Friday 22 November 2013

Natural Learning vs Institutional Learning

A while back I was lucky enough to spend some time with Roger Schank after a conference he spoke at.  More and more I am finding that a thorough appreciation of human learning, cognition, and memory is essential to the study of AI.  His books, especially Scripts Plans Goals and Understanding and Teaching Minds, are timeless essentials.

His core messages at the event were around how to improve our education system, something I take every opportunity to contribute to, and the contrast he drew between how people actually learn and how we structure education was really arresting as a way to pose the core problem:

Natural learning is:
  • Voluntary
  • Interest driven
  • Goal driven
  • Depends of failure
  • Fun
Which, on the surface, certainly resonates.

Institution learning is:
  • Involuntary
  • Based on the school's goals, not the individuals goals
  • Individuals interests ignored
  • Failure seen as bad
  • Not fun
I know a lot of people whose school experiences support at least a handful of these.

Makes you think.  I guess I will joining that PTA after all - although I guess you need to be a P first!

Saturday 17 August 2013

The business value of technology

There's already plenty of material on articulating the value of technology in a business sense, but it tends to be quite - I don't know - corporate I guess, and most focus on the justification of a particular framework or product.  That might be quite appropriate in an IT environment, but less helpful in more visceral engineering endeavors.

So what's the justification for good method, design, and computer science?  How do you map that academia to real business concerns?  A while back I made a handy reference guide for my business buddies:

I don't believe that's comprehensive, but I do believe it's prototypical.

Sunday 6 January 2013

AI and Travel

There’s been a bit of coverage recently about our Natural Language Processing search beta, but all that’s being talked about is the semantic search element.  The journey we’re embarking upon is much more ambitious than that, so I want to take a few minutes to fill in the blanks.

NLP is an important ingredient to this product, but it is not the product.  The ‘product’ is a goal-oriented artificial intelligence specialized for solving travel retail problems.  We need natural language only to provide a human-like interface into that intelligence.


We’re modeling our AI on the human-human interactions that travel agents have with real people.  But first let’s talk a little more about the concept of agency.  ‘Agency’ has a few meanings, the most important one here being an actor able to interact with the world.  We develop agents ourselves every day – subprocesses of the mind which are trained to take unsupervised control of complex tasks for which we have developed some proficiency.

Driving a car is an example most of us can relate to.  When you first start, you have to consciously direct all your actions.  Hands at 10 and 2, check the mirror, engage the clutch, watch the speedometer.  After a few years (hey – I’m a slow learner) you develop what you probably call the ‘knack’ for it and you can drive around listening to music or holding a basic conversation.  Those things you had to think so carefully about have receded back from your conscious focus, delegated to a specialized agent who frees up your attention for other things.  You can use that comfortably in any ‘like’ scenario – ie you don’t need to develop a new one when you exchange your Toyota for a Honda.  Neat mental tool and fundamental to learning.


What we’re trying to do at Expedia is mimic this feat of human intelligence with machine reasoning, to give the level of personalized service and helpful, relevant support that a customer would receive from a real, live agent.

That’s why what we’re doing here is so much more than a semantic search service; it is more like a conversation which enables a customer to start with their intent (a beach holiday, a romantic break, cheap ski vacation etc) and, through an iterative exchange of ideas in question-answer format, end up with the most suitable travel arrangements made.

This isn’t a straightforward journey.  I was recently lucky enough to spend some time with Dr. Steven Pinker discussing this at length, and we concluded that we understand the “A” but we don’t understand the “I” so this kind of project is always part research.  You have to be optimistic to be a computer scientist!

But search can only really get a little bit better before we have to make the leap to AI.  To society this is the move from easy simple access to information to the delegation of problems to agents.  Perhaps now is the right time to touch on the bigger picture, what the future might look like:

The future of search

You won't see this page anymore.  The whole search space will be superseded by a network of generalized intelligences and specialized intelligences, and search engines like Google [as you know it today] will become the back end for that network, no longer a user-facing experience.

