ZF Intensive course on Machine Learning: brainstorm

(From Rick Donnelly)

We talked earlier this year about the idea of a focused workshop on machine learning in travel forecasting. Eric Miller and I have batted this idea around over drinks on and off again for two years, as I’ve been keen to put something like this together with him using Transport Canada funding. Alas, we've been crazy busy enough that it stubbornly remains right beyond our focus. Then the idea surfaced again in your list of possible Zephyr activities. I've talked to several folks since then to get an idea of what might be entailed.

The goal of such a workshop would be to impart enthusiasm and skills through one or more focused projects that participants would complete together and showcase to the profession. My vision for this is a mashup of successes in other realms, including the TRB Chan Wui & Yunyin Rising Star Workshop, Data Science for Social Good (DSSG), the Santa Fe Institute's Complex Systems Summer School (CSSS), the TNU Computational Psychiatry course (which I’m very seriously considering for myself this summer), and of course the week-long discrete choice modeling courses at MIT, Sydney, and in Europe. And those are only the ones I’m familiar with. They’re each aimed at different audiences, of course, but I see several common threads: Competition appears to be stiff for attendance at all of them Leading firms or labs in each industry compete for the graduates All are heavily underwritten by grants or corporate sponsors The course cost is very pricey by our industry standards (although worth every penny of it) All are held in academic settings, with several offering on-campus accommodations All have top faculty and industry leaders teaching the course Most are designed for graduates of other programs, and for high-achievers rather than mainstream students Most have production of either individual or group portfolios as key goal Many are geared towards cultivating academics rather than practitioners Although all are full-time intensives, the length of the courses I looked at varied considerably. They range from one week to an entire summer. My guess is that a two-week course would be ideal for us, although 3-4 weeks doesn't appear out of the question (but pricier).

I don't see any of these factors as particularly onerous for us to match other than the financial hurdles. Ideally Transport Canada or FHWA would put together a scholarship program to cover cost of public sector participants. Doing so would ease the financial pain and make the scholarship more palatable to state and local governments, who would view accepting private money from any source as problematic gifts. Obtaining the seed money will admittedly be challenging, but I’d rather abandon this idea only after exhausting a concerted effort rather than assuming it as foregone conclusion if that is only serious objection.

I like the DSSG model the best. Like the CSSS, it is 12 weeks in duration. However, they are unique in their approach of combining cram courses in data science with real-world projects. They’re trying to emulate TED by expanding the course to select locations and wanting to have mini-conferences showcasing their projects. I’d really like to emulate that in our realm. We could put together something close to DSSG — but a quarter of the duration — like “data science for transport planning” (DSTP), which would have more immediate appeal to broader audience in our profession, or a more focused “machine learning for transport planning” (MLTP). In either case we’d present a series of lectures and labs on fundamentals taught be industry experts (read: from outside our profession), followed by one or more projects that the participants would work on. If we used Uber or Lyft data, perhaps combined with other public or private data, we’d have endless possibilities for very practical projects that participants could adapt over and over in their careers.

I see this as a great target for young and emerging professionals in our field. If we had 16-20 participants we'd be probably be running at max capacity. If we could team up a few data science educators with a few of us (to run the projects) we might have a great shot at this. I haven’t thought much about venues for this, although several come quickly to mind — including partnering with DSSG or SFI. I’m guessing we could put together a 2-3 week program that would cover the essentials, which is as long as most public sector folks would be able to escape for in any case. I’m still keen to see if we can get Transport Canada to underwrite an initial offering. If we succeeded there we could export it it far cheaper than we’d be able to create it. I’m not sure we can get the same federal sponsorship in the USA, if for no other reason than lack of sense of urgency at USDOT to get smart about this.

I don’t know where to take this from here. Almost everyone was evasive about their actual costs, so we’ll need to dig deeper for that or work up cost estimate ourselves if we take this to next step. I’ve spent a lot of time investigating these programs, but don’t understand ZF well enough to know how to take this from idea statement like this to solid proposal we can shop around, or whether Zephyr board of directors will support this. How do we take the next step?

ZephyrTransport / management

ZF Intensive course on Machine Learning: brainstorm #14