Thrilled to announce our company, 🚀
We are in an exciting era of human-computer collaboration evolving the way we will reason with, process and generate information.
At Essential AI, we are passionate on advancing capabilities in planning, reasoning,
New Paper:
Stand-Alone Self-Attention in Vision Models
Can attention work as a stand-alone primitive for vision models?
We develop a pure self-attention model by replacing the spatial convolutions in a ResNet by a simple, local self-attention layer.
Life update: For those who haven’t heard, I left Google Brain!
I’m grateful for the 6+ years I spent there, the peers and friends that are inspiring and the opportunities to push on some of the most important problems in AI.
I’m excited to announce our new startup Adept with the mission to build useful general intelligence. We are a research and product lab that is enabling humans and computers to work together collaboratively.
Hello, world! Adept is a new AI research and product lab that aims to build useful general intelligence. What does that mean? Read on for a short introduction, or see our full launch announcement here:
Our new paper “Image Transformer”, extends self-attention from the original Transformer to much longer sequences on Image Generation and Super-Resolution tasks. It beats previous SOTA autoregressive models like PixelCNN, PixelRNN.
Our new paper “Image Transformer”, extends self-attention from the original Transformer to much longer sequences on Image Generation and Super-Resolution tasks. It beats previous SOTA autoregressive models like PixelCNN, PixelRNN.
New Paper:
Stand-Alone Self-Attention in Vision Models
Can attention work as a stand-alone primitive for vision models?
We develop a pure self-attention model by replacing the spatial convolutions in a ResNet by a simple, local self-attention layer.
1/7 We built a new model! It’s called Action Transformer (ACT-1) and we taught it to use a bunch of software tools. In this first video, the user simply types a high-level request and ACT-1 does the rest. Read on to see more examples ⬇️
Check out an early preview of the models we are building at
The future of interaction with computers is going to be massively redefined, starting with a few tools to anything you do with them. Come shape this future at
@AdeptAILabs
- we are hiring.
We made a fun video of some of the earliest things our system can do! If you want to help us build useful general intelligence, please reach out -- we are hiring.
Check out this paper on how de-noising decoders are useful for downstream vision tasks like segmentation. I'm excited about the future of a single, unified model across modalities and tasks.
Announcing decoder denoising pretraining for semantic segmentation:
Take a U-Net, pretrain the encoder on classification, pretrain the decoder on denoising, and fine-tune on semantic segmentation. This achieves a new SoTA on label efficient segmentation.
We presented our work on Image Transformer today at ICML’2018. We show self attention based techniques for Image Generation. Great experience with super colleagues
@ashVaswani
and
@dustinvtran
and others :)
We develop a non-autoregressive machine translation model whose accuracy almost matches a strong greedy autoregressive baseline Transformer, while being 3.3 times faster at inference. Joint work with
@ashVaswani
@nikiparmar09
Aurko Roy
Mesh TensorFlow: A sneak peak of what we've been working on with Noam Shazeer,
@topocheng
and others, to make model parallelism easier to scale to really large models. More details to come soon..
Pre-training on large amounts of text using a Transformer and then performing similar or better on a whole range of NLP tasks using little fine tuning! Impressive results , showing again that increasing compute to train bigger pre-trained models are helpful across various tasks.
5/ A most notable feature of Phenaki is its ability to generate long videos with *prompts changing over time*. These can be thought of as stories, where the user narrates and creates dynamically changing scenes. Here's a 2 minute (!) video from :
Check out the Image Transformer by
@nikiparmar09
@ashVaswani
others and me. Talk at 3:20p @ Victoria (Deep Learning). Visit our poster at 6:15-9:00p @ Hall B
#217
!
Super excited to be on this journey with
@ashVaswani
along with our incredible team
@mrinal_iyer_01
,
@ag_i_2211
,
@samcwl
,
@andrewhojel
and Varun Desai.
