As we increase the number of features in the SAE, hierarchical features tend to split into finer features ("math" may split into "algebra", "geometry", etc.), a phenomenon referred to as feature splitting. However, we show that sparse decomposition and splitting of hierarchical features is not robust. Specifically, we show that seemingly monosemantic features fail to fire where they should, and instead get "absorbed" into their children features. We coin this phenomenon feature absorption, and show that it is caused by optimizing for sparsity in SAEs whenever the underlying features form a hierarchy. We introduce a metric to detect absorption in SAEs, and validate our findings empirically on hundreds of LLM SAEs. Our investigation suggests that varying SAE sizes or sparsity is insufficient to solve this issue. We discuss the implications of feature absorption in SAEs and some potential approaches to solve the fundamental theoretical issues before SAEs can be used for interpreting LLMs robustly and at scale.

In this paper, we take another approach to diffusion model acceleration.

Sign language translation (SL T) aims to interpret sign video sequences into text-based natural language sentences.

GAN-based image attribute editing firstly leverages GAN Inversion to project real images into the latent space of GAN and then manipulates corresponding latent codes. Recent inversion methods mainly utilize additional high-bit features to improve image details preservation, as low-bit codes cannot faithfully reconstruct source images, leading to the loss of details. However, during editing, existing works fail to accurately complement the lost details and suffer from poor editability. The main reason is they inject all the lost details indiscriminately at one time, which inherently induces the position and quantity of details to overfit source images, resulting in inconsistent content and artifacts in edited images. This work argues that details should be gradually injected into both the reconstruction and editing process in a multi-stage coarse-to-fine manner for better detail preservation and high editability. Therefore, a novel dual-stream framework is proposed to accurately complement details at each stage. The Reconstruction Stream is employed to embed coarse-to-fine lost details into residual features and then adaptively add them to the GAN generator. In the Editing Stream, residual features are accurately aligned by our Selective Attention mechanism and then injected into the editing process in a multi-stage manner. Extensive experiments have shown the superiority of our framework in both reconstruction accuracy and editing quality compared with existing methods.

Wikipedia articles are hierarchically organized through categories and lists, providing one of the most comprehensive and universal taxonomy, but its open creation is causing redundancies and inconsistencies. Assigning DBPedia classes to Wikipedia categories and lists can alleviate the problem, realizing a large knowledge graph which is essential for categorizing digital contents through entity linking and typing. However, the existing approach of CaLiGraph is producing incomplete and non-fine grained mappings. In this paper, we tackle the problem as ontology alignment, where structural information of knowledge graphs and lexical and semantic features of ontology class names are utilized to discover confident mappings, which are in turn utilized for finetuing pretrained language models in a distant supervision fashion. Our method SLHCat consists of two main parts: 1) Automatically generating training data by leveraging knowledge graph structure, semantic similarities, and named entity typing. 2) Finetuning and prompt-tuning of the pre-trained language model BERT are carried out over the training data, to capture semantic and syntactic properties of class names. Our model SLHCat is evaluated over a benchmark dataset constructed by annotating 3000 fine-grained CaLiGraph-DBpedia mapping pairs. SLHCat is outperforming the baseline model by a large margin of 25% in accuracy, offering a practical solution for large-scale ontology mapping.

Deep Gaussian Processes (DGP) enable a non-parametric approach to quantify the uncertainty of complex deep machine learning models. Conventional inferential methods for DGP models can suffer from high computational complexity as they require large-scale operations with kernel matrices for training and inference. In this work, we propose an efficient scheme for accurate inference and prediction based on a range of Gaussian Processes, called the Tensor Markov Gaussian Processes (TMGP). We construct an induced approximation of TMGP referred to as the hierarchical expansion. Next, we develop a deep TMGP (DTMGP) model as the composition of multiple hierarchical expansion of TMGPs. The proposed DTMGP model has the following properties: (1) the outputs of each activation function are deterministic while the weights are chosen independently from standard Gaussian distribution; (2) in training or prediction, only O(polylog(M)) (out of M) activation functions have non-zero outputs, which significantly boosts the computational efficiency. Our numerical experiments on real datasets show the superior computational efficiency of DTMGP versus other DGP models.