State-of-the-art results in large language models (LLMs) often rely on scale, which becomes computationally expensive. This has sparked a research agenda to reduce these models' parameter count and computational costs without significantly impacting their performance. Our study focuses on transformer-based LLMs, specifically targeting the computationally intensive feedforward networks (FFN), which are less studied than attention blocks. We consider three candidate linear layer approximations in the FFN by combining efficient low-rank and block-diagonal matrices. In contrast to many previous works that examined these approximations, our study i) explores these structures from the training-from-scratch perspective, ii) scales up to 1.3B parameters, and iii) is conducted within recent Transformer-based LLMs rather than convolutional architectures. We first demonstrate they can lead to actual computational gains in various scenarios, including online decoding when using a pre-merge technique. Additionally, we propose a novel training regime, called \textit{self-guided training}, aimed at improving the poor training dynamics that these approximations exhibit when used from initialization. Experiments on the large RefinedWeb dataset show that our methods are both efficient and effective for training and inference. Interestingly, these structured FFNs exhibit steeper scaling curves than the original models. Further applying self-guided training to the structured matrices with 32\% FFN parameters and 2.5$\times$ speed-up enables only a 0.4 perplexity increase under the same training FLOPs. Finally, we develop the wide and structured networks surpassing the current medium-sized and large-sized Transformer in perplexity and throughput performance. Our code is available at \url{https://github.com/CLAIRE-Labo/StructuredFFN/tree/main}.

Despite the high diagnostic accuracy of Magnetic Resonance Imaging (MRI), using MRI as a Point-of-Care (POC) disease identification tool poses significant accessibility challenges due to the use of high magnetic field strength and lengthy acquisition times. We ask a simple question: Can we dynamically optimise acquired samples, at the patient level, according to an (automated) downstream decision task, while discounting image reconstruction? We propose an ML-based framework that learns an active sampling strategy, via reinforcement learning, at a patient-level to directly infer disease from undersampled k-space. We validate our approach by inferring Meniscus Tear in undersampled knee MRI data, where we achieve diagnostic performance comparable with ML-based diagnosis, using fully sampled k-space data. We analyse task-specific sampling policies, showcasing the adaptability of our active sampling approach. The introduced frugal sampling strategies have the potential to reduce high field strength requirements that in turn strengthen the viability of MRI-based POC disease identification and associated preliminary screening tools.

As part of an ongoing worldwide effort to comprehend and monitor insect biodiversity, this paper presents the BIOSCAN-5M Insect dataset to the machine learning community and establish several benchmark tasks. BIOSCAN-5M is a comprehensive dataset containing multi-modal information for over 5 million insect specimens, and it significantly expands existing image-based biological datasets by including taxonomic labels, raw nucleotide barcode sequences, assigned barcode index numbers, and geographical information. We propose three benchmark experiments to demonstrate the impact of the multi-modal data types on the classification and clustering accuracy. First, we pretrain a masked language model on the DNA barcode sequences of the BIOSCAN-5M dataset, and demonstrate the impact of using this large reference library on species- and genus-level classification performance. Second, we propose a zero-shot transfer learning task applied to images and DNA barcodes to cluster feature embeddings obtained from self-supervised learning, to investigate whether meaningful clusters can be derived from these representation embeddings. Third, we benchmark multi-modality by performing contrastive learning on DNA barcodes, image data, and taxonomic information. This yields a general shared embedding space enabling taxonomic classification using multiple types of information and modalities. The code repository of the BIOSCAN-5M Insect dataset is available at https://github.com/zahrag/BIOSCAN-5M.

Sign Language Recognition (SLR) is a fast-growing field that aims to fill the communication gaps between the hearing-impaired and people without hearing loss. Existing solutions for Persian Sign Language (PSL) are limited to word-level interpretations, underscoring the need for more advanced and comprehensive solutions. Moreover, previous work on other languages mainly focuses on manipulating the neural network architectures or hardware configurations instead of benefiting from the aggregated results of multiple models. In this paper, we introduce PenSLR, a glove-based sign language system consisting of an Inertial Measurement Unit (IMU) and five flexible sensors powered by a deep learning framework capable of predicting variable-length sequences. We achieve this in an end-to-end manner by leveraging the Connectionist Temporal Classification (CTC) loss function, eliminating the need for segmentation of input signals. To further enhance its capabilities, we propose a novel ensembling technique by leveraging a multiple sequence alignment algorithm known as Star Alignment. Furthermore, we introduce a new PSL dataset, including 16 PSL signs with more than 3000 time-series samples in total. We utilize this dataset to evaluate the performance of our system based on four word-level and sentence-level metrics. Our evaluations show that PenSLR achieves a remarkable word accuracy of 94.58% and 96.70% in subject-independent and subject-dependent setups, respectively. These achievements are attributable to our ensembling algorithm, which not only boosts the word-level performance by 0.51% and 1.32% in the respective scenarios but also yields significant enhancements of 1.46% and 4.00%, respectively, in sentence-level accuracy.

