Mole End

Some blog posts are for the ages, others are ephemeral. The Mole acknowledges this post's place in the latter category, yet still feels it appropriate to write today about what he heard yesterday even if it is relevant only until tomorrow.

Last night's meetup — yes, the third this week — was hosted at Yelp under the auspices of SF Bayarea Machine Learning; Ameet Talwalkar demonstrated the slideware version of MLBase.  One of the lessons the Mole learned at some point between late academia and early industry was to speak only of what he had done, not what he was going to do.  Practitioners like and want and need to know what has already been accomplished, preferably accompanied with a link to download that version of the code.  A broad agenda to build a better mousetrap is greeted with skepticism.  So while the Mole wishes the MLBase team all the best, he will reserve judgment about the efficacy of their remarkable new universal can opener until he can install it somewhere.

Oddly and ironically, the Mole noted some small evidence for one of the bluntly stated propositions of the talk the previous day.  The speaker there alleged that the machine learning algorithms in open source tools invariably lag the state of the art.  The algorithms promised for the first release of MLBase fit that pattern; the open question is whether they will be only a starting point or representative of what the system will deliver.  

Expect a comment later this summer when the first public version of MLBase is released.  The Mole hopes to be proven wrong in his doubts.

The Mole is a retiring and self-effacing sort, quick to minimize his own acheivements (such as they are), so he will stipulate without hesitation that Dr. Alexander Gray is more accomplished and more knowledgeable about machine learning than he is.  But the Mole has also known quite a few members of the National Academies -- and even one Nobel laureate -- who have not found it necessary to proclaim their intellectual superiority quite so sneeringly.

The Mole has also seen and heard people describe their technical innovations in ways that were more compelling than blending self-congratulation and hand-waving.  At the very least a table of timings showing product X running much faster than product Y would tend to induce a belief that the presenter had run the the two products head-to-head.

So whence this little jeremiad?  The Big Data Gurus meetup last night at Samsung titled "Real-World Machine Learning on Big Data: Which Method(s) Should You Use?".  The taxonomic analyses (ML tasks and their corresponding methods, parametric vs. non-parametric methods) were interesting -- sufficiently so that the Mole will go digging through YouTube to find a version of the talk and copy the information down since the speaker indicated an intention not to share his slides.  And a couple of hints dropped during Q&A appear to be well worth investigating further:  a forthcoming report from the National Academies on massive data (did he mean "New Tools for the Analysis of Massive Data", which is a year and a half overdue?) and the CMU TETRAD project led by philosopher Peter Sprirtes.

Overall it was a very good pitch for BigML, or wise.io, or Precog, or Alteryx, or any other competitor to Skytree, and good motivation to go contribute to the open source big data machine learning tools so derisively dismissed.

Last night the Mole attended a pleasant talk by Mark Noworolski on Streetline's smart parking system.  (The talk was presented through the good offices of the IoTSiliconValley meetup, capably organized by Drew Johnson and Elle Wood, and hosted by Hacker Dojo.  </shoutouts>)  While the talk was not particularly technically demanding, the Mole was quite impressed by the expansiveness of Streetline's vision.  

Stipulating up front that the devil is in the details and with a tip of the hat to Streetline's six years of hard work, the general technical architecture was exactly what the Mole expected:  

• Wireless sensor nodes in or on the street,
• Connected as a mesh network to wireless repeaters off the ground,
• Connected to wired backbone nodes,
• Feeding up into Streetline's cloud,
• Delivered via web and mobile applications.

Two strategic decisions are particularly interesting:

• Streetline has built almost the full vertical stack.
  • While the components in the sensor nodes are COTS, even the firmware is bespoke.  (Sorry, Contiki, TinyOS, et al.)
  • Everything from there up is Streetline's:  the hardware, the analytic systems (though there is some sort of partnership with IBM Cognos), the delivery platforms, etc.
  • At the analytic and application layers they do rely on (unspecified) open source tools, frameworks, and languages.
• Streetline is pursuing a pure an almost pure aaS play:  apart from a de minimis installation fee, their business model is to charge a recurring per-month per-space fee.

The verticality is interesting because it indicates a belief that the increased development costs will be repaid by the improvements made possible on the feature side (better reliability, control of enhancements, etc.).  This runs against the grain of trendy lean startup MVP notions, reflecting some combination of the grownup big company backgrounds of Streetline's founders and the stolidity of the customers.  

The service play is interesting because of the capital exposure Streetline faces.  The Mole would love to see their internal numbers:  either they're in a position to loan substantial amounts of money to municipalities for an extended period or they've laid their hands on the goose that lays the golden eggs.  If they can recover the costs of installing a network quickly their margins will tilt sharply upwards the next day.  That allows them to balance strong cash flows with significant ongoing product R&D.

So where is the "expansive vision"?  Streetline have thought deeply about the needs of both their paying customers (i.e., municipalities, private parking operators, etc.) and the end users of the system.  They're even evolving best-practice guidance for how to use the technology (the Mole asked about this), advising cities, for example, not to zero meters when cars pull out.  Although technically possible, the bad feeling engendered among drivers (who feel that finding some time on a meter is a "God-given right") isn't worth the extra revenue.  

They're also working towards a broader sensor ecosystem that will incorporate both Streetline technology and others'.  In addition to creating additional functionality such as traffic flow sensors and management tools they are thinking about opening up (the Mole hopes he got this right...) a network level API that would allow other sensors to ship their data into the Streetline cloud.

In the medium term the Mole predicts moderate chaos as vertically integrated players like Streetline jostle not only with direct competitors (FastPrk, Fybr) and niche endeavors (SFpark), and with lateral moves from nearby domains (Sensys Networks, Iteris), but also with parallel innovations (Parkopedia).  Plan to install multiple parking apps on your smartphone...

And one final nugget of great value:  in the context of advocating "version early, version everything", Mark pointed the crowd to Tom Preston-Werner's Semantic Versioning.  Read it.  Follow it.