For the Transformer architecture [59], the attention mechanism constitutes the primary computational bottleneck, since computing the self-attention scores of queries and keys has quadratic scaling in the sequence length. Scaling attention to longer context will unlock new capabilities (modeling and reasoning over multiple long documents [24, 43, 50] and files in large codebases [30, 48]), new modalities (high-resolution images [11], audio [23], video [25]), and new applications (user interaction with long history [53], agent workflow with long horizon [62]). This has generated significant interest in making attention faster in the long-context regime, including by approximation [14, 27, 56] and software optimization ([17, 29, 45]), or even alternative architectures [22, 42, 55]. In this work, we build on the work of Dao et al. [17] on developing exact-attention algorithms that integrate knowledge of the GPU's execution model and hardware characteristics into their high-level design. In [17], Dao et al. introduced FlashAttention, a novel tiling strategy for parallelizing attention that eliminates intermediate reads/writes to slow global memory through fusing all of the attention operations into a single GPU kernel. Dao [15] restructured the algorithm as FlashAttention-2 to also parallelize over the sequence length dimension and perform the inner loop of the forward pass over blocks of the key and value matrices, thus improving the occupancy and distribution of work on the GPU. However, we observe that FlashAttention-2 nonetheless achieves poor utilization on newer GPUs relative to optimized matrix-multiplication (GEMM) kernels, such as 35% vs. 80-90% on the Hopper H100 GPU. Partially, this may be attributed to implementation-level differences, such as not using Hopper-specific instructions in place of Ampere ones when targeting the Tensor Cores. Several work such as ThunkerKitten [52] and cuDNN 9 [39] has shown that with Hopper-specific instructions and tile-based abstractions, one can speedup attention computation and simplify the implementation.

This paper introduces a scale-invariant methodology employing \textit{Fractal Geometry} to analyze and explain the nonlinear dynamics of complex connectionist systems. By leveraging architectural self-similarity in Deep Neural Networks (DNNs), we quantify fractal dimensions and \textit{roughness} to deeply understand their dynamics and enhance the quality of \textit{intrinsic} explanations. Our approach integrates principles from Chaos Theory to improve visualizations of fractal evolution and utilizes a Graph-Based Neural Network for reconstructing network topology. This strategy aims at advancing the \textit{intrinsic} explainability of connectionist Artificial Intelligence (AI) systems.

Discourse phenomena in existing document-level translation datasets are sparse, which has been a fundamental obstacle in the development of context-aware machine translation models. Moreover, most existing document-level corpora and context-aware machine translation methods rely on an unrealistic assumption on sentence-level alignments. To mitigate these issues, we first curate a novel dataset of Chinese-English literature, which consists of 160 books with intricate discourse structures. Then, we propose a more pragmatic and challenging setting for context-aware translation, termed chapter-to-chapter (Ch2Ch) translation, and investigate the performance of commonly-used machine translation models under this setting. Furthermore, we introduce a potential approach of finetuning large language models (LLMs) within the domain of Ch2Ch literary translation, yielding impressive improvements over baselines. Through our comprehensive analysis, we unveil that literary translation under the Ch2Ch setting is challenging in nature, with respect to both model learning methods and translation decoding algorithms.

This paper presents "Emotion Talk," a system designed to provide continuous emotional support through audio messages for psychological assistance. The primary objective is to offer consistent support to patients outside traditional therapy sessions by analyzing audio messages to detect emotions and generate appropriate responses. The solution focuses on Portuguese-speaking users, ensuring that the system is linguistically and culturally relevant. This system aims to complement and enhance the psychological follow-up process conducted by therapists, providing immediate and accessible assistance, especially in emergency situations where rapid response is crucial. Experimental results demonstrate the effectiveness of the proposed system, highlighting its potential in applications of psychological support.

In this work, we address the problem of eavesdropping on digital video displays by analyzing the electromagnetic waves that unintentionally emanate from the cables and connectors, particularly HDMI. This problem is known as TEMPEST. Compared to the analog case (VGA), the digital case is harder due to a 10-bit encoding that results in a much larger bandwidth and non-linear mapping between the observed signal and the pixel's intensity. As a result, eavesdropping systems designed for the analog case obtain unclear and difficult-to-read images when applied to digital video. The proposed solution is to recast the problem as an inverse problem and train a deep learning module to map the observed electromagnetic signal back to the displayed image. However, this approach still requires a detailed mathematical analysis of the signal, firstly to determine the frequency at which to tune but also to produce training samples without actually needing a real TEMPEST setup. This saves time and avoids the need to obtain these samples, especially if several configurations are being considered. Our focus is on improving the average Character Error Rate in text, and our system improves this rate by over 60 percentage points compared to previous available implementations. The proposed system is based on widely available Software Defined Radio and is fully open-source, seamlessly integrated into the popular GNU Radio framework. We also share the dataset we generated for training, which comprises both simulated and over 1000 real captures. Finally, we discuss some countermeasures to minimize the potential risk of being eavesdropped by systems designed based on similar principles.

