As promised, we're continuing to push our research 🔬 into
#MusicFX
DJ to make the experience more engaging and controllable--putting the human deeper in the infinite loop.
First up, we've increased the responsiveness and musical diversity, and added a
We just launched Magenta Studio: a suite of free music-making tools using our machine learning models! They're available as an
@ableton
plugin or as standalone Electron apps. So excited to see what you make with them!!
#MadeWithMagenta
👉 Download them at
We are releasing mT5: A massively-multilingual version of T5 that supports over 💯 languages! mT5 was pre-trained on a multilingual version of C4 and achieves SoTA on many cross-lingual NLP tasks.
📜Pre-print:
💾Code/models:
Check out our
@GoogleAI
blog post discussing the T5 paper from October and some fun follow-up work.
Not only did T5 redefine SoTA for many NLP tasks 📈, it can also tell you how to make a peanut butter banana sandwich 🥪 even Elvis would be proud of.
Introducing the MAESTRO dataset (172 hours of aligned MIDI and audio piano performances) and Wave2Midi2Wave: a factorized approach to modeling and generating piano performances.
w/
@fjord41
Andriy Stasyuk
@iansimon
@huangcza
@sedielem
Read 👉
Listen 👇
As promised, we have made the Text-To-Text Transfer Transformer (T5) models much easier to fine-tune for new tasks, and we just released a Colab notebook where you can try it yourself on a free TPU!
👇
(1/3)
We just updated Magenta Studio for
@Ableton
!! 🎉
The big change is a 5th plugin called Drumify which generates 🥁 grooves to accompany melodies, chord progressions, & basslines using
#ML
.
Check it out in action 🎬👇
Learn more and try it yourself at !
Lots of exciting work to share today about
#MusicVAE
on the
#Magenta
blog: a full description, a JavaScript library built on
@deeplearnjs
, 3 browser-based apps to try it yourself, and examples galore!
Excited to share my newest work
#MusicVAE
for interpolating and sampling melodies, beats, and three-part song segments from a VAE! Listen to samples and create your own in the
#colab
notebook (link in YT description)
w/
@jesseengel
@deck
#magenta
#nips2017
For a year, the T5 team has collab'd with FLAX and JAX to build a successor to our research library, using it to train models at many scales 📈 on TPU...
...and now you can too!
T5X is still in rapid development, but you can use it or find inspiration at !
Last summer,
@GoogleMagenta
intern
@jon_gillick
recorded MIDI of professional drummers to train MusicVAEs (GrooVAEs🥁). This awesome work is core Magenta Studio and was accepted to
#icml2019
🎉
Today we release the dataset, code and colab on our blog: 👀
New preprint: How Much Knowledge Can You Pack into the Parameters of a Language Model?
We show that T5 outperforms all previous open-domain QA systems *without using any external knowledge or context*.
Joint work w/
@colinraffel
& Noam Shazeer.
(1/5)
After a 2 year hiatus from music to work on LLMs, I'm psyched to be back at it with
#MusicLM
and
#SingSong
.
#SingSong
has been blowing my mind for weeks and I'm ecstatic to finally be able to share it with the world.
Looking for ways to bring it closer to you...
Excited to share SingSong, a system which can generate instrumental accompaniments to pair with input vocals!
📄
🔊
Work co-led by myself,
@antoine_caillon
, and
@ada_rob
as part of
@GoogleMagenta
and the broader MusicLM project 🧵
Here is a real-world 🚲 example (not in the paper) for T5 Small (~60M params). The VeLO-trained model reaches the same loss as >5x steps of Adafactor and takes only ~1.5x as long per step in wall time 🕓...a >3x speedup⏫!!
Pretty amazing for such an OOD task!
UPDATE: We have spent the past month “fine-tuning” our approach for Closed Book QA (CBQA, no access to external knowledge) w/ T5 and now our appendix is overflowing with interesting results and new SoTAs on open domain WebQuestions and TriviaQA!
