enjoying this awesome tutorial on conformal prediction by @ml_angelopoulos & @stats_stephen: https://t.co/TBuVvn27yC. what a nice, straightforward framework to think of UQ in practical terms.
— Kyunghyun Cho (@kchonyc) August 10, 2022
there’s an associated tutorial paper as well: https://t.co/76PNfSIYlJ.
Just saw this comprehensive thesis on deep learning for OOD data drop on arXiv: "Towards Robust Deep Learning using Entropic Losses" https://t.co/MKXq6hbwWZ.
— Sebastian Raschka (@rasbt) August 10, 2022
Lots to read here, but in a nutshell, it looks like the proposed IsoMax+ performs well and is relatively easy to use 👌 pic.twitter.com/aF3Kpyd6QN
Atlas: a not-so-large language model (11B parameters) that beats the big guys at question answering and fact checking.
— Yann LeCun (@ylecun) August 8, 2022
The main difference is that it can retrieve facts from a corpus.
Paper: https://t.co/WlV3rbuY30 https://t.co/7bsNXfqV6l
A very good paper I came across this morning by the @DeepMind researchers. For the past five years Transformers have been one of the most dominant approaches to Deep Learning problems, especially in the #NLP domain.
— Bojan Tunguz (@tunguz) August 5, 2022
1/5 pic.twitter.com/XRQODHdQn3
Masked Vision and Language Modeling for Multi-modal Representation Learning
— AK (@_akhaliq) August 4, 2022
abs: https://t.co/zpOExcblUH pic.twitter.com/siunmulnng
1. On measuring causal contributions via do-interventions https://t.co/SjJVL5hCw0
— Amit Sharma (@amt_shrma) July 29, 2022
Shapley value, but each possible change in input is a do-intervention. Simple, principled idea for attribution. By @yonghanjung @patrickbloebaum @eliasbareinboim et al.
Check out a new hybrid approach that leverages both ML and rule-based semantic engines to build a Transformer-based semantic code completion model, which we show can improve developer efficiency. Read more ↓ https://t.co/bu6dTdPgII
— Google AI (@GoogleAI) July 26, 2022
[5/6] For large(r) datasets (50k), however, the gap between tree-based ML and deep learning is much smaller.
— Sebastian Raschka (@rasbt) July 23, 2022
However, the authors argue that large tabular datasets are rare. Do you agree? (I think it is somewhat true in academic collaboration but in industry?) pic.twitter.com/zNqdPkfzHq
And the deep learning vs conventional machine learning for tabular data continues!
— Sebastian Raschka (@rasbt) July 23, 2022
A new paper looks at 45 mid-sized datasets (10k examples) and finds that tree-based models (XGBoost & random forests) still outperform deep neural networks on tabular datasets. [1/6] pic.twitter.com/Xp1BDSHNtk
Contrary to what grad school wants you to believe, staying up late to work on your ML projects is actually bad (for the environment).
— Sebastian Raschka (@rasbt) July 22, 2022
Training a model at midnight increases CO2 emissions by 7.7% compared to 6:00 am. https://t.co/W1b3Vc9JxR pic.twitter.com/F2bWpX2tSe
Time and object detectors are running by so fast!
— Sebastian Raschka (@rasbt) July 20, 2022
Feels like I was using YOLOv3 just a few years ago. This month, YOLOv7 was released!
It's 1500% faster than a SWIN Transformer-based R-CNN (at the same accuracy) and 150% faster than YOLOv5!https://t.co/fdHnzNPj2n pic.twitter.com/byeT9ftAbT
We investigate what features of tabular data explain the difference, for this, we modify tabular data to narrow the gap.
— Gael Varoquaux (@GaelVaroquaux) July 19, 2022
Smoothing the outcome in feature space narrows the gap: deep architectures struggle with irregular patterns. Tree models do not care about smoothness
6/9 pic.twitter.com/NlyvuzwlgD
