Late last year at #corl2022, I gave a talk highlighting some of our recent #waymo #research on behavior modeling topics.
If you are interested in behavior prediction (aka motion forecasting) and agent modeling, check it out!
youtube.com/watch?v=RpiN3L
1/2
Follow
Click to Follow dcower
Dillon Cower
@dcower
Software Engineer . Before: on WebRTC, Duo, spatial audio, ambisonics, VR media. 
: @dcower@mastodon.gamedev.place
: ratan.band
: @dcower@mastodon.gamedev.place
: ratan.bandDillon Cowerβs Tweets
Excited to share SingSong, a system which can generate instrumental accompaniments to pair with input vocals!
πarxiv.org/abs/2301.12662
πg.co/magenta/singso
Work co-led by myself, , and as part of and the broader MusicLM project π§΅
140
1,085
3,128
Show this thread
Following our recent expansion of 24x7 rider-only operations across all of SF and in downtown PHX, weβre forging ahead with advancing capabilities for future scope, e.g., freeways. Exciting road ahead! Hereβs a snippet from my recent ride from our South Bay HQ to the SF office:
14
56
317
Expanded our rider-only (no human driver) 24x7 operations to all of SF and doubled the downtown PHX area including airport. Proud and inspired by what the team accomplished this year, leveraging some mind-blowing AI/ML advances. More exciting milestones ahead⦠onward!
3
22
72
Today weβre proud to announce:
β‘οΈ Weβre the first AV company to launch a commercial service to an airport
β‘οΈ We just doubled our #DTPHX territory in a matter of weeks
β‘οΈ Weβre now rider-only 24/7 in all of #SF
Weβre leading the way forward. blog.waymo.com/2022/12/wheels
GIF
12
67
183
This clip is from one of my rider-only (fully autonomous) rides in San Francisco. Excited for SF residents to experience this. It's really remarkable how well it works. Sign up and try it for yourself!
Quote Tweet
Receiving the @californiapuc driverless pilot permit means we can now bring this rider-only experience to San Francisco residents. 24/7; no human driver. Download the Waymo One app and join the waitlist to ride with us! bit.ly/3G8EdhO
0:25
3.1K views
2
3
Show this thread
SF, whoβs ready to ride? ππ€
After receiving the driverless pilot permit from the , Waymo One is opening to members of the public in San Francisco. Available 24/7βwithout anyone in the driverβs seat: cpuc.ca.gov/news-and-updat
33
199
473
Show this thread
Seeing Beyond the Brain: Conditional Diffusion Model with Sparse Masked Modeling for Vision Decoding
abs: arxiv.org/abs/2211.06956
project page: mind-vis.github.io
github: github.com/zjc062/mind-vis
25
307
876
ZerO Initialization: Initializing Neural Networks with only Zeros and Ones
, Florian SchΓ€fer,
tl;dr: theoretical version of twitter.com/ducha_aiki/sta
Thanks to to sharing the paper.
openreview.net/forum?id=1AxQp
This Tweet is unavailable.
3
22
99
Now that this prediction has come to pass, I should explain why I thought this was a reasonable prediction and why a lot of other predictions about what might happen were or are implausible.
Some necessary background for this is the state of Twitter over the past few years: twitter.com/altluu/status/
This Tweet is unavailable.
10
73
298
Show this thread
Today, we present our paper on Google Search Ads CTR model at ORSUM , Seattle.
arxiv.org/abs/2209.05310
We highlight ML techniques suited to *online learning* that go well beyond traditional accuracy improvements.
orsum.inesctec.pt/orsum2022/prog
A short thread:
1/n
6
97
385
Show this thread
Over the next six months, those in the Bellevue, Washington area may see our fifth-generation #WaymoDriver as we conduct a variety of tests that will help us deepen our understanding of weather as we expand our Driverβs capabilities. π€ππ§οΈ
GIF
2
20
95
Diffusion for music synthesis!
We trained a βnotes2audioβ pipeline to synthesize audio from multi-instrument MIDI notes.
Listen π: g.co/magenta/spec-d
Play πΌ: g.co/magenta/spec-d
Code π©βπ»: g.co/magenta/spec-d
Read π : arxiv.org/abs/2206.05408
1/
read image description
ALT
33
637
2,727
Show this thread
Looking for #lofi tunes to help you end the workweek or commute home on the right note? π§πΆπ€π Weβve got just the thing for you (h/t Luke from our Product team):
4
22
56
the way to tell that you have matured as an engineer is to examine your reaction when you see an opportunity to come up with a design for a nice generic subsystem. if your first instinct is to start actively looking for ways to avoid building it - you're ready.
2
5
57
Show this thread
Safe and responsible autonomous driving, from Waymo. Built with TensorFlow & Keras (+many other technologies) twitter.com/Waymo/status/1
This Tweet is unavailable.
