Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometrycs.AI updates on arXiv.org arXiv:2503.01822v2 Announce Type: replace-cross
Abstract: Sparse Autoencoders (SAEs) are widely used to interpret neural networks by identifying meaningful concepts from their representations. However, do SAEs truly uncover all concepts a model relies on, or are they inherently biased toward certain kinds of concepts? We introduce a unified framework that recasts SAEs as solutions to a bilevel optimization problem, revealing a fundamental challenge: each SAE imposes structural assumptions about how concepts are encoded in model representations, which in turn shapes what it can and cannot detect. This means different SAEs are not interchangeable — switching architectures can expose entirely new concepts or obscure existing ones. To systematically probe this effect, we evaluate SAEs across a spectrum of settings: from controlled toy models that isolate key variables, to semi-synthetic experiments on real model activations and finally to large-scale, naturalistic datasets. Across this progression, we examine two fundamental properties that real-world concepts often exhibit: heterogeneity in intrinsic dimensionality (some concepts are inherently low-dimensional, others are not) and nonlinear separability. We show that SAEs fail to recover concepts when these properties are ignored, and we design a new SAE that explicitly incorporates both, enabling the discovery of previously hidden concepts and reinforcing our theoretical insights. Our findings challenge the idea of a universal SAE and underscores the need for architecture-specific choices in model interpretability. Overall, we argue an SAE does not just reveal concepts — it determines what can be seen at all.
arXiv:2503.01822v2 Announce Type: replace-cross
Abstract: Sparse Autoencoders (SAEs) are widely used to interpret neural networks by identifying meaningful concepts from their representations. However, do SAEs truly uncover all concepts a model relies on, or are they inherently biased toward certain kinds of concepts? We introduce a unified framework that recasts SAEs as solutions to a bilevel optimization problem, revealing a fundamental challenge: each SAE imposes structural assumptions about how concepts are encoded in model representations, which in turn shapes what it can and cannot detect. This means different SAEs are not interchangeable — switching architectures can expose entirely new concepts or obscure existing ones. To systematically probe this effect, we evaluate SAEs across a spectrum of settings: from controlled toy models that isolate key variables, to semi-synthetic experiments on real model activations and finally to large-scale, naturalistic datasets. Across this progression, we examine two fundamental properties that real-world concepts often exhibit: heterogeneity in intrinsic dimensionality (some concepts are inherently low-dimensional, others are not) and nonlinear separability. We show that SAEs fail to recover concepts when these properties are ignored, and we design a new SAE that explicitly incorporates both, enabling the discovery of previously hidden concepts and reinforcing our theoretical insights. Our findings challenge the idea of a universal SAE and underscores the need for architecture-specific choices in model interpretability. Overall, we argue an SAE does not just reveal concepts — it determines what can be seen at all. Read More
While satellite constellations — such as Starlink — are resilient, 2,000 drones could cut communications to a region the size of Taiwan, researchers find. Read More
Is the new privacy protocol helping malicious actors more than Internet users? Read More
The Korean National Police have arrested four individuals suspected of hacking over 120,000 IP cameras across the country and then selling stolen footage to a foreign adult site. […] Read More
OpenAI’s AI-powered ChatGPT is down worldwide with users receiving errors when attempting to access chats, with no reasons currently given. […] Read More
The Federal Trade Commission (FTC) is proposing that education technology provider Illuminate Education to delete unnecessary student data and improve its security to settle allegations related to an incident in 2021 that exposed info of 10 million students. […] Read More
India’s Department of Telecommunications (DoT) has issued directions to app-based communication service providers to ensure that the platforms cannot be used without an active SIM card linked to the user’s mobile number. To that end, messaging apps like WhatsApp, Telegram, Snapchat, Arattai, Sharechat, Josh, JioChat, and Signal that use an Indian mobile number for uniquely […]
Microsoft has confirmed that the KB5070311 preview update is triggering bright white flashes when launching the File Explorer in dark mode on Windows 11 systems. […] Read More
Israeli entities spanning academia, engineering, local government, manufacturing, technology, transportation, and utilities sectors have emerged as the target of a new set of attacks undertaken by Iranian nation-state actors that have delivered a previously undocumented backdoor called MuddyViper. The activity has been attributed by ESET to a hacking group known as MuddyWater (aka Mango Read More
7 ChatGPT Tricks to Automate Your Data TasksKDnuggets This article explores how to transform ChatGPT from a chatbot into a powerful data assistant that streamlines the repetitive, the tedious, and the complex.
This article explores how to transform ChatGPT from a chatbot into a powerful data assistant that streamlines the repetitive, the tedious, and the complex. Read More
