A Machine Learning Approach for Denoising and Upsampling HRTFs

The demand for realistic virtual immersive audio continues to grow, with Head-Related Transfer Functions (HRTFs) playing a key role. HRTFs capture how sound reaches our ears, reflecting unique anatomical features and enhancing spatial perception. It has been shown that personalized HRTFs improve localization accuracy, but their measurement remains time-consuming and requires a noise-free environment. Although machine learning has been shown to reduce the required measurement points and, thus, the measurement time, a controlled environment is still necessary. This paper proposes a method to address this constraint by presenting a novel technique that can upsample sparse, noisy HRTF measurements. The proposed approach combines an HRTF Denoisy U-Net for denoising and an Autoencoding Generative Adversarial Network (AE-GAN) for upsampling from three measurement points. The proposed method achieves a log-spectral distortion (LSD) error of 5.41 dB and a cosine similarity loss of 0.0070, demonstrating the method's effectiveness in HRTF upsampling.

Emergence and Evolution of Interpretable Concepts in Diffusion Models

Diffusion models have become the go-to method for text-to-image generation, producing high-quality images from noise through a process called reverse diffusion. Understanding the dynamics of the reverse diffusion process is crucial in steering the generation and achieving high sample quality. However, the inner workings of diffusion models is still largely a mystery due to their black-box nature and complex, multi-step generation process. Mechanistic Interpretability (MI) techniques, such as Sparse Autoencoders (SAEs), aim at uncovering the operating principles of models through granular analysis of their internal representations. These MI techniques have been successful in understanding and steering the behavior of large language models at scale. However, the great potential of SAEs has not yet been applied toward gaining insight into the intricate generative process of diffusion models. In this work, we leverage the SAE framework to probe the inner workings of a popular text-to-image diffusion model, and uncover a variety of human-interpretable concepts in its activations. Interestingly, we find that even before the first reverse diffusion step is completed, the final composition of the scene can be predicted surprisingly well by looking at the spatial distribution of activated concepts. Moreover, going beyond correlational analysis, we show that the discovered concepts have a causal effect on the model output and can be leveraged to steer the generative process. We design intervention techniques aimed at manipulating image composition and style, and demonstrate that (1) in early stages of diffusion image composition can be effectively controlled, (2) in the middle stages of diffusion image composition is finalized, however stylistic interventions are effective, and (3) in the final stages of diffusion only minor textural details are subject to change.

Understanding the Repeat Curse in Large Language Models from a Feature Perspective

Large language models (LLMs) have made remarkable progress in various domains, yet they often suffer from repetitive text generation, a phenomenon we refer to as the "Repeat Curse". While previous studies have proposed decoding strategies to mitigate repetition, the underlying mechanism behind this issue remains insufficiently explored. In this work, we investigate the root causes of repetition in LLMs through the lens of mechanistic interpretability. Inspired by recent advances in Sparse Autoencoders (SAEs), which enable monosemantic feature extraction, we propose a novel approach, "Duplicatus Charm", to induce and analyze the Repeat Curse. Our method systematically identifies "Repetition Features" -the key model activations responsible for generating repetitive outputs. First, we locate the layers most involved in repetition through logit analysis. Next, we extract and stimulate relevant features using SAE-based activation manipulation. To validate our approach, we construct a repetition dataset covering token and paragraph level repetitions and introduce an evaluation pipeline to quantify the influence of identified repetition features. Furthermore, by deactivating these features, we have effectively mitigated the Repeat Curse.

Scaling sparse feature circuit finding for in-context learning

Sparse autoencoders (SAEs) are a popular tool for interpreting large language model activations, but their utility in addressing open questions in interpretability remains unclear. In this work, we demonstrate their effectiveness by using SAEs to deepen our understanding of the mechanism behind in-context learning (ICL). We identify abstract SAE features that (i) encode the model's knowledge of which task to execute and (ii) whose latent vectors causally induce the task zero-shot. This aligns with prior work showing that ICL is mediated by task vectors. We further demonstrate that these task vectors are well approximated by a sparse sum of SAE latents, including these task-execution features. To explore the ICL mechanism, we adapt the sparse feature circuits methodology of Marks et al. (2024) to work for the much larger Gemma-1 2B model, with 30 times as many parameters, and to the more complex task of ICL. Through circuit finding, we discover task-detecting features with corresponding SAE latents that activate earlier in the prompt, that detect when tasks have been performed. They are causally linked with task-execution features through the attention and MLP sublayers.

Application of Deep Generative Models for Anomaly Detection in Complex Financial Transactions

This study proposes an algorithm for detecting suspicious behaviors in large payment flows based on deep generative models. By combining Generative Adversarial Networks (GAN) and Variational Autoencoders (VAE), the algorithm is designed to detect abnormal behaviors in financial transactions. First, the GAN is used to generate simulated data that approximates normal payment flows. The discriminator identifies anomalous patterns in transactions, enabling the detection of potential fraud and money laundering behaviors. Second, a VAE is introduced to model the latent distribution of payment flows, ensuring that the generated data more closely resembles real transaction features, thus improving the model's detection accuracy. The method optimizes the generative capabilities of both GAN and VAE, ensuring that the model can effectively capture suspicious behaviors even in sparse data conditions. Experimental results show that the proposed method significantly outperforms traditional machine learning algorithms and other deep learning models across various evaluation metrics, especially in detecting rare fraudulent behaviors. Furthermore, this study provides a detailed comparison of performance in recognizing different transaction patterns (such as normal, money laundering, and fraud) in large payment flows, validating the advantages of generative models in handling complex financial data.

