"On Empirical Comparisons of Optimizers for Deep Learning" => "As tuning effort grows without bound, more general optimizers should never underperform the ones they can approximate" https://t.co/a76zy4kce3 pic.twitter.com/hUIGMyshkC
— Sebastian Raschka (@rasbt) October 16, 2019
"ExBert - A Visual Analysis Tool to Explore Learned Representations in Transformers Models" -- this looks like a nice one! Tool: https://t.co/YZqHvUBD5h and paper: https://t.co/NEJG4LKGzP pic.twitter.com/q96JWrWfD2
— Sebastian Raschka (@rasbt) October 15, 2019
We've trained an AI system to solve the Rubik's Cube with a human-like robot hand.
— OpenAI (@OpenAI) October 15, 2019
This is an unprecedented level of dexterity for a robot, and is hard even for humans to do.
The system trains in an imperfect simulation and quickly adapts to reality: https://t.co/O04izt3KvO pic.twitter.com/8lGhU2pPck
This new activation function has seen quite a bit of success in the @fastdotai community already, including generating a forum discussion with over 500 posts (including many from the author of the paper)! https://t.co/lL05utISsp https://t.co/6LU0CiCXrg
— Jeremy Howard (@jeremyphoward) October 14, 2019
Transferability is a major challenge in street navigation. Our work @ICCV19 proposed a cross-view policy learning approach that utilises top-down aerial view imagery to enable agents to learn faster and better in unseen street areas. https://t.co/wRgWdbyXeR pic.twitter.com/9sHHMj2X1r
— DeepMind (@DeepMindAI) October 14, 2019
Mish: Self Regularized Non-Monotonic Activation Function
— ML Review (@ml_review) October 13, 2019
By @DigantaMisra1
𝑓(𝑥)=𝑥⋅𝑡𝑎𝑛ℎ(𝑠𝑜𝑓𝑡𝑝𝑙𝑢𝑠(𝑥))
Increased accuracy over Swish/ReLU
Increased performance over Swish
Githubhttps://t.co/RYzuj0xhDN
ArXivhttps://t.co/YJKTd4yKvr pic.twitter.com/zlyQ0hwggt
NEW POST: My paper with @N8V_TXN on opioid prescribing rates in Texas is now out!
— Julia Silge (@juliasilge) October 12, 2019
I used #rstats tools like the tidyverse, tidycensus, and huxtable! ✨https://t.co/d14gWUu1lC pic.twitter.com/MMAKqFtt1z
Linking emotions to behaviors through deep transfer learning
— hardmaru 😷 (@hardmaru) October 11, 2019
Interesting work from psychology field that uses deep learning as a tool to help quantify the relationship between emotions and behaviors.https://t.co/yEYwmsNeAP pic.twitter.com/HkESKW8DMy
https://t.co/tbNxduCZNp is something was launched for researchers but it will be a game changer for applied AI. They run models inference on their GPUs so you get a trusted info about:
— Piotr Czapla (@PiotrCzapla) October 10, 2019
- Inference time on V100
- The performance vs paper
- How hard is to run the model pic.twitter.com/gzl99VMlHV
sotabench 🏅🛠—This is a great new service by @paperswithcode!
— Sebastian Ruder (@seb_ruder) October 10, 2019
It enables benchmarking open-source ML models and comparing the results to what was reported in the paper—directly from GitHub repos. Should be a step towards more replicability in ML!https://t.co/5diJXFTFxX pic.twitter.com/j1jn1Tdljt
Announcing TorchBeast, an IMPALA-inspired @pytorch platform for distributed RL research. Used in a growing number of projects here at @facebookai. Project lead by Heinrich Küttler, with major effort by @nntsn et al.
— Edward Grefenstette (@egrefen) October 9, 2019
Paper: https://t.co/HMYTtesSqC
Code: https://t.co/vadpUsyvaZ
This is a nice diagram by Zhengyan Zhang and @BakserWang that shows how many recent pretrained language models are connected. The GitHub repo contains a full list of relevant papers: https://t.co/uQNRGqMJAA pic.twitter.com/gJeJkUKAvR
— Sebastian Ruder (@seb_ruder) October 8, 2019