Large language models with vision capabilities (VLMs), e.g., GPT-4o and Gemini 1.5 Pro are powering countless image-text applications and scoring high on many vision-understanding benchmarks. We propose BlindTest, a suite of 7 visual tasks absurdly easy to humans such as identifying (a) whether two circles overlap; (b) whether two lines intersect; (c) which letter is being circled in a word; and (d) counting the number of circles in a Olympic-like logo. Surprisingly, four state-of-the-art VLMs are, on average, only 56.20% accurate on our benchmark, with \newsonnet being the best (73.77% accuracy). On BlindTest, VLMs struggle with tasks that requires precise spatial information and counting (from 0 to 10), sometimes providing an impression of a person with myopia seeing fine details as blurry and making educated guesses. Code is available at: https://vlmsareblind.github.io/

We propose CoNSAL (Combining Neural networks and Symbolic regression for Analytical Lyapunov function) to construct analytical Lyapunov functions for nonlinear dynamic systems. This framework contains a neural Lyapunov function and a symbolic regression component, where symbolic regression is applied to distill the neural network to precise analytical forms. Our approach utilizes symbolic regression not only as a tool for translation but also as a means to uncover counterexamples. This procedure terminates when no counterexamples are found in the analytical formulation. Compared with previous results, CoNSAL directly produces an analytical form of the Lyapunov function with improved interpretability in both the learning process and the final results. We apply CoNSAL to 2-D inverted pendulum, path following, Van Der Pol Oscillator, 3-D trig dynamics, 4-D rotating wheel pendulum, 6-D 3-bus power system, and demonstrate that our algorithm successfully finds their valid Lyapunov functions. Code examples are available at https://github.com/HaohanZou/CoNSAL.

In this paper, we revisit SAT encodings of the partial-ordering based ILP model for the graph coloring problem (GCP) and suggest a generalization for the bandwidth coloring problem (BCP). The GCP asks for the minimum number of colors that can be assigned to the vertices of a given graph such that each two adjacent vertices get different colors. The BCP is a generalization, where each edge has a weight that enforces a minimal'distance' between the assigned colors, and the goal is to minimize the'largest' color used. For the widely studied GCP, we experimentally compare the partial-ordering based SAT encoding to the state-of-the-art approaches on the DIMACS benchmark set. Our evaluation confirms that this SAT encoding is effective for sparse graphs and even outperforms the state-of-the-art on some DIMACS instances. For the BCP, our theoretical analysis shows that the partial-ordering based SAT and ILP formulations have an asymptotically smaller size than that of the classical assignment-based model. Our practical evaluation confirms not only a dominance compared to the assignment-based encodings but also to the state-of-the-art approaches on a set of benchmark instances. Up to our knowledge, we have solved several open instances of the BCP from the literature for the first time.

Information extraction from handwritten documents involves traditionally three distinct steps: Document Layout Analysis, Handwritten Text Recognition, and Named Entity Recognition. Recent approaches have attempted to integrate these steps into a single process using fully end-to-end architectures. Despite this, these integrated approaches have not yet matched the performance of language models, when applied to information extraction in plain text. In this paper, we introduce DANIEL (Document Attention Network for Information Extraction and Labelling), a fully end-to-end architecture integrating a language model and designed for comprehensive handwritten document understanding. DANIEL performs layout recognition, handwriting recognition, and named entity recognition on full-page documents. Moreover, it can simultaneously learn across multiple languages, layouts, and tasks. For named entity recognition, the ontology to be applied can be specified via the input prompt. The architecture employs a convolutional encoder capable of processing images of any size without resizing, paired with an autoregressive decoder based on a transformer-based language model. DANIEL achieves competitive results on four datasets, including a new state-of-the-art performance on RIMES 2009 and M-POPP for Handwriting Text Recognition, and IAM NER for Named Entity Recognition. Furthermore, DANIEL is much faster than existing approaches. We provide the source code and the weights of the trained models at \url{https://github.com/Shulk97/daniel}.

Large language models (LLMs) have shown remarkable capabilities, but still struggle with processing extensive contexts, limiting their ability to maintain coherence and accuracy over long sequences. In this work, we introduce EM-LLM, a novel approach that integrates key aspects of human episodic memory and event cognition into LLMs, enabling them to effectively handle practically infinite context lengths while maintaining computational efficiency. When needed, these events are retrieved through a two-stage memory process, combining similarity-based and temporally contiguous retrieval for efficient and human-like access to relevant information. Experiments on the LongBench dataset demonstrate EM-LLM's superior performance, outperforming the state-of-the-art InfLLM model with an overall relative improvement of 4.3% across various tasks, including a 33% improvement on the PassageRetrieval task. Furthermore, our analysis reveals strong correlations between EM-LLM's event segmentation and human-perceived events, suggesting a bridge between this artificial system and its biological counterpart. This work not only advances LLM capabilities in processing extended contexts but also provides a computational framework for exploring human memory mechanisms, opening new avenues for interdisciplinary research in AI and cognitive science. For contemporary pre-trained large language models (LLMs), the context window serves as the primary mechanism to incorporate domain-specific, private, or common up-to-date information. These limitations stem from inherent challenges in Transformer-based architectures. Recent studies have shown that Transformers struggle with extrapolating to contexts longer than their training window size (Kazemnejad et al., 2024). On top of this, employing softmax attention over extended token sequences requires substantial computational resources for each token generation, and the resulting attention embeddings risk becoming excessively noisy and losing their distinctiveness (Tworkowski et al., 2023). To mitigate those challenges, recent works have focused on retrieval-based methods, either in the form of in-context augmentation (e.g., RAG-based techniques (Lewis et al., 2020; Gao et al., 2024)) or via retrieval of previously-inferred key-value pairs (KV) within individual attention heads (Wu et al., 2022; Tworkowski et al., 2023; Bertsch et al., 2023).