We believe that a small, focused team of motivated individuals can create outsized breakthroughs. If you want to work on some
Check out HaloNets, local self-attention that is efficient and gets strong results on ImageNet!
*Faster runtimes
*Parameter efficient
*85.6% Top-1 Accuracy
(1/5) In our recent CVPR paper, we develop a new family of parameter-efficient local self-attention models, HaloNets, that outperform EfficientNet in the parameter-accuracy tradeoff on ImageNet. .
I am at
#NeurIPS18
in Montreal. PM me if you’d like to talk about Generative Models, Model Parallelism or anything else!
We will be presenting our poster on Mesh TensorFlow along with
@topocheng
on Dec 4th, Poster Session A,
#136
.
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)
Further studies show that self-attention is the most useful in later layers while convolutions better capture lower-level features. Combining their will be an interesting research direction.
Overall my
#NIPS2017
experience has been amazing for a first timer. I leave with lots of inspiration and good connections and some non-partying for the near future.
#NIPS2017
Amazing results by OpenAI on using the GPT Language Model across so many tasks!
It's surprising to see how far, simple unsup LM models with scale, can take us :)
We've trained an unsupervised language model that can generate coherent paragraphs and perform rudimentary reading comprehension, machine translation, question answering, and summarization — all without task-specific training:
Enter
@AdeptAILabs
' and its brilliant co-founders
@jluan
and
@nikiparmar09
and
@ashVaswani
who building machines to work together with people in the driver's seat: discovering new solutions, enabling more informed decisions, and giving us more time for the work we love.
Our human evaluation result is 3x better than Pixel Super Resolution. We beat ppl by a significant margin on ImageNet and match PixelCNN++ on Cifar10 image generation.
The neural networks we build are designed for the hardware we have, and the hardware we build is designed for the neural networks we have. Help break the wheel — build networks that influence the design of future hardware.
A nice collaboration between
@GoogleAI
researchers and engineers and
@MayoClinic
to use Mesh TensorFlow to deal with very high resolution images that show up when doing machine learning on some kinds of medical imaging data.
Effective Machine Learning With Cloud TPUs. Tune in to find out more about the different tasks you can run on TPUs. Bonus: Performance numbers of various models in terms of time+cost including the Transformer and Image Transformer!!
The Universe in your hand 🔭💫
Learn how researchers from
@UCBerkeley
and
@NERSC
are using a new simulation code, FlowPM, to create fast numerical simulations of the Universe in TensorFlow.
Read the blog →
Yoshua Bengio, Geoffrey Hinton and Yann LeCun, the fathers of
#DeepLearning
, receive the 2018
#ACMTuringAward
for conceptual and engineering breakthroughs that have made deep neural networks a critical component of computing today.
@jekbradbury
The ImageNet experiments are also on 32x32 resolution. The current full autoregressive nature prevents us from going to bigger images. But we want to expand this to bigger images and stronger conditioning in the future
This model outperforms the baseline on ImageNet classification with fewer parameters and FLOPS. On COCO object detection, it matches the mAP of a baseline RetinaNet while having 39% fewer FLOPS and 34% fewer parameters.
Stephen Hawking once said, "I'm not afraid of death, but I'm in no hurry to die. I have so much I want to do first." RIP Professor Hawking, and may we all strive to live as fully as he did.
@MortezDehghani
@USC
@GoogleBrain
I wish I could :) really packed on time, spent the day at ISI. Next time around, I’ll plan well in advance and inform you. Apologies :)
Today we’re sharing our AI principles and practices. How AI is developed and used will have a significant impact on society for many years to come. We feel a deep responsibility to get this right.
Reviews on
#iPhonex
after 2 weeks of use
Good: amazing camera, colors and sharpness brilliant, good battery life, Face ID unlock works mostly
Bad: charging slow, feels a bit heavy, Face ID struggles in low battery mode, when using it in bed and when phone is a bit further away