wendykan

Mom, ML eng @google, previously at @kaggle. PhD in bio eng, BS/MS in electrical eng

[wendykan @wendykan]

https://t.co/I1K9PeCYAo 🤣🤣🤣

[wendykan @wendykan]

@__apf__ Love this! My 4.5 year old was my +1 at a party w all grownups, quietly played on his own or whoever wanted to entertain him, didn't complain for lack of attention. Super sweet kid and fun to talk to. Sometimes he's difficult but he's still better than all of my ex bfs ❤️❤️

[wendykan @wendykan]

This week the other two women announced they're leaving my team 😭😭😭 now I'm the only woman as TL leading a bunch of guys. What's worse, I cried in the team meeting after their announcement 🤦‍♀️. Please send help.

[wendykan @wendykan]

I made a simple app to generate kids bedtime stories using gpt-3.5 :) It was great fun! Thanks @chrisgorgo for the inspiration. https://t.co/DTY6LiWdT9

[wendykan @wendykan]

🙌

[Anthony Goldbloom @antgoldbloom]

.@benhamner and I have started a newco. We're building a large repository of external datasets to help data scientists to find, join and evaluate external data.

We're looking for a full-stack web dev and a hybrid data engineer/machine learner.

DM me if interested.

[wendykan @wendykan]

@chrisgorgo @OpenAI I like this! How did you build it?

[wendykan @wendykan]

@benhamner This! I deal with a lot of URLs these days, all other stuff I'm happy to use external libraries for experiment and then switch to internal for production. urllib is the one 3rd-party library that is far far inferior than the internal library at google.

[wendykan @wendykan]

@eigenhector I'm excited too, already heard some interesting ideas/plans. But seems like funding this year could be competitive given this environment.

[wendykan @wendykan]

@rachel_l_woods Agreed. This paper https://t.co/Z3xgtvR2rI from @huggingface said prompting is often worth 100s of data points on average (see graph). A good prompt adds much more value than 100s of data points for tine tuning.

[wendykan @wendykan]

I started playing with LLM for evaluating my team's use cases. Prompt tuning is so fun but also so dumb at the same time. But this is definitely the most fun I've had at work in a long time.

[wendykan @wendykan]

+@hsuantienlin, Cho-Jui Hsieh.

[arXiv Daily @Arxiv_Daily]

Reducing Training Sample Memorization in GANs by Training with Memorization Rejection
https://t.co/P2pQYxbrkS
by Andrew Bai et al. including @wendykan
#GenerativeAdversarialNetwork #ComputerScience

[wendykan @wendykan]

@DSaience Haha thanks for the convo! I was pretty surprised too since it felt like I've learned nothing recently 😅

[wendykan @wendykan]

Recently I found my ML knowledge/understanding improved. Mildly shocked because I've been spending less time doing hands-on work, but a ton more meetings, and fighting with people in design doc comments.

[wendykan @wendykan]

@DSaience That said I still do hands on work though, just not as much. Now I do maybe 4-8 hr/week, which is not enough for data cleaning 😅 so I focus more on tooling and not production work

[wendykan @wendykan]

@DSaience I think so! It's the iterative process of asking-listening-suggesting-discussing-deciding that distilled technical knowledge without actually doing, I guess.

[wendykan @wendykan]

@DSaience good question, maybe the ability to see the big picture? Things like "should we invest in few-shot learning" and "should we train this embedding with more data sources, is it even worth it to combine data" and having to debate/defend my decisions.

[wendykan @wendykan]

I told @PhilCulliton when I'm sick I have a tendency to want to quit my job.

“Yes, you should quit... for today!” said Phil. Thanks for saving my day!

[wendykan @wendykan]

@SeeRunLove It's going pretty well! Just pulled up the doc and the most recent item was "Pick Apples - make apple dessert or apple cider" and we've been doing that 😋🍎

[wendykan @wendykan]

I keep forgetting things I need to discuss with my husband, to a point that I opened a 1-1 doc, and wrote down bullet points of things we want to cover/discuss/decide "next time we sync"... I can't tell if I'm proud or sad about this.

[wendykan @wendykan]

@DynamicWebPaige ensembling?

[wendykan @wendykan]

oof, model overfit :( data distribution is not stationary even for a short timeframe :(