Why now is the time for dialog

I'm working on a task-oriented dialog product and things are going surprisingly well from a business standpoint. It turns out that existing techniques are sufficient to substitute some portion of commercial dialog interactions from human to machine mediated, with tremendous associated cost savings which exceed the cost of developing the automatic systems. Here's the thing that is puzzling: the surplus is so large that, as far as I can tell, it would have been viable to do this 10 years ago with then-current techniques. All the new fancy AI stuff helps, but only to improve the margins. So how come these businesses didn't appear 10 years ago?

I suspect the answer is that a format shift has occurred away from physical transactions and voice mediated interactions to digital transactions and chat mediated interactions.

The movement away from voice is very important: if we had to try and do this using ASR, even today, it probably wouldn't work. Fortunately, today you chat with your cable company rather than talking to them. That shift was motivated by cost savings: a human agent can handle multiple concurrent chat sessions more easily than multiple concurrent voice conversations. However it requires most of your customers to have a computer, smartphone, or other device rather than an old-school telephone.

The continuing dominance of e-commerce over physical stores is also a factor (RIP Sears). In e-commerce, human salespersons increasingly assist customers in transactions via live chat interfaces. Once again, what starts as a more effective way of deploying human resources becomes the vector by which automation increasingly handles the workload.

The end game here is that the number of people employed in retail goes down, but that their compensation goes up. That is because the machines will increasingly handle the routine aspects of these domains, leaving only the long tail of extremely idiosyncratic issues for the humans to resolve. Handling these non-routine issues will require more skill and experience and therefore demand higher compensation (also, an increasing part of the job will be to structure the torso of non-routine issues into something that the machines can handle routinely, i.e., teaching the machines to handle more; this is analogous to programming and will also demand higher compensation).

Microsoft starts an ML blog and an ML product

My employer, Microsoft, has started a new blog around ML and also announced a new product for ML.

The blog is exciting, as there are multiple ML luminaries at Microsoft who will hopefully contribute. Joseph Sirosh is also involved so there will presumably be a healthy mix of application oriented content as well.

The product is also exciting. However if you are an ML expert already comfortable with a particular toolchain, you might wonder why the world needs this product. Those who work at large companies like Microsoft, Google, Facebook, or Yahoo are presumably aware that there is an army of engineers who maintain and improve the systems infrastructure underlying the data science (e.g., data collection, ingest and organization; automated model retraining and deployment; monitoring and quality assurance; production experimentation). However if you've never worked at a startup then you aren't really aware of how much work all those people are doing to enable data science. If those functions become available as part of a service offering, than an individual data scientist with a hot idea has a chance of competing with the big guys. More realistically, given my experience at startups, the individual data scientist will have a chance to determine that their hot idea is not so hot before having to invest large amount of capital developing infrastructure :)

Of course there is a lot more that has to happen for “Machine Learning as a Service” to be fully mature, but this product announcement is a nice first step.

Stealth No More!

The startup I'm currently at publicly launched today. It's a social image sharing site called LoveIt. This is a crowded space at the moment, but we've tried to throw in some innovative new features. One machine learning related bit that I worked on is the recommendation system; here's an example screenshot with the recommendations in the bottom right hand side.



The image for a mashup by DJ Earworm (who is totally awesome!). In this case the second recommendation is a music collection which is very sensible, but the first recommendation is more questionable (focusing on the costume ball aspect). Hopefully the system will get better as we generate more behavioral data exhaust. I have noticed image recommendation is more forgiving than text recommendation: images have less precise meaning so people are more willing to invent why a quirky recommendation makes sense.

Conceptually the system is heavily Elkan inspired. The implementation is a combination of Elasticsearch and Vowpal Wabbit, strung together with Erlang. The tricky part is getting it to compute something quickly (circa 100ms), and both Elasticsearch and Vowpal Wabbit are excellent pieces of software in this regard!

The Bigger Picture

When I first started on the internet, the most common demand for machine learning I encountered was for optimizing performance marketing (the other big one would have been algorithmic search, but southern California wasn't a major player in that space). Nowadays there are many big smart companies focused on the science of advertising. In my opinion, if you have some machine learning acumen and some plucky post-series-A startup claiming to revolutionize internet advertising with a new algorithm attempts to recruit you, run the other way! There are probably still many smaller exits to be had in this space selling to the major ad networks, but unless you have a large equity share it won't change your life.

Fortunately there is a new nexus of ubiquitous machine learning need: content recommendation, personalization, summarization, and visualization. This is driven by the intersection of several trends, including the rise in user-generated content, social networks, and smartphones. For example, Twitter has turned everybody into an intelligence analyst lost in a sea of intercepts. Technologies that can scan all of Twitter and surface the (personalized) good stuff in real-time would be very interesting. Furthermore, as Google has proven, if you position yourself as a trusted discovery tool for users you can easily monetize. Thus if you get a recruiting call from a startup claiming to attack such problems, my advice is to seriously consider it.