探索全球前沿学术脉络

Ensembling Sparse Autoencoders

arXiv:2505.16077v2 Announce Type: replace Abstract: Sparse autoencoders (SAEs) are used to decompose neural network activations into human-interpretable features. Typically, features learned by a single SAE are used for downstream applications. However, it has recently been shown that a single SAE captures only a limited subset of features that can be extracted from the activation space. Motivated by this limitation, we introduce and formalize SAE ensembles. Furthermore, we propose to ensemble multiple SAEs through naive bagging and boosting. In naive bagging, SAEs trained with different weight initializations are ensembled, whereas in boosting SAEs sequentially trained to minimize the residual error are ensembled. Theoretically, naive bagging and boosting are justified as approaches to reduce reconstruction error. Empirically, we evaluate our ensemble approaches with three settings of language models and SAE architectures. Our empirical results demonstrate that, compared to an expanded SAE that matches the number of features in the ensemble, ensembling SAEs improves the reconstruction of language model activations along with SAE stability. Additionally, on downstream tasks such as concept detection and spurious correlation removal, SAE ensembles achieve better performance, showing improved practical utility.

Cross-Device Adaptation of Mirai for Mammography-Based Breast Cancer Risk Prediction

Fine-tuning can adapt pretrained medical imaging models to new clinical datasets, but device-specific domain shifts may limit generalizability. We evaluated Mirai, a mammography-based deep learning model for breast cancer risk prediction, in a large screening cohort containing Hologic and General Electric (GE) full-field digital mammography systems, including GE Premium View (GE PV) and Tissue Equalization (GE TE) post-processing software. Native Mirai showed lower performance on TE images than on Hologic or PV images. Fine-tuning on TE images improved TE performance, particularly for short-term risk prediction, but substantially reduced performance on Hologic images, consistent with catastrophic forgetting. To mitigate this effect, we developed a device-invariant model using interleaved multi-device sampling and conditional adversarial training. This approach largely restored Hologic performance while maintaining improved TE performance, providing better robustness across heterogeneous imaging platforms. Comparison of cumulative and annual risk AUCs over a five-year time horizon further showed that performance gains were driven mainly by short- and intermediate-term predictions. These findings highlight both the value and dangers of device-specific fine-tuning and support balanced domain-adaptation strategies for deploying mammography-based risk models across diverse clinical imaging environments.