Mixpanel Essay Contains Several Smart Software Gems

August 11, 2016

I read “The Hard Thing about Machine Learning.” The essay explains the history of machine learning at Mixpanel. Mixpanel is a business analytics company. Embedded in the write up are several observations which I thought warranted highlighting.

The first point is the blunt reminded that machine learning requires humans—typically humans with specialist skills—to make smart software work as expected. The humans have to figure out what problem they and the numerical recipes are supposed to solve.  Mixpanel says:

machine learning isn’t some sentient robot that does this all on its own. Behind every good machine learning model is a team of engineers that took a long thoughtful look at the problem and crafted the right model that will get better at solving the problem the more it encounters it. And finding that problem and crafting the right model is what makes machine learning really hard.

The second pink circle in my copy of the essay corralled this observation:

The broader the problem, the more universal the model needs to be. But the more universal the model, the less accurate it is for each particular instance. The hard part of machine learning is thinking about a problem critically, crafting a model to solve the problem, finding how that model breaks, and then updating it to work better. A universal model can’t do that.

I think this means that machine learning works on quite narrow, generally tidy problems. Anyone who has worked with the mid 1990s Autonomy IDOL system knows that as content flows into a properly trained system, that “properly trained” system can start to throw some imprecise and off-point outputs. The fix is to retrain the system on a properly narrowed data set. Failure to do this would cause users to scratch their heads because they could not figure out how their query about computer terminals generated outputs about railroad costs. The key is the word “terminal” and increasingly diverse content flowing into the system.

The third point received a check mark from this intrepid reader:

Correlation does not imply causation.

Interesting. I think one of my college professors in 1962 made a similar statement. Pricing for Mixpanel begins at $600 per month for four million data points.

Stephen E Arnold, August 11, 2016

Comments

Comments are closed.

  • Archives

  • Recent Posts

  • Meta