gmail, InboxVudu can take advantage of Google features without breaking. Google, for instance, introduced a priority inbox feature many years ago that predicts which conversations are important, and its accuracy at this point is very high. My own experience with InboxVudu improved when I switched my inbox settings to the priority setup.
Google's APIs have made it easy for developers to build products that hook into Gmail's functionality. InboxVudu is not a replacement for the many getting things done (GTD) -inspired Gmail and task list solutions out there, from Mailbox to Todist, Handle and the just-released Sortd. Gmail itself has introduced many innovations recently including tabbed inboxes and its new Inbox app. Inbox takes a slightly different approach by turning structured information in email messages into "cards" similar to those in Google Search, Google Now and other Android apps. Given the fact that Google's own Prediction API includes features like sentiment analysis, spam detection, message routing and classification, it is likely that it will eventually incorporate InboVudu-like features into Gmail itself.
This is the risk that any startup makes when it builds its product on top of a major offering by Google or Facebook or another major tech company. The flip side, of course, is that these are the major flow channels in place. If you want to help fix email, Gmail is the place to start. And if it can work for Gmail it can likely work for Slack or other messaging services. Parakweet's gambit is that working on understanding intent in natural language will have value beyond a specific platform.
Interestingly, Parakweet's first product, BookVibe uses sentiment analysis of social media feeds (first Twitter and now Facebook) to make book recommendations. Sentiment analysis, supplied by the Google Prediction API, is an aspect of Guesswork's offerings, as well, as I shall discuss in part 2 of this story.
