T. @ Techtonique
teeeeeeeee.bsky.social
tech, stats, machine learning, computer simulation, numerical optimization https://github.com/Techtonique
👤 Joined on 10 December 2024
92
Followers
14
Following
4795
Total Posts
📜 Kullanıcı Geçmişi
- 10 Ara 2024, 17:30
🎉 Bluesky'e katıldı - İlk kullanıcı adı:
@teeeeeeeee.bsky.social
🚫 Blokladığı Kullanıcılar
❤️ Beğendiği Gönderiler
the rare comeback story where the thing actually comes back. respect to whatever server gremlins got evicted
Show Post
KRR surrogates are underused for this. In my experience the hard part is choosing the kernel bandwidth to match the local structure of the original model rather than global behavior. Gaussian kernels tend to over-smooth near decision boundaries.
Show Post
Old thread but to me very interesting. Advanced, but relatively simple, and interesting notes about confidence interval calculations. On my list of things to try next..
Show Post
Not exactly. You don't need to integrate a method into the package. You can use a "post-hoc" method instead. These are separate methods applied to a model after it has been trained to explain it's predcitions. A great package for neural networks (if you are using PyTorch) is captum: captum.ai
Show PostI don't have much experience with R, but I know a good package is IML: cran.r-project.org/web/packages... However, this one focuses on permutation-based XAI methods, not gradient-based ones as you have suggested.
Show Post