David Thompson David Thompson

Deep Learning and Systemic Insight

June 8, 2015

A number of months ago, I staked the following claim:

Through a combination of thoughtful physical space planning  coupled with elements of rich mobile, network, and sensor data we can  engineer the randomness of human interaction, in the hope of enriching  for serendipitous outcomes. Such outcomes will be driven by engaged  actors contextualizing previously unknown but knowable  information/data/knowledge.

There’s an emerging community of thinkers/doers exploring the  intersection of data collection, modeling, and people analytics - in the  service of engineering randomness. In the context of ‘work’, this  paragraph represented the best synthesis of my noodling at that time on  the topic.

Wait, why is work in single quotation marks?

When I talk about 'work’, I typically refer to actors engaged in  activities that contribute towards something that is bigger than any one  of them. Interestingly this is inclusive of profit and non-profit  'work’ … and it’s also inclusive of, perhaps, civic engagement.

Now, hold that thought for a moment, while we turn to Deep Learning …

Earlier this year, some researchers from Google DeepMind, published some research in Nature which demonstrated 'Human-level control through deep reinforcement learning’.  In short, a deep learning architecture was fed images from classic  Atari games. In addition the architecture was fed the game score at that  time. In general, as the number of learning iterations is increased,  the architecture determines ever-increasingly optimal strategies for  increasing the game score.

For example, in Breakout, the 'machine’ learns the 'shimmy’ (the  strategy of applying a little noise to position of the paddle), before  figuring out that popping the ball over the back of the wall is a quick  way to get rid of a lot of blocks, and increase your score. A  mind-blowing, accessible, and visual introduction to this research can  be found here.

Think about this for a moment. An algorithmic infrastructure, given  minimal input, subsequently determines the underlying rules of a system -  and then learns to navigate that system with a view to increasing an  objective function.

Now, hold that for a moment while we turn to distributed computing …

With services like Amazon Web Services we have unprecedented compute  power available to us, but it’s actually really difficult to write  software that works optimally across hundreds-of-thousands/millions of  machines. Adding compute power works when the underlying problems are  separable, or embarrassingly parallel, but less so for problems that are  intractably complex - you know, like social systems.

Interestingly, companies like [Improbable](www.improbable.io) are exploring exactly such architectures - with a view to simulating complex social problems. Improbable CEO Herman Narula explores this in a recent a16z podcast and alludes to the potential power of enriching simulated worlds with sensor and Internet of Things-enabled devices …

Now, let’s bring it all together …

What if we:

  1. Find a space to explore (space could mean workplace, city etc.)

  2. Create a sensor rich data abstraction of the workings of the space

  3. Build a simulation of the space, enriched by the rich sensor data

Such a simulated environment might be a real boon to exploring  cause/effect in a way that mirrored the true complexity of the social  nature of the problem(s) faced in spaces containing people.

Now, imagine the following (and I admit this is all tenuously sci-fi  stuff …): what if those simulations were fed into a deep learning  architecture? Could a machine then 'learn’ the unwritten rules for how  people navigate the underlying space? (I fully appreciate that the  generation of the simulation data is an additional unnecessary step -  this is perhaps most useful when the simulation is exploring the effect  of an as-yet unimplemented intervention - prospective questions might  explore whether 'the rules’ changed in (un)expected ways).

Such algorithmically generated insights could be amazingly powerful  tools for objectively exploring the spaces that contain our interactions  - and, most excitingly, all of the pieces needed are currently  available, or in active development. It’s an exciting time to be  thinking about people analytics.

Thanks for listening,

DT

Read More