(1/7)
New preprint: How Much Knowledge Can You Pack into the Parameters of a Language Model?
We show that T5 outperforms all previous open-domain QA systems *without using any external knowledge or context*.
Joint work w/
@colinraffel
& Noam Shazeer.
(1/5)
I've found myself referring to this table in the appendix of our recent paper () A LOT so I thought I'd point it out for those who might benefit from it. Basic upshot is that ED:MLM works really well, especially for classification tasks, even in 0shot.
I'm really proud of the work our team continues to do to make generative AI a transformative tool for human creativity!
*Music AI Tools* will bring unprecedented creative capability and control to more people, developed responsibly with artists throughout the process.
Thrilled to share
#Lyria
, the world's most sophisticated AI music generation system. From just a text prompt Lyria produces compelling music & vocals. Also: building new Music AI tools for artists to amplify creativity in partnership w/YT & music industry
This is paving the way for a very dangerous practice by media organizations. DALL-E mini is pretty obvious, but once models at the level of
#dalle2
and
#imagen
are widely available, we are in trouble.
Very excited about this new work demonstrating the benefits of multitask finetuning / instruction tuning!
Also happy to continue the trend of open-sourcing our T5 model variants via
#t5x
for everyone to use and build upon (even if you use
@huggingface
:) -- more to come!
New paper + models!
We extend instruction finetuning by
1. scaling to 540B model
2. scaling to 1.8K finetuning tasks
3. finetuning on chain-of-thought (CoT) data
With these, our Flan-PaLM model achieves a new SoTA of 75.2% on MMLU.
Speaking totally from my own perspective, while I'm technically impressed with
@OpenAI
's jukebox, I have many concerns from an ethical perspective. Why was the GPT-2 launch treated with such sensitivity but not this?
I'm so excited about the program we've put together for Saturday's
#NeurIPS2019
ML for Creativity and Design Workshop 3.0.
Aside from the amazing accepted talks and posters, we have a diverse set of invited speakers I want to highlight in this thread.
After 2 years focused on infra, it was a joy to dive back into research with
@colinraffel
and the friendly folks at
@BigscienceW
while taking OSS T5X for a spin.
Our systematic study fills some gaps in our knowledge of LLM design for zero-shot. Hope you find it useful!
🧐 When targeting zero-shot use, should you train a T5, a PrefixLM, or a GPT? What if you plan to leverage multitask finetuning (à la T0)?
🤩 In , we explore how architectures & pretraining objectives impact zero-shot performance.
⬇️ Thread time!
We used T5X and SeqIO to train PaLM 🌴, and also released a short paper on this infrastructure alongside the model. Check it out at .
Huge thanks to the many contributors and my co-lead authors
@hwchung27
@anselmlevskaya
@mishra_gaurav_
@jekbradbury
!
Introducing the 540 billion parameter Pathways Language Model. Trained on two Cloud
#TPU
v4 pods, it achieves state-of-the-art performance on benchmarks and shows exciting capabilities like mathematical reasoning, code writing, and even explaining jokes.
Exciting new work!
Biggest takeaway IMHO is that you should switch to finetuning the new👶Flan-T5 checkpoints over the OG🧓T5 ones for best results.
Congrats to
@ShayneRedford
on leading this project and helping us continue the tradition of opensourcing our T5 work!
✨New Paper✨What’s the best completely public competitor to
#ChatGPT
?
Flan-T5 beats all public models we tested:
Flan-T5 3B ▶️ T0++ 3B ▶️ OPT-IML 175B ▶️ GLM-130B ▶️ Flan 2021 3B ▶️ NIv2 3B
We release the
@GoogleAI
🌟Flan Collection🌟data + methods for Instruction Tuning!
1/
"It’s tempting to analogize implementations to scientific experiments, but usually they aren’t. It’s tempting to think of them as engineering solutions, but they usually aren’t. I think “implementations” are best understood as design solutions..."
👆👇
SOOOOO excited this album is finally out!! Musical parts were composed using MusicVAE and NSynth bleeps all over it!
