mathematics Thief & Chef "through the window of differential equations, mathematics sees the light in the real world" / "通过微分方程的窗子,数学家看到现实世界的光" (Jiang Zehan)
Seattle
Joined June 2008
- Tweets14.9K
- Following637
- Followers4.4K
- Likes6.8K
deep Manifold retweeted
12. Targeted Neuron Modulation via Contrastive Pair Search
🔑 Keywords: Contrastive Neuron Attribution, Instruction-Tuned Language Models, Behavioral Steering, MLP Neurons, Alignment Fine-Tuning
💡 Category: Natural Language Processing
🌟 Research Objective:
- The study aims to identify a subset of MLP neurons that distinguish harmful prompts from benign ones in instruction-tuned language models in order to achieve effective behavioral steering without compromising output quality.
🛠️ Research Methods:
- The research introduces Contrastive Neuron Attribution (CNA), which identifies specific neurons responsible for distinguishing harmful prompts, utilizing forward passes without gradients or auxiliary training.
💬 Research Conclusions:
- The implementation of neuron-level intervention via CNA enables effective behavioral steering without the quality tradeoffs seen in other methods, such as those operating on the residual stream. It shows significant reduction in refusal rates while maintaining output fluency and coherence.
👉 Paper link: huggingface.co/papers/2605.1…
deep Manifold retweeted
Five easy things you can do to combat cancel culture:
1. Raise awareness by talking to others about it.
2. Support its victims, even if privately.
3. Donate to organizations fighting it.
4. Vote for candidates who oppose it.
5. Boycott companies that cave in to it.
deep Manifold retweeted
Wow wow wow! @GoogleColab now supports the fastai library directly :D That means you can run all the lessons at course.fast.ai, and do your own research and development with fastai, all for free. That's... that's just amazing.
To get started: course.fast.ai/start_colab.h…
deep Manifold retweeted
There's a very interesting Python package for outlier detection called PyOD (Python Outlier Detection). Might be useful for the next fraud/bot detection project or for data cleaning in general. github.com/yzhao062/pyod
deep Manifold retweeted
@fastdotai has been really helpful in doing my experiments with medical images. So I will like to share some of the work I have been doing: github.com/renato145/fastai_…
deep Manifold retweeted
I've been getting used to gganimate and thought it would be useful to put together some illustrations of what various causal inference methods *actually do to data* and how they work. Here, for example, is what it means to control for a (binary) variable
deep Manifold retweeted
Classical causal inference takes for granted that a problem's expressed in the right abstractions. Representation learning today yields no causal insight but can learn abstractions from perceptual data. What is the best recent work you have seen attempting to bridge these worlds?
deep Manifold retweeted
Visual Causal Feature Learning:
arxiv.org/abs/1412.2309
Multi-Level Cause-Effect Systems:
arxiv.org/abs/1512.07942
Unsupervised Discovery of El Nino Using Causal Feature Learning on Microlevel Climate Data:
arxiv.org/abs/1605.09370
deep Manifold retweeted
Pinpoint a range of rare genetic disorders by analyzing pictures of people’s faces. Can detect many syndromes including Down, Holoprosencephaly, & others. 91% top-10 accuracy, trained on 17k points. nature.com/articles/s41591-0…
deep Manifold retweeted
Think you need to learn Bayesian Analysis? Read this first peadarcoyle.com/2019/01/01/t…
deep Manifold retweeted
Fluke Detection Redux, now with
✅more @fastdotai
✅less code 👍🤖
✅quicker train time (16% decrease in train time with same # of epochs)
✅better results (0.85 vs 0.82 IoU)
✅still only 300 train examples
(targets are in green, predictions in red)
github.com/radekosmulski/wha…
deep Manifold retweeted
ml4a has a brand new demos page! including javascript in-browser demos and all the figures for the book.
a section for @ml5js demos is in the works for early-mid january.
ml4a.github.io/demos/
deep Manifold retweeted
Clean and performant PyTorch implementation of Generative Query Networks by Shohei Taniguchi: github.com/iShohei220/torch-… Pixyz implementation by Shohei Taniguchi and Masahiro Suzuki: github.com/masa-su/pixyzoo/t… Pixyz (deep generative modeling library): github.com/masa-su/pixyz 👌
deep Manifold retweeted
We released #PyTorch code of two NeurIPS papers on 3D image synthesis (github.com/junyanz/VON) and 3D scene derendering and manipulation (github.com/ysymyth/3D-SDN). Two projects are part of our ongoing efforts toward integrating 3D graphics and deep generative models.@jiajunwu_cs
deep Manifold retweeted
10 Exciting Ideas of 2018 in NLP: A collection of 10 ideas that I found exciting and impactful this year—and that we'll likely see more of in the future.
ruder.io/10-exciting-ideas-o…
deep Manifold retweeted
A detection result should always have at least two numbers: precision/recall, type I/type II, selectivity/specificity, for at least one reasonable operating point. Or else give the whole damn ROC curve.
deep Manifold retweeted
Pre-trained network for image super resolution (in Keras): github.com/idealo/image-supe…
An evening project would be to export it to TF.js to run in the browser on user-uploaded photos
deep Manifold retweeted
Did you know: fastai includes a computer vision library that lets you write your own transformations that are automatically optimized in a fused pipeline, get random augmentation with a single annotation, and work on images, masks, and sets of coordinates?
docs.fast.ai/vision.transfor…
deep Manifold retweeted
Brought to you by NYU, NYU-Shanghai, and Amazon AWS.
DGL (Deep Graph Library) - Clean and efficient library to build graph neural networks including GCN, TreeLSTM and graph generative models. Includes auto-batching and other tricks for speed.
github.com/dmlc/dgl
