@StatsSimon

|| Biostat prof @ UW | interests in pedogogy, critical theory, chess (pretty bad at it) | he/him ||

Joined May 2019
  • Tweets180
  • Following141
  • Followers322
4Media
What's everyone's favorite package/approach to engage with tensorflow/keras in R? (trying to teach a short course, and don't want to require ppl to know python; unfortunately reticulate->miniconda seems to have a bug in windows =[)
1
2
4
I got my 3rd quarter performance review today...
14
Just came across a nugget of wisdom from a mathematician that I thought I would share: "A drunk man will find his way home, but a drunk bird may get lost forever" - Shizuo Kakutani
1
2
By far, the most accurate representation of machine learning pipelines in the real world 🍿 😀
205
1,549
419
10,942
She may look cute, but her favorite activity is eye gouging. no joke. (to be fair, it was a rough week for everyone...)
1
18
Wanted to share a lovely paper primarily by @StatTZhang that was just accepted at Annals of Statistics on efficient non-parametric regression using stochastic optimization! arxiv.org/abs/2104.00846 Less technical summary to follow... (1/n)

A Sieve Stochastic Gradient Descent Estimator for Online...

The goal of regression is to recover an unknown underlying function that best links a set of predictors to an outcome from noisy observations. In nonparametric regression, one assumes that the...

arxiv.org
1
2
16
more replies
Some neat points: Polynomial terms do not need to be orthogonal wrt our data! In fact this works super generally for regression functions in RKHS/Sobolev Ellipsoids. Even if we don't care about the online setting, this is a very computationally efficient approach. (7/n)
1
1
What does this say about deep reinforcement learning? Perhaps we can consider starting with a simple network and scheduling additions to the topology as we go --- though possible that we want to decrease step size for those additions. (8/n) Sorry this thread was so long!
1
2
She's no longer allowed to play with the neighborhood squirrels...
26
Why aren't there any good golden snitch costumes for infants? Also, how dare you suggest I plan to dress my baby up as the golden snitch while I watch the secrets of dumbledore... (when it is streamable)
3
Using familiar terms, you conflate causality and correlation in this offhand tweet. CRT does the hard work of identifying causal structures that lead to systemic inequities… you compare it with a non-rigorous opinion piece that treat root causes as personal opinions.
Judea Pearl
@yudapearl28 Apr 2022
We must remember that the holocaust was preceded by a successful program of "critical race theory". In April, 1938, Nobel Prize winner J. Stark explained in the British weekly "Nature" why Jewish professors at German universities must be replaced by "under-represented" Germans.
5
1
1
35
Have you ever had an SVM substantially outperform a reasonably tuned logistic regression?
abhishek
@abhi1thakur4 Apr 2022
When was the last time you used SVM?🤔 What was the performance like?
2
This is quite useful (especially for people looking to get into ML); my summary of the thread is that the ability to engage in simple, fit-for-purpose modeling, and then validate+communicate your models trumps all else.
Mark Tenenholtz
@marktenenholtz24 Mar 2022
Universities do a terrible job teaching machine learning. Not only do they give you critically out-of-date information, but they focus most of their time on the least important aspects. Here 5 things everyone in industry WISHES your professor taught you:
1
3
I think something intimated, but unwritten, is the value in understanding the FEW situations where a very simple modeling technique will NOT suffice. Most of the time, at the modeling stage, glm(...) or glmnet(...), or xgboost(...) is totally fine.
1
3
awesome work by Jean on how to get the most out of neural networks (with less work) for high dimensional biomedical problems!
Jean Feng@Jean_J_Feng15 Mar 2022
Our paper came out on Pi day! We describe how to train and tune neural networks for high-dimensional data using only TWO hyper-parameters (doi.org/10.1002/sam.11579). We think it's easier, so we've called it EASIER-net. @StatsSimon

Ensembled sparse‐input hierarchical networks for high‐dimensional datasets

In high-dimensional datasets where the number of covariates far exceeds the number of observations, the most popular prediction methods make strong modeling assumptions. Unfortunately, these methods...

onlinelibrary.wiley.com
3
Load more