Did you try out the new Bing Search? Or maybe you fiddled around with Google AI~Overviews? These might sound familiar because the modern-day search stack has recently evolved to include retrieval-augmented generation (RAG) systems. They allow searching and incorporating real-time data into large language models (LLMs) to provide a well-informed, attributed, concise summary in contrast to the traditional search paradigm that relies on displaying a ranked list of documents. Therefore, given these recent advancements, it is crucial to have an arena to build, test, visualize, and systematically evaluate RAG-based search systems. With this in mind, we propose the TREC 2024 RAG Track to foster innovation in evaluating RAG systems. In our work, we lay out the steps we've made towards making this track a reality -- we describe the details of our reusable framework, Ragnar\"ok, explain the curation of the new MS MARCO V2.1 collection choice, release the development topics for the track, and standardize the I/O definitions which assist the end user. Next, using Ragnar\"ok, we identify and provide key industrial baselines such as OpenAI's GPT-4o or Cohere's Command R+. Further, we introduce a web-based user interface for an interactive arena allowing benchmarking pairwise RAG systems by crowdsourcing. We open-source our Ragnar\"ok framework and baselines to achieve a unified standard for future RAG systems.

Accommodating long sequences efficiently in autoregressive Transformers, especially within an extended context window, poses significant challenges due to the quadratic computational complexity and substantial KV memory requirements inherent in self-attention mechanisms. In this work, we introduce SPARSEK Attention, a novel sparse attention mechanism designed to overcome these computational and memory obstacles while maintaining performance. Our approach integrates a scoring network and a differentiable top-k mask operator, SPARSEK, to select a constant number of KV pairs for each query, thereby enabling gradient-based optimization. As a result, SPARSEK Attention offers linear time complexity and constant memory footprint during generation. Experimental results reveal that SPARSEK Attention outperforms previous sparse attention methods and provides significant speed improvements during both training and inference, particularly in language modeling and downstream tasks. Furthermore, our method can be seamlessly integrated into pre-trained Large Language Models (LLMs) with minimal fine-tuning, offering a practical solution for effectively managing long-range dependencies in diverse applications.

Current trends in audio anti-spoofing detection research strive to improve models' ability to generalize across unseen attacks by learning to identify a variety of spoofing artifacts. This emphasis has primarily focused on the spoof class. Recently, several studies have noted that the distribution of silence differs between the two classes, which can serve as a shortcut. In this paper, we extend class-wise interpretations beyond silence. We employ loss analysis and asymmetric methodologies to move away from traditional attack-focused and result-oriented evaluations towards a deeper examination of model behaviors. Our investigations highlight the significant differences in training dynamics between the two classes, emphasizing the need for future research to focus on robust modeling of the bonafide class.

In this paper, we propose novel quaternion activation functions where we modify either the quaternion magnitude or the phase, as an alternative to the commonly used split activation functions. We define criteria that are relevant for quaternion activation functions, and subsequently we propose our novel activation functions based on this analysis. Instead of applying a known activation function like the ReLU or Tanh on the quaternion elements separately, these activation functions consider the quaternion properties and respect the quaternion space $\mathbb{H}$. In particular, all quaternion components are utilized to calculate all output components, carrying out the benefit of the Hamilton product in e.g. the quaternion convolution to the activation functions. The proposed activation functions can be incorporated in arbitrary quaternion valued neural networks trained with gradient descent techniques. We further discuss the derivatives of the proposed activation functions where we observe beneficial properties for the activation functions affecting the phase. Specifically, they prove to be sensitive on basically the whole input range, thus improved gradient flow can be expected. We provide an elaborate experimental evaluation of our proposed quaternion activation functions including comparison with the split ReLU and split Tanh on two image classification tasks using the CIFAR-10 and SVHN dataset. There, especially the quaternion activation functions affecting the phase consistently prove to provide better performance.

Recently, machine unlearning, which seeks to erase specific data stored in the pre-trained or fine-tuned models, has emerged as a crucial protective measure for LLMs. However, unlearning approaches for LLMs that have been considered thus far have focused on the removal of independent data points and have not taken into account that the stored facts are logically connected to one another and form an implicit knowledge graph. To facilitate the development of structural unlearning methods, which are essential for the practical application of unlearning, we propose PISTOL, a pipeline for compiling multi-scenario datasets for benchmarking structural LLM unlearning. Additionally, leveraging sample datasets synthesized using PISTOL, we conducted benchmarks with four distinct unlearning methods on both Llama2-7B and Mistral-7B models. This analysis helps to illustrate the prevailing challenges in effectively and robustly removing highly inter-connected data, batched data, or data skewed towards a specific domain. It also highlights the choice of pre-trained model can impact unlearning performance. This work not only advances our understandings on the limitation of current LLMs unlearning methods and proposes future research directions, but also provides a replicable framework for ongoing exploration and validation in the field.

Deep learning has achieved impressive results in nuclei segmentation, but the massive requirement for pixel-wise labels remains a significant challenge. To alleviate the annotation burden, existing methods generate pseudo masks for model training using point labels. However, the generated masks are inevitably different from the ground truth, and these dissimilarities are not handled reasonably during the network training, resulting in the subpar performance of the segmentation model. To tackle this issue, we propose a framework named DoNuSeg, enabling \textbf{D}ynamic pseudo label \textbf{O}ptimization in point-supervised \textbf{Nu}clei \textbf{Seg}mentation. Specifically, DoNuSeg takes advantage of class activation maps (CAMs) to adaptively capture regions with semantics similar to annotated points. To leverage semantic diversity in the hierarchical feature levels, we design a dynamic selection module to choose the optimal one among CAMs from different encoder blocks as pseudo masks. Meanwhile, a CAM-guided contrastive module is proposed to further enhance the accuracy of pseudo masks. In addition to exploiting the semantic information provided by CAMs, we consider location priors inherent to point labels, developing a task-decoupled structure for effectively differentiating nuclei. Extensive experiments demonstrate that DoNuSeg outperforms state-of-the-art point-supervised methods. The code is available at https://github.com/shinning0821/MICCAI24-DoNuSeg.