A truly groundbreaking human/AI collaboration involving many amazing artists, researchers, and ML models!
Check out the updated MusicVAE paper for
#icml2018
featuring new attribute vector results w/ Hinton diagrams courtesy of
@colinraffel
and an example page thanks to
@fjord41
!
paper:
examples:
Very happy that the Creativity Workshop was accepted for a third year at NeurIPS, but even happier to see this wonderful breakdown of the selection process by the selection committee!
While things may be heating up in the competitive space of AI and LLMs, it's important that we share resources and artifacts for the benefit of science and progress!
Also important to have talented students like
@ShayneRedford
remind us of this fact.
Today we’re releasing a new collection of tasks, templates and methods for instruction tuning of
#ML
models. Training on this collection can enable language models to reason more competently over arbitrary, unseen tasks. Learn all about it at:
We just updated Magenta Studio for
@Ableton
!! 🎉
The big change is a 5th plugin called Drumify which generates 🥁 grooves to accompany melodies, chord progressions, & basslines using
#ML
.
Check it out in action 🎬👇
Learn more and try it yourself at !
@yacht
is making some truly groundbreaking art with collaborators across the world of machine learning and art as
@TheUniverse
discussed at
#io19
yesterday (). Lots more exciting stuff to come from them so stay tuned!!!
Kudos to
@OpenAI
for making this type of generation impossible with their model and API. That will slow things down a bit...
In the meantime it is important to solidify journalistic standards on this and distribution mechanisms (like Yahoo and Google News) to enforce them.
I recently gave a talk at
@GrayAreaorg
for the
#AlgorithmicArtAssembly
covering the rapid progress in generative models in recent years and how
@GoogleMagenta
is exploring their use as creative tools.
...and it's now available on YouTube! Take a 👀:
If you're at
#abletonLoop
today, come hear
@jesseengel
and I talk about how you can use now use Magenta's machine learning models as part of your
@ableton
workflow at 11:45!
Just added much improved support for exporting SavedModels to the T5 codebase. This makes it much easier and faster to do inference on your fine-tuned model!
Check out the new example at the bottom of the colab:
A few of you at
#ismir2019
asked me what I've been up to. Well, I took a sorta-but-not-really-hiatus from Magenta to work on NLP!
The result was T5, which has been a very rewarding experience. Now, I'm looking forward to bringing some new insights back to music generation.
New paper! We perform a systematic study of transfer learning for NLP using a unified text-to-text model, then push the limits to achieve SoTA on GLUE, SuperGLUE, CNN/DM, and SQuAD.
Paper:
Code/models/data/etc:
Summary ⬇️ (1/14)
Check out our new blog post on music transformer that generates minute-long pieces that I find myself listening to again! It also generates compelling harmonizations and realizes chord progressions.
Which is your favorite? Ours is the Jazz 🎷
👂
13 years ago yesterday I asked awrness to marry me and 13 years ago today we went to the courthouse in durhamnc to make it official! A few weeks later we moved to the Bay Area and have been there ever since. Today,…
Also, I am 1000% hiring PhD students this round! If you want to work on
- open models
- collaborative/decentralized training
- building models like OSS
- coordinating model ecosystems
- mitigating risks
you should definitely apply! Deadline is Friday 😬
First year CS students should take a "history of computing" course and
@TheUniverse
's "Broad Band" should be required reading.
I deputize
@justinesherry
to help make this a reality :)
Very excited to have been a small part of this amazing project.
More work to be done to make this optimizer the go-to for giant (1B+) models, but it's in sight!
Tired of tuning your neural network optimizer? Wish there was an optimizer that just worked? We’re excited to release VeLO 🚲, the first hyperparameter-free learned optimizer that outperforms hand-designed optimizers on real-world problems:
🧵
Algorithmic Art Assembly is a brand new conference and music festival, focused on algorithmic tools and processes. Join us for two days of talks and two nights of live music featuring artists using algorithms for aesthetics.
