Adam Li
@adam2392
Followers
363
Following
9K
Media
33
Statuses
2K
Postdoc @Columbiacompsci. Comp neuro, dyn systems and causal data science. Previously @ucsdbe, @ucsdmathdept, @JHUBME @ICM_JHU. Passionate about #opensource
New York, USA
Joined October 2014
I take the @LIRR to work everyday and pay $250+ for a monthly “peak” ticket for over 1 year now. I have never had the train come on time a single time. Not a single time. You get to work at least 10-15 mins later than estimated on a daily basis. How has this become acceptable?
55
8
373
🚨 $100 GIVEAWAY 🚨 At MaxoutDeals: 📦 We post deals 💳 You order w/ Amazon cashback (5–7%) 💵 We pay commissions → double earnings! Now we’re giving back 👇 Win $100 + invite friends & fam to join MaxoutDeals! How to enter: 1⃣ Follow @DealsCashouts 2⃣ Like ❤️ & Retweet 🔁
10
46
47
Customer service: expect update in 3-5 business days. 2 weeks later: expect update in 3-5 business days. LOL what?
0
0
0
Given a sufficient set of heterogenous distributions, our theory predicts when representations are disentangleable. For example, in (a), when we perturb the representation of digit color, the digit and the bar color rarely change. 6/n
0
0
0
Empirically, we demonstrate how these disentangled representations can be used to robustly edit an image. We train a G-constrained neural network as a proxy model to map input images to latent causal variables: bar color, digit color, and the number of the digit. 5/n
1
0
0
We learn a proxy graphically constrained model that estimates the mixing function, and the corresponding latent variable distributions using data arising from multiple distributions. 4/n
1
0
0
We study disentangled causal representation learning from three axes: Input assumptions, input data, and output disentanglement. We develop a general symbolic algorithm for identifying when latent variables are disentangleable in the learned representations. 3/n
1
0
0
Our work studies when learned causal representations are disentangled, and demonstrates robust image-editing capabilities given a neural network model trained on only images. https://t.co/lGgdUvpm4O 2/n
1
0
0
Are you interested in provable disentangled causal representation learning and obtaining interpretable and controllable representations? We (Yushu Pan, Adam Li, and @eliasbareinboim) are presenting #NeurIPS2024 work at #5105 on Thursday December 12th, at 11AM-2PM PST. 1/n
1
0
4
The 2024 scikit-learn user survey is out! Please join this structured dialogue with the scikit-learn team to better guide and prioritize decision-making about the development of the project: https://t.co/VZdaNzRGMk
docs.google.com
This survey is being conducted by the scikit-learn survey team to ensure that scikit-learn evolves in a way that benefits its user community. Participation in this survey is voluntary and it offers...
1
15
26
New paper in Imaging Neuroscience by Russell A. Poldrack, Krzysztof J. Gorgolewski, et al: The past, present, and future of the brain imaging data structure (BIDS) https://t.co/97QyI12fsX
4
164
454
It’s unclear to myself why I choose to still fly @united to EWR. Experience is about as bad as @SpiritAirlines lmao. Might be finally time to pay more to go somewhere else.
1
0
0
He inadvertently created an icon of one of the biggest deceptions in our lifetime. I wish it weren’t true. Recycling was invented by the plastic industry as a way to transfer responsibility for waste to the individual. There are little to no economics in recycling. Of the
In 1970, the Container Corporation of America organized a design competition to create a symbol for recycled paper. Gary Anderson, a 23-year-old engineering student at the University of Southern California, submitted his design and won the contest. The design, which consisted of
196
813
4K
If you compare patient data in hospital A vs hospital B, certain distributions differ while others do not. Graphical conditions on the selection diagram capture these invariances. 4/n
1
0
1
We propose a framework for characterizing the invariances we expect to see in our multi-domain datasets when given a selection diagram. What does this mean in the context of data? 3/n
1
0
1
Dealing with multi-domain data is a problem studied under the lens of transportability. Very general conditions about when we can transport causal effects from one domain to another have previously been established. These have been formalized using the selection diagram. 2/n
1
0
1
Interested in causal discovery? Do you have access to observations and experiments arising from multiple domains (e.g. different lab settings, humans and bonobos, different countries)? See here for a summary of our #NeurIPS2023 paper and a short🧵:
1/5 Happy to share some of our latest work that will be presented this week at NeurIPS in New Orleans! The authors would be delighted to see you at the poster session and talk more about our current work and future challenges! Tue 6:15 pm (Poster Session 2) "Causal discovery
1
1
12
Hey #neuroscience twitter. Does anyone know where to find details of NIH's requirement for data sharing upon publication? I would like to convince collaborators to allow us to openly share data that has been valuably curated and collected.
1
0
2