I'm thrilled to announce our company,
@essential_ai
. We believe that breakthroughs in AI will unlock the most profound tools for thought, advancing humanity's collective knowledge and capability.
After 5+ wonderful years in Google Brain, working at the forefront of ML alongside inspiring colleagues, I'm excited to share my new adventure. We started Adept with the mission to build the future of human-computer collaboration. .
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 ⬇️
We’re building the future of useful and intelligent machines. Our early neural networks can make plots, query databases and fetch internet data! If you’re excited to work on fundamental research like learning from human feedback and building sample efficient models, please apply!
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.
(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. .
Essential AI is dedicated to building technology that lowers the barrier to any enterprise workflow, time consuming or complex, that can be performed on a computer.
We believe that a small, focused team of motivated individuals can create outsized breakthroughs. If you want to work on some of the most important problems in AI, please apply here:
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.
(3 / 5) In previous work (), we used pixel-centered windows, similar to convolutions. Here, we develop a block centered formulation for better efficiency on matrix accelerators.
(4/5) When applied to detection and instance segmentation, our local self-attention improves on top of strong convolutional baselines. Interestingly, local self-attention with 14x14 receptive fields performs nearly as well as 35x35.
(2/5) In addition to strong results on ImageNet, we also see promising improvements (up to 4.4x inference speedups) over strong baselines when pretrained on ImageNet-21k with comparable settings.
Our self-attention model outperforms the Resnet baseline on ImageNet classification and matches RetinaNet on object detection with fewer FLOPS and parameters.
@jekbradbury
@arvind_io
@nikiparmar09
Thanks! We haven't inspected the latents as yet but that is probably the most exciting thing to do at this point. It's definitely on our TODO.