March 22 – 23rd
Tickets here:
Looking forward to
@TheUniverse
(of
@YACHT
) speaking on the band's recent work involving machine learning and participating in a discussion with her,
@jesseengel
, and another (surprise!) artist on Wednesday. You can watch live on YouTube!
(2) Claire Evans AKA
@YACHT
talks about deep learning as a tool in the creative process of YACHT's new album – with
@Google
Magenta's
@ada_rob
and
@jesseengel
Tune in live 5/8 12:30pm PT or visit us in-person on Stage 4. Learn more ↓
So happy that the Magenta.js piano transcription found its way into one of these accessibility projects thanks to
@yotammann
. Really amazing and inspiring work by everyone involved!
"Unmixer is a web interface where users can upload music, extract loops, remix them, and mash-up loops from different songs. To extract loops with source separation, we use a nonnegative tensor factorization method improved with a sparsity constraint."
One great thing about having a kid is you occasionally rediscover amazing artifacts like a
@philipglass
scored animation about the geometry of circles from
@sesamestreet
circa 79
The EfficientQA
@NeurIPSConf
Competition () is offering GCP grants and office hours to help expand participation amongst underrepresented groups.
Please apply (even if you're new to QA) at by July 30 to be considered for a grant!
In concurrent work (also posted to arXiv today) our friends at Google Research released the incredibly cool REALM, an open-book QA system that is trained end-to-end to retrieve documents and answer questions. They outperform us on NQ and WQ.
To see how to train T5 on a simplified variant of this problem with a free TPU, check out our colab notebook:
Questions, comments, and feedback welcome!
(5/5)
I think my biggest concern is conditioning on specific artists without their permission, which is something that has rubbed me the wrong way in past work as well. It seems even more problematic now that it's this good.
I just reheard this very impressive WaveNet-based singing synthesis from Merlijn Blaauw and
@JordiBonada
during
@emiliagogu
's excellent keynote at the
#ICML2018
music workshop, and I was once again blown away. Take a listen if you haven't already.
We built a really cool machine learning drummer that accompanies your piano in real-time!!
This is DrumBot! It’s a web app, so it can come along to your next gig. 🎉
🥁
📝 Blog post:
🎧 Live action:
These models have some journalistic value in making news article illustrations, but not photorealistic images that could be mistaken for photographic evidence.
@sarahookr
was the first AI Resident I mentored, but I am confident in saying that over the years I've learned much more from her than she ever did from me!
I'm very excited to see what incredible things For AI will do under her leadership!
I'm excited to finally share what I have been working on.
Today we are officially launching Cohere For AI
@forai_ml
a non-profit research lab that aims to reimagine how, where, and by whom research is done.
It may not be as flashy, but I am going to continue focusing my work in the music space on interactive tools to enhance creativity for professionals and amateurs alike!
Morning
#ICML2018
. Come check out the awesome set of talks at the workshop on Machine Learning for Music!
I'll be presenting the recently launched Magenta.js library and talking about how it is engaging artists, musicians, and creative coders.
Announcing new project: GANHarp
This is an experimental musical instrument based on Magenta's GANSynth and Magenta.js.
It synthesizes timbres from latent space, and then lets you morph between them as you play.
#MadeWithMagenta
🎹
Read/watch a bit more about
@GoogleMagenta
's I/O collaboration with
@theflaminglips
on Fruit Genie and long-term work with
@YACHT
on their forthcoming album. Still lots more to come!
🔨🍓
Also, learn how to build your own Fruit Genie at !
Some fun work by our team in the Google Doodle today to celebrate Bach's birthday! Super proud of
@huangcza
who did a lot of this work as my intern 3 years ago (w/ very little help from me) and continues to do amazing things!!
Highly recommended that anyone interested in starting/continuing deep learning research (especially those from underrepresented groups and non-traditional backgrounds) APPLY!
Applications for the 2020 Google AI Residency program are now open! Visit for application information. To learn more about the research accomplishments of the 2019 alumni, check out the post below!