#TRBAM through a neural network
Note: This is a personal post and all opinions expressed here are my own — though if my employer has an opinion on “Secision Embinem Ridesharing Skop?”, I’d love to hear it.
Each January, the Transportation Research Board Annual Meeting (TRBAM) convenes over 10,000 practitioners to discuss every transportation-related topic imaginable: from road and rail infrastructure, to transit and land use, to equity and safety, and everything in between. I was thrilled to attend my third TRBAM last week, but as always, going to this conference is akin to drinking from the proverbial firehose — there are far too many sessions in just my field for one person to attend, let alone the hundreds of others. At over 450 pages, even the mere list of sessions, talks, and posters rivals my copy of Moby Dick in its length and technical inscrutability.
Since I’m a fan of Janelle Shane’s AI-generated recipe titles, April Fool’s pranks, and other delightful weirdness, this massive, semi-structured text document got me thinking: what would a neural network do with a TRBAM training set? I’ve been wanting to dip my toe in the machine learning space for a while now, so I extracted and cleaned the text, installed char-rnn, and gave it a go. Not surprisingly, the results were pretty amusing — especially after spending a week immersed in the transportation world.
Presentation Titles
As a first step, I trained a simple model on a list of just over 5,000 talk and poster titles. Sampling the model with low “temperature” (i.e. not letting it be too creative) resulted in a bunch of permutations of the most common conference buzzwords: lots of asphalt, pavement, and autonomous vehicles:
The Impact of Autonomous Vehicles on the Controlled Traffic Signal Projects
A Comparison of Predictive Model for Asphalt Mixtures
A Comparison of Asphalt Mixture Modeling of Connected and Automated Vehicles
Sensitivity of Social Conditions for Asphalt Mixtures
The Impact of Measuring Performance of Automated Vehicles Using Cranking
Lightning Talk: An Emission Study of Resilience on Connected Vehicle Scale Considering Probe for Pavement Systems
Turning up the temperature produced decidedly more interesting results (including less-common buzzwords and some intriguing neologisms):
Pedestrian Modified Random Highways
Study of Driving?
Using Supply Goundability in the Caramon Methods
Intergation of Doof Applications on Developing New York Networks
Soilt Prestant Crosswalk
Rase Simulation Considering the Effect of Hazardous Bikeshare, Thuid ections and Crossings
Hyperloop for Sensitivity of Warbal Vehicles
Laboratory Evaluation of Pavement Lust Impact Estimation
Improvements of New York?, Rail-Time
On Lane Behavior to Identify the Weaving and Convisuolang and Technology
Pavement Recycling Safety Office Control Distraction
Connected and Automated Vehicles: A Predictive Contasting of Propided Cracker Solutions
Dogulation Model for Characterization of Study and Local Road Factors
Development of the Travel Control of Blotian Pavements?
Effects of Porking Monitoring of the Cracking Test for Farture Examples
Secision Embinem Ridesharing Skop?
Crossy Crashes
Attracting Light Puttoring Planning in Origin–Destination Technology Utilization
Enhancing Tame Floating the Trougle-Based Effects of Povestion of Compaction Problem
An Importance of Horizontal Level of Freeway Experiences
Travel Simulator of Pavement Mental Analysis: Development of Electric Intersections and Celling Car Follities?
Finally, given that bikesharing and bicycles were popular topics in some TRBAM sessions (and are topics of personal interest to me), I tried several rounds of priming the model with either Bikeshar or Bicycl and letting it finish the title:
Bikeshare Policies for Self-Healings
Bicycle Bikeshare Sharing
Bikeshare Technology for Asphalt Concrete Taxis: The Case of Houthor in California’s Safe Aggregate
Bikeshare Transportation: A Florida
Bikeshare Control for Ship Beet Friction Evaluation
Bikeshare Bikeshare
Bikeshare Developing Airline Cot Grids
Bikesharing: A Case Study of Wrong Technologies [Ed.’s note: 😞]
Bicycling?: Evidence from New Yarking Evaluation and Travel Recognition Systems
Bicycle Portland in the Simulated Gradient Internet
Bicycle Cressing Using Connected Automated Bridges
Bicyclists in Connected and Automated Vehicles: Shateline Systems
Bicycles System Toward Texas Using Mapping Teramelation with Commercial Strength
Names and Affiliations
Since presenters were listed after each talk and poster title (in the format [Name]/[Organization]), it was straightforward to extract these and train another model. The neural network came up with a long list of esteemed, lesser-known researchers and institutions:
Marge Matnoran/University of Illinois, Thing
Beargen Hister/University of National Authority, Inc.