MIB: A Mechanistic Interpretability Benchmark

How can we know whether new mechanistic interpretability methods achieve real improvements? In pursuit of meaningful and lasting evaluation standards, we propose MIB, a benchmark with two tracks spanning four tasks and five models. MIB favors methods that precisely and concisely recover relevant causal pathways or specific causal variables in neural language models. The circuit localization track compares methods that locate the model components - and connections between them - most important for performing a task (e.g., attribution patching or information flow routes). The causal variable localization track compares methods that featurize a hidden vector, e.g., sparse autoencoders (SAEs) or distributed alignment search (DAS), and locate model features for a causal variable relevant to the task. Using MIB, we find that attribution and mask optimization methods perform best on circuit localization. For causal variable localization, we find that the supervised DAS method performs best, while SAE features are not better than neurons, i.e., standard dimensions of hidden vectors. These findings illustrate that MIB enables meaningful comparisons of methods, and increases our confidence that there has been real progress in the field.

A Real-time Anomaly Detection Method for Robots based on a Flexible and Sparse Latent Space

The growing demand for robots to operate effectively in diverse environments necessitates the need for robust real-time anomaly detection techniques during robotic operations. However, deep learning-based models in robotics face significant challenges due to limited training data and highly noisy signal features. In this paper, we present Sparse Masked Autoregressive Flow-based Adversarial AutoEncoders model to address these problems. This approach integrates Masked Autoregressive Flow model into Adversarial AutoEncoders to construct a flexible latent space and utilize Sparse autoencoder to efficiently focus on important features, even in scenarios with limited feature space. Our experiments demonstrate that the proposed model achieves a 4.96% to 9.75% higher area under the receiver operating characteristic curve for pick-and-place robotic operations with randomly placed cans, compared to existing state-of-the-art methods. Notably, it showed up to 19.67% better performance in scenarios involving collisions with lightweight objects. Additionally, unlike the existing state-of-the-art model, our model performs inferences within 1 millisecond, ensuring real-time anomaly detection. These capabilities make our model highly applicable to machine learning-based robotic safety systems in dynamic environments. The code will be made publicly available after acceptance.

Uncovering Branch specialization in InceptionV1 using k sparse autoencoders

Sparse Autoencoders (SAEs) have shown to find interpretable features in neural networks from polysemantic neurons caused by superposition. Previous work has shown SAEs are an effective tool to extract interpretable features from the early layers of InceptionV1. Since then, there have been many improvements to SAEs but branch specialization is still an enigma in the later layers of InceptionV1. We show various examples of branch specialization occuring in each layer of the mixed4a-4e branch, in the 5x5 branch and in one 1x1 branch. We also provide evidence to claim that branch specialization seems to be consistent across layers, similar features across the model will be localized in the same convolution size branches in their respective layer.

Interpreting the Linear Structure of Vision-language Model Embedding Spaces

Vision-language models encode images and text in a joint space, minimizing the distance between corresponding image and text pairs. How are language and images organized in this joint space, and how do the models encode meaning and modality? To investigate this, we train and release sparse autoencoders (SAEs) on the embedding spaces of four vision-language models (CLIP, SigLIP, SigLIP2, and AIMv2). SAEs approximate model embeddings as sparse linear combinations of learned directions, or "concepts". We find that, compared to other methods of linear feature learning, SAEs are better at reconstructing the real embeddings, while also able to retain the most sparsity. Retraining SAEs with different seeds or different data diet leads to two findings: the rare, specific concepts captured by the SAEs are liable to change drastically, but we also show that the key commonly-activating concepts extracted by SAEs are remarkably stable across runs. Interestingly, while most concepts are strongly unimodal in activation, we find they are not merely encoding modality per se. Many lie close to - but not entirely within - the subspace defining modality, suggesting that they encode cross-modal semantics despite their unimodal usage. To quantify this bridging behavior, we introduce the Bridge Score, a metric that identifies concept pairs which are both co-activated across aligned image-text inputs and geometrically aligned in the shared space. This reveals that even unimodal concepts can collaborate to support cross-modal integration. We release interactive demos of the SAEs for all models, allowing researchers to explore the organization of the concept spaces. Overall, our findings uncover a sparse linear structure within VLM embedding spaces that is shaped by modality, yet stitched together through latent bridges-offering new insight into how multimodal meaning is constructed.

Training Autoencoders Using Stochastic Hessian-Free Optimization with LSMR

Hessian-free (HF) optimization has been shown to effectively train deep autoencoders (Martens, 2010). In this paper, we aim to accelerate HF training of autoencoders by reducing the amount of data used in training. HF utilizes the conjugate gradient algorithm to estimate update directions. Instead, we propose using the LSMR method, which is known for effectively solving large sparse linear systems. We also incorporate Chapelle & Erhan (2011)'s improved preconditioner for HF optimization. In addition, we introduce a new mini-batch selection algorithm to mitigate overfitting. Our algorithm starts with a small subset of the training data and gradually increases the mini-batch size based on (i) variance estimates obtained during the computation of a mini-batch gradient (Byrd et al., 2012) and (ii) the relative decrease in objective value for the validation data. Our experimental results demonstrate that our stochastic Hessian-free optimization, using the LSMR method and the new sample selection algorithm, leads to rapid training of deep autoencoders with improved generalization error.