paper-with-me

홈 › Papers

Scale Alone Does not Improve Mechanistic Interpretability in Vision Models

2023-07-11 · NeurIPS 2023 11 · Roland S. Zimmermann, Thomas Klein, Wieland Brendel

In light of the recent widespread adoption of AI systems, understanding the internal information processing of neural networks has become increasingly critical. Most recently, machine vision has seen remarkable progress by scaling neural networks to unprecedented levels in dataset and model size. We here ask whether this extraordinary increase in scale also positively impacts the field of mechanistic interpretability. In other words, has our understanding of the inner workings of scaled neural networks improved as well? We use a psychophysical paradigm to quantify one form of mechanistic interpretability for a diverse suite of nine models and find no scaling effect for interpretability - neither for model nor dataset size. Specifically, none of the investigated state-of-the-art models are easier to interpret than the GoogLeNet model from almost a decade ago. Latest-generation vision models appear even less interpretable than older architectures, hinting at a regression rather than improvement, with modern models sacrificing interpretability for accuracy. These results highlight the need for models explicitly designed to be mechanistically interpretable and the need for more helpful interpretability methods to increase our understanding of networks at an atomic level. We release a dataset containing more than 130'000 human responses from our psychophysical evaluation of 767 units across nine models. This dataset facilitates research on automated instead of human-based interpretability evaluations, which can ultimately be leveraged to directly optimize the mechanistic interpretability of models.

📄 PDF Abstract BibTeX arXiv:2307.05471

Code (0)

등록된 구현이 없습니다.

Methods 이 논문이 사용한 방법론

Auxiliary Classifier Auxiliary Classifiers are type of architectural component that seek to improve the convergence of very deep networks. They are classifier heads we attach to layers before the…
Local Response Normalization Local Response Normalization is a normalization layer that implements the idea of lateral inhibition. Lateral inhibition is a concept in neurobiology that refers to the…
Convolution A convolution is a type of matrix operation, consisting of a kernel, a small matrix of weights, that slides over input data performing element-wise multiplication with the…
Dense Connections Dense Connections, or Fully Connected Connections, are a type of layer in a deep neural network that use a linear operation where every input is connected to every output…
Dropout Dropout is a regularization technique for neural networks that drops a unit (along with connections) at training time with a specified probability $p$ (a common value is…
Softmax The Softmax output function transforms a previous layer's output into a vector of probabilities. It is commonly used for multiclass classification. Given an input vector $x$…
1x1 Convolution A 1 x 1 Convolution is a convolution with some special properties in that it can be used for dimensionality reduction,…
Inception Module An Inception Module is an image model block that aims to approximate an optimal local sparse structure in a CNN. Put simply, it allows for us to use multiple types of filter…

Similar Papers 제목 키워드 기반

From Observation to Insight: Mechanistic World Models and the Quest for Autonomous Discovery

2026-07-14 · Ingmar Posner, Anson Lei, Bernhard Schölkopf arxiv

Recent advances in foundation models have transformed AI for Science, enabling remarkably accurate predictive performance across domains ranging from protein folding to weather forecasting. Yet prediction alone does not …

Representation LearningWeather Forecasting

Mechanistic Interpretability of Antibody Language Models Using SAEs

2025-12-05 · Rebonto Haque, Oliver M. Turnbull, Anisha Parsan, Nithin Parsan 외 arxiv

Sparse autoencoders (SAEs) are a mechanistic interpretability technique that have been used to provide insight into learned concepts within large protein language models. Here, we employ TopK and Ordered SAEs to investig…

Automated Attention Pattern Discovery at Scale in Large Language Models

2026-04-04 · Jonathan Katzy, Razvan-Mihai Popescu, Erik Mekkes, Arie van Deursen 외 arxiv

Large language models have found success by scaling up capabilities to work in general settings. The same can unfortunately not be said for interpretability methods. The current trend in mechanistic interpretability is t…

Whether, Not Which: Mechanistic Interpretability Reveals Dissociable Affect Reception and Emotion Categorization in LLMs

2026-03-15 · Michael Keeman arxiv

Large language models appear to develop internal representations of emotion -- "emotion circuits," "emotion neurons," and structured emotional manifolds have been reported across multiple model families. But every study …

Entanglement as Memory: Mechanistic Interpretability of Quantum Language Models

2026-03-27 · Nathan Roll arxiv

Quantum language models have shown competitive performance on sequential tasks, yet whether trained quantum circuits exploit genuinely quantum resources -- or merely embed classical computation in quantum hardware -- rem…