False faced: can a simple classifier tell when a model is manipulating its activations to evade probes?
来源: LessWrong ·
This experiment was a submission to my Statistics & Probability course Bayesian classifiers project. TL;DR: We test whether simple classifiers can tell when a model is hiding concepts in its activations to evade probes. We find that, for Bayesian classifiers, the probabilities are enough to separate most evading rows from the rest on concepts seen in training, but the correlation aspect through the pairs of dimensions that the Dependent version adds over the Naive approach does not measurably help. None of the classifiers work on a new concept, because they learn the trigger word that elicits
常见问题
什么是 False faced: can a simple classifier tell when a model is manipulating its activations to evade probes?？
False faced: can a simple classifier tell when a model is manipulating its activations to evade probes? is a 文章 discovery curated from LessWrong. This experiment was a submission to my Statistics & Probability course Bayesian classifiers project. TL;DR: We test whether simple classifiers can tell when a model is hiding…
如何访问或体验此内容？
You can experience this content directly on its original host (LessWrong) by clicking the "Open Original" button.
此内容来源于哪里？
Curated from LessWrong by the OnSuggest team to surface the most interesting and serendipitous content on the web.
相关发现
所有 文化与思想 发现 →I Buy Bag After Bag of These Crunchy Tortilla Chips (They’re the Absolute Best)
They have the perfect crunch. READ MORE...
Thanksgiving Delight
I became a pecan-lover sitting under a tree. When I was growing up on a farm in Arkansas, we’d gather pecans by the bucketful every November. I’d sometimes take a break, back against the trunk of our enormous pecan tree (the biggest in the state, according to my dad), and squeeze two pecans in my hand until the shells cracked, sneaking a snack. Every year my grandma would turn some of this bounty into a pecan pie for Thanksgiving dinner. READ MORE...
The Gordon Ramsay-Approved Cookware Is on Major Sale for Black Friday
Black Friday deals can help you bring home fan favorites for a fraction of the price, especially if you want the best of the best cookware and cutlery for your kitchen. And one favorite brand you don’t want to miss this holiday season is HexClad. READ MORE...
How to Watch a VOD Documentary on Any Screen: TV, Phone, and Computer Setup
A practical guide to finding a documentary’s official VOD service and watching it on a supported TV, phone, tablet or computer, with casting, HDMI and troubleshooting steps. The post How to Watch a VOD Documentary on Any Screen: TV, Phone, and Computer Setup appeared first on DocumentaryTube .
Can You Stream IDFA Films at Home? How to Find a Film Legally
IDFA has links to selected films from past editions, but home viewing depends on each film’s current streaming rights and your location. Here’s how to check what you can watch. The post Can You Stream IDFA Films at Home? How to Find a Film Legally appeared first on DocumentaryTube .
How to Stream On Patrol: Live and Find Episodes From Your Area
Peacock’s official route is REELZ in Channels at 9 p.m. ET Fridays and Saturdays, with next-day on-demand episodes for eligible subscribers. Here’s how to check provider availability and search episode descriptions for local departments. The post How to Stream On Patrol: Live and Find Episodes From Your Area appeared first on DocumentaryTube .