Morgan Herlocker

@morganherlocker

geometry, graphs, flowers, etc. - writes about privacy & algorithms - & creator of - previously - he/him

Oakland, CA
Vrijeme pridruživanja: veljača 2011.

Tweetovi

Blokirali ste korisnika/cu @morganherlocker

Jeste li sigurni da želite vidjeti te tweetove? Time nećete deblokirati korisnika/cu @morganherlocker

  1. The effects of this are playing out with "smart cities" and "enterprise AI" hitting the wall of disappointment building their own telemetry-driven systems. The effective data driven systems you see day to day were carefully chosen and scoped. There are no silver bullets.

    Prikaži ovu nit
    Poništi
  2. Model creators have spent a lot of effort making these systems seem magic and omniscient, when they are often more human-centered and limited than they might seem at first glance.

    Prikaži ovu nit
    Poništi
  3. Models and "AI" are influenced by user perceptions of accuracy and completeness just as much as actual measured metrics, and I think this is often overlooked.

    Prikaži ovu nit
    Poništi
  4. In the case of congestion, we would often painstakingly avoid reporting "red" when we were not sure, and user testing showed that people don't complain about false predictions but are livid when they notice a missed prediction (no "red" when there is a jam).

    Prikaži ovu nit
    Poništi
  5. Overall, I think the takeaway of this piece should be that these models are unreliable, not because of some evil entity or AI ham-handedly controlling society, but because modeling behavior is difficult and non-deterministic. They are still accurate enough to be useful.

    Prikaži ovu nit
    Poništi
  6. Should be noted that there is a naturally occurring "wagon full of phones" scenario which is "buses/trains full of phones", and it is difficult to filter reliably.

    Prikaži ovu nit
    Poništi
  7. Also, while everyone was talking policy implications of modality classification error rates, I think they missed google's bidirectional edge speed aggregation. I suspect this has more impact on false congestion prediction rates overall, not that I disagree w/ the design choice.

    Prikaži ovu nit
    Poništi
  8. *dude walks down street with wagon full of 100 burner phones*

    Prikaži ovu nit
    Poništi
  9. "Our model can predict the correct mode 99.1% of the time, representing state of the art accuracy."

    Prikaži ovu nit
    Poništi
  10. Having spent a bunch of time developing modality classification models, I have spent an absurd amount of time imagining (and worrying about) this exact scenario. Predicting vehicle type at low speeds is quite difficult.

    Prikaži ovu nit
    Poništi
  11. proslijedio/la je Tweet
    31. sij

    OpenStreetMap in Milwaukee is amazing. You have curb edges! Sidewalks! Medians! Trees! Building footprints! Green space! How is it not this good everywhere?!

    Poništi
  12. 28. sij

    hot take: the amazon ring is bad, but for different reasons than the big report this week

    Poništi
  13. 28. sij

    first pass visualizing the incremental graph normalization algorithm in the new road network conflation library I've been working on

    Poništi
  14. proslijedio/la je Tweet
    21. sij

    We're hiring! is looking for a software engineer. Help us build open source tools to reinvent how cities collect and share transportation data. 👇

    Poništi
  15. 23. sij

    most exciting map tech of 2020 is the new MS flight simulator

    Poništi
  16. 23. sij

    added a debug log "terminal" for printing values alongside the geometric display - style can be driven by data, which is helpful for debugging many types of geospatial algorithms

    Poništi
  17. 21. sij
    Prikaži ovu nit
    Poništi
  18. 21. sij

    I write something like this for nearly every algorithm-oriented library I write, so it made sense to package up the flow into something reusable. I may be the only user, but it will be open source with a cross-language workflow so others can use it too.

    Prikaži ovu nit
    Poništi
  19. 21. sij

    Working on a new way to visualize and debug geospatial algorithms. Idea is to have a minimal animation grammar for the drawing and manipulation of features, so you can write a stream of actions to a file and create a sharable html bundle.

    Prikaži ovu nit
    Poništi
  20. 21. sij

    The bike_id rotation patch to GBFS has passed by unanimous vote, meaning realtime tracking of riders by reconstructing trips will no longer be possible. Big thanks to everyone who contributed to the discussion and provided valuable feedback.

    Poništi

Čini se da učitavanje traje već neko vrijeme.

Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.

    Možda bi vam se svidjelo i ovo:

    ·