While there are many attempts to map Auto ML to deep learning, to automate the tedious hyperparam tuning, here's a smart alt. via good initialization schemes: "Dynamical Isometry and a Mean Field Theory of CNNs: How to Train 10,000-Layer Vanilla ConvNets" https://t.co/zt9o89uP0K
— Sebastian Raschka (@rasbt) June 15, 2018
Autoregressive Quantile Networks for Generative Modeling: https://t.co/MImdcFiYxj
— DeepMind (@DeepMindAI) June 15, 2018
Text to speech with few data. https://t.co/JiX6TglXAw
— Nando de Freitas (@NandoDF) June 15, 2018
Differentiable Compositional Kernel Learning for Gaussian Processes #ICML 2018
— ML Review (@ml_review) June 15, 2018
By @ssydasheng @Guodzhhttps://t.co/aBisxDTp2h pic.twitter.com/AhPa71rRhm
Improving Consistency-Based Semi-Supervised Learning with Weight Averaging
— Andrew Gordon Wilson (@andrewgwils) June 15, 2018
Our new paper (+code!): https://t.co/5dRLFqgJlh
By analyzing loss geometry, we achieve record semi-supervised results, including 95% accuracy on CIFAR-10 with only 4000 labels! pic.twitter.com/EfqDhr3qC3
I want to flag this fascinating arxiv (by friends incl. @andrea_e_martin), showing a neural net architecture that can both learn _and_ generalise. If you follow @GaryMarcus then you'll know that deep nets don't generalise, so this seems a real advance. https://t.co/aZvZWARWQF
— Hugh Rabagliati (@hugh_rab) June 14, 2018
'A Probabilistic U-net for Segmentation of Ambiguous Images': a cool new paper by my colleagues at DeepMind on how to deal with uncertainty in segmentation models. https://t.co/PIVQ1CAOuh #DeepLearning #MachineLearning pic.twitter.com/6JJr68mZzD
— Nenad Tomasev (@weballergy) June 14, 2018
Playing around with cutout (https://t.co/qI4FtcWJlh), which, mixed with a bit of super-convergence, gets us to training cifar10 to 95% accuracy in just short of 30 minutes on a single GPU. pic.twitter.com/Rb2Sr0cmL0
— Sylvain Gugger (@GuggerSylvain) June 14, 2018
"Fast and Scalable Bayesian Deep Learning by Weight-Perturbation in Adam," Khan et al.: https://t.co/cOOinpU3Fo
— Miles Brundage (@Miles_Brundage) June 14, 2018
This is incredible. "Neural Best-Buddies: Sparse Cross-Domain Correspondence" https://t.co/RvZmflafAJ https://t.co/FAbRzwRao0
— hardmaru (@hardmaru) June 13, 2018
I wanted to finish talking to the journal, but the media articles keep coming and even @ylecun is repeating the claims... time for some #openpeerreview!
— Luke Oakden-Rayner (@DrLukeOR) June 13, 2018
This paper is flawed. Their comparison underestimates human performance systematically. See whole thread for explanation: https://t.co/RaHWho2bY9
Meta-learning enables fast learning, but needs hand-engineered meta-training tasks. Can we get the tasks themselves automatically? Our first attempt at this for RL: unsupervised meta-reinforcement learning:https://t.co/ePkIsMcMjZ
— Sergey Levine (@svlevine) June 13, 2018
w/ Abhishek Gupta, Ben Eysebach, @chelseabfinn
