Profile cover photo
Profile photo
Michael Tetelman
513 followers -
Life is a chain of accidents, but it is a most probable chain of accidents
Life is a chain of accidents, but it is a most probable chain of accidents

513 followers
About
Michael's posts

Post has shared content
This is really cool!
Some fun work on automatic model design trough evolution: you can evolve (last year's) state-of-the-art CIFAR models from scratch using 100 exaFLOPs in under 300 hours without baking any assumptions into the architecture search. Given that it took us years to get there through 'Graduate Student Descent', it's not a bad start!

Post has shared content
+Esteban Real, +Jon Shlens, Xin Pan, and +Vincent Vanhoucke in the Google Brain team and Stefano Mazzocchi in another team at Google Research just released a new public dataset called YouTube-BoundingBoxes, consisting of 5 million human annotated bounding boxes across 380,000 video segments.

Deep learning models for handling video rather than just static images are likely to be the next frontier for computer vision research, and this large dataset is likely to be an important new tool in assessing the effectiveness of a wide variety of video models in the areas of localization, detection, and object tracking.

There's a more detailed associated Arxiv paper that describe the dataset and the methodology used for collecting it in more detail: https://arxiv.org/pdf/1702.00824v1.pdf

Nice work, everyone!

Post has shared content
Creating Human-Level AI by Yoshua Bengio, Université de Montréal.

Post has shared content
Single-image depth prediction with FCRN:

Code and models are now also available in TensorFlow.

Post has shared content

Post has shared content
Call for Papers: 1st Conference on Robot Learning

Submission deadline: June 28, 2017
Paper acceptance notification: September 1, 2017
Conference dates: November 13-15, 2017

Post has shared content
Please check out our newly released dataset SceneNet RGB-D. This is work from the Dyson Robotics Lab at Imperial College, led by +John McCormac and +Ankur Handa in collaboration with +Stefan Leutenegger and myself. It's a photorealistic (ray-traced) dataset of 5 million indoor images split into 1fps sequences of 300 images from cameras following smooth, random trajectories. The large number of different layouts are filled with statistically sampled collections of objects in random but physically convincing positions. We provide many types of ground truth, including camera path, semantic segmentation, object instance segmentation, depth and optical flow. The dataset could be used for many things, but is primarily targeted at semantic scene understanding, where the scale and variety we include makes it feasible to train RGB-D segmentation networks from scratch. Let us know what you think it's useful for!

Post has shared content
This talk is so full of wisdom. Lots of great high level advice about how to build deep learning systems and run an ML organisation.

Post has shared content
Using videos to learn what objects are in an unsupervised way must be the way to go. I expect few-shot learning of object-label mapping to follow:
http://xxx.lanl.gov/pdf/1612.06370v1


Post has shared content
Quite interesting work on SLAM-enabled deep reinforcement learning from Oxford. See the paper at https://arxiv.org/abs/1612.00380
Photo
Wait while more posts are being loaded