Specialized intelligences will know how solve specific problems – they'll have what we call domain expertise – such as changing a washer in a tap, making a candle, or – ahem – planning a vacation.  They'll know how to organize loads of dissimilar data and services into the logical relationships which allow us to achieve those tasks and only those tasks.  Generalized intelligences will be responsible for marshaling these specialized intelligences, so that we don't even have to keep an index of the specialized guys.  So any time you want something you'll consult your generalized intelligence which acts kind of like your e-majordomo; interpreting your wishes, dividing them up into tractable problems, finding solvers and delegating problems to them, then assembling the answer which carries the most confidence and presenting it back to you in human language.  Kind of like how 'people' organizations work today – there are specialists who can undertake specific tasks for you and generalists who can route you to the right specialist (and sometimes have some supervisory function).  It is a pattern that we're used to.

Example; you're going to change your spark plugs (because you have a classic car – we'll all be hydrogen fuel cells by then!) so, assuming you're not an expert mechanic, you'll first look up the general principles – disconnect HT cables, unscrew old plugs, set gap on new plugs, screw in new plugs, reconnect HT cables.  Then you look up the specific details for your vehicle – Haynes manual kind of stuff – how to remove the rocker cover for easier access, how to check the cam timing etc, and then you get your tools and parts together.  You need to go any find the right plugs (obviously) but also need the correct size socket driver and gaskets and grease etc.  That's quite an assembly of information and collecting of items etc from [potentially] lots of different databases and shops.  Or you could just pose the problem to your personal AI and head straight out to the garage.  Perhaps you’ll also be receiving step-by-step instructions, via your HUD, overlaid in real time on the engine itself as you look at it.

This is long term view.  It will happen piecemeal, with sites gradually becoming more intelligent and starting to offer experiences which allow you to pose your problem, rather than hunt out information and evaluate it yourself.  Imagine an Amazon where you could just say “I have a leaky tap” and (perhaps after some Q&A to narrow down the problem) you’d be shown a pipe wrench, a pack of 1/2 inch washers, and some DIY guides showing how to apply those tools without flooding your kitchen.  Right now it shows me a book called “Death and Other Things” by Christopher Hall and a mains powered household gas detector.  Today the onus is on me to know that I need a pipe wrench etc and go looking for those items individually, accumulating as I go.

We kind of already do this.  Back to our car analogy for a second, the complex system that is the modern motor vehicle already contains a number of these agents today.  When I was young (oh no – I have become my parents) my first car had manual everything.  I had to change gears, which meant developing a feel for torque and engine speed.  I had to set the choke, which meant developing an awareness of fuel/air mixture, and I had to switch the radiator fan on and off, forcing me to have explicit knowledge of engine temperature.

In other words, driving a car used to require many more proficiencies than it does today.  The mechanical complexities are managed for us by clever (well, just clever enough…) homunculi so all we need to do is point it in the right direction and push one lever to go faster and another to slow down.

Back to travel

The most important question to ask about any advancement in technology is how it will improve the human experience.  A future in which specialized intelligences take away more of our common problems is essentially connecting intent to effect by fewer intermediary steps.  Appeals to the lazy in all of us.

From our adventures in machine learning specifically, we expect to be able to benefit the traveler and the travel business by:

  • A better user experience; expert guidance through travel planning instead of imposing a significant research burden upon the traveler.
  • More intuitive interface which can be entirely cohesive across dissimilar devices.
  • Higher conversion rates on sites and apps.
  • Faster support for travelers in-trip delivered at a lower cost.

To do this well any machine learning algorithm needs a body of knowledge to train it.  The more comprehensive that body of knowledge, the more confidence you can have in the answers an AI will produce.  Expedia is the world’s largest travel company and has been selling to and supporting millions of travelers for over 15 years.  That experience is captured in petabytes of data generated by instrumenting every aspect of the travel experience.  This puts us in a really good position to do something meaningful for the web travel universe with artificial intelligence, as well as contribute to machine learning and machine reasoning disciplines.

That’s what our innovation labs programme is all about benefiting our partners and contributing to our profession.