Kelrix Barnigerzi/U.S. Department of New Jersey
Batelon Bargheet/Bad Rola Department of Transportation
Uhe Indrester/University of Now York, LLC
David Benter/Transportation Department of Transportation
Sare Sanen/University of Ternsportation (FHWA)
Carin Car/University of Terntate of Transportity
Tenn Sane/University of Texan Aditerrantity of Transportition (FHA)
I especially liked its take on California-based organizations:
Laganmos XcLanply/University of Cultfornia
Arde Stitacher/University of California, Brongeley
Kegon Smahs/University of California, Burkerey
Rachanl Carler/University of California, California
Savew Chahnalin/University of California, Mankelley
Qie Hingcey/University of National California
Owren Kenson/University of New Bork Carifornia
Buttiinal Fling/California Scape University
Jasesh Schillop/Caryfonnia Department of Transportation
Session Descriptions
Having achieved some success (or at least entertainment) in reproducing session titles and affiliations, I decided to try session descriptions. This is significantly more difficult, since the descriptions are more complex (consisting of a title, leader/affiliation, sponsor(s), and an optional freeform description), and there are far fewer of them — roughly 800, as opposed to the 5,000+ talks/posters. Even so, the neural network did a surprisingly good job at reconstructing the basic format. Here’s one result from a simple model, sampled at low temperature:
State DOTs and Project Construction and Projects
John Shanne, University of National, presiding
Sponsored By Standing Committee on Transportation Data and Information Systems
This session will explore the projects and and the project development and the transportation agencies and the survey and the project design and and and the the project development and and analysis and state and the presentations of the projects and the projects and are a services and projects and the project development and …. [continues ad infinitum]
After trying a few different model variations and cranking up the temperature, some amusing (and, sometimes, not entirely unrealistic) descriptions result:
Mobility and Emerging the Automated Vehicles
Jahis Parlinos, Fasters Geometria, presiding
Sponsored By Standing Committee on Transportation Data and Information Systems, Standing Committee on Highway Traffic Monitoring
This session will provide experts and papers and a concrete to be the international performance of the travel agencies, and papers that will be used to emerge state and solitionally, and connectivity of their and transportation infrastructure.Asphalt?
Reathing Binkrals, Charging Transportation Systems, presiding
Sponsored By Standing Committee on Transportation Planning and Energy and Visualization
Session will be will be a facilitated to develop and activities in the transportation research response on proressing practices and incorporating marking sessions (160).Transportation Infrastructure Pavements and Public Transportation Sores
Yannifley, Texas Department of Transportation (AUA), presiding
Sponsored By Papers And Environmental InstituteProjects and Project Modeling and Projects
Palatic Use Malara, University of Carifornia, University of Technology, presiding
Sponsored By Standing Committee on Network BradgesBonk 2519: Haf Might Practices
Mily Soniorni, University of Nowcrorster, presiding
Sponsored By Standing Committee on Surface Crais and Underground Emergency From TransportationAmerican Productivity and Size Discussion
Ading Reelkillo, University of New Performance, presiding
Sponsored By Standing Committee on ADD40, Subcommittee on Community Testing
In this session force on emergency transportation planning considerations that socual apply hand as the construction.Concrete Freeways
Elest Departments of Concrete, Inc., presiding
Sponsored By Standing Committee on Maintenance and Recent ResearchTransportation Experience May Response Asphalt Toons for Smark Dent Roadways
New Garbini, Federal Highway Administration (FHWA), presiding
Sponsored By Standing Committee on Travel Many Highway Administration Institute and Futures
Conclusion
While we may not be ready to turn over the entire TRBAM agenda to a neural network anytime soon, I was impressed by how even a rudimentary first stab at this technique was able to (a) quickly distill some of the most common recurring themes at this massive conference (similar to a word cloud, but less annoying), and (b) produce somewhat readable (often hilarious, and sometimes strikingly plausible) talk titles and session descriptions. But its predictive power has yet to be tested; we’ll have to wait and see if “dogulation models” are a hot topic at next year’s TRBAM!
As noted above, this post (and the methodology behind it) were inspired by Janelle Shane’s terrific AiWeirdness blog. This might be a one-off for me, so be sure to check out her site and mailing list for a regularly-updated treasure trove of bizarre and hilarious neural network behavior!