4
9
144
Introducing LocoProp, a new framework that reconceives a neural network as a modular composition of layersβeach of which is trained with its own weight regularizer, target output and loss functionβyielding both high performance and efficiency. Read more β goo.gle/3bdKOhA
GIF
read image description
ALT
16
221
973
Agreed! I feel like an underinvested area in ML is benchmark design. Progress is catalyzed by the existence of benchmarks and, I suspect, stagnates in their absence.
1
4
15
- Yes34.2%
- No65.8%
926 votesFinal results
6
20
34
Show this thread
Tired of finding and writing finite state machines by hand? Why not find them automatically using differentiable programming?
Really cool post: google-research.github.io/self-organisin and trivial numpy-like Jax code!
3
61
286
Show this thread
Weβre looking forward to participating in this year, both in person and online! Hereβs a preview of some of our sessions, including recent state-of-the-art work in autonomous driving research that weβll be presenting: blog.waymo.com/2022/06/%20Way
GIF
3
5
25
Show this thread
Ever wondered why deep learning is always done on array data?π€ Happy to announce our work:
From data to functa: Your data point is a function and you can treat it like one
πarxiv.org/abs/2201.12204 w/ , to appear in ICML22
GIF
14
91
519
Show this thread
βFinish the cat drawingβ viral meme tweet has replies with all sorts of nice, creative βout of the boxβ thinking.
I use #Dalleβs inpainting function to do this task, and was impressed at what it can do. Here is the output using the prompt βcatsβ
π§΅An entire thread of results π twitter.com/memesiwish/sta
This Tweet is unavailable.
8
370
1,394
Show this thread
Alright, fine: it's getting enough traction that I think I need to address this paper as a certified Grumpy Linguist in NLP.
(I generally try to avoid peer reviewing students in public but at this point it's definitely *already* very public, so π€·ββοΈ.)
Quote Tweet
DALLE-2 has a secret language.
"Apoploe vesrreaitais" means birds.
"Contarra ccetnxniams luryca tanniounons" means bugs or pests.
The prompt: "Apoploe vesrreaitais eating Contarra ccetnxniams luryca tanniounons" gives images of birds eating bugs.
A thread (1/n)
Show this thread
8
121
462
Show this thread
DALLE-2 has a secret language.
"Apoploe vesrreaitais" means birds.
"Contarra ccetnxniams luryca tanniounons" means bugs or pests.
The prompt: "Apoploe vesrreaitais eating Contarra ccetnxniams luryca tanniounons" gives images of birds eating bugs.
A thread (1/n)π§΅
208
3,846
9,326
Show this thread
Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding
project page: gweb-research-imagen.appspot.com
sota FID(7.27 on COCO), without ever training on COCO, human raters find Imagen samples to be on par with the COCO data itself in image-text alignment
30
916
2,683
Show this thread
How many languages can we support with Machine Translation? We train a translation model on 1000+ languages, using it to launch 24 new languages on Google Translate without any parallel data for these languages.arxiv.org/abs/2205.03983 Technical π§΅below: 1/18
7
113
344
Show this thread
We're excited to present a new method to render Signed Distance Functions (SDFs) in a differentiable manner, enabling high-fidelity image-based shape reconstruction. This is joint work with and and will be presented at SIGGRAPH'22. (1/8)
13
347
1,605
Show this thread
transformer inference performance is becoming increasingly important and there's not as much lore on it, so here is a lot of lore that i think fully models llm inference performance
carolchen.me/blog/transform
4
61
370
Show this thread
Weβre taking the next step in our journeyβboth in San Francisco and in Phoenix. Last week, our employees began taking fully autonomous rides in the City by the Bay, and soon weβll be expanding our Waymo One Trusted Tester program into Downtown Phoenix. blog.waymo.com/2022/03/taking
GIF
15
155
277
2.8 million images were used to build a grid of Block-NeRFs and create the largest neural scene representation to date, capable of rendering an entire neighborhood in San Francisco. Dive in to the latest research from Waymo and Google Research: waymo.com/research/block
21
284
1,110
𧡠Making a cpu using an analog modular synthesizer
I'd appreciate if you might share this, it was fun but writing it up was a lot of work :3
GIF
read image description
ALT
103
1,388
3,679
Show this thread
IMO the most important thing to work on for gamedev technology and services is how to reduce the cost of making games without needing to sacrifice production value. I canβt speak for Epicβs entire strategy but that is my personal drive. twitter.com/mikeBithell/st
This Tweet is unavailable.
6
43
358
Show this thread
I'm starting a (free) newsletter on financial infrastructure, with the first issue shipping Friday.
If you'd like to get it when it comes out:
41
100
763
Show this thread
The bar for singing voice synthesis research demos has been raised to "sing your paper's abstract" twitter.com/_akhaliq/statu
This Tweet is unavailable.
1
29
146
"Often wrong but never in doubt"
Sample Standard Deviation (SD) vs Standard Err (SE)
You want an estimate mΜ of m=πΌ(x) from N independent samples xα΅’. A typical choice is the average or "sample" mean
But how stable is this? The Standard Error (SE) tells how stable it is
1/6
3
31
206
Show this thread