We formally study the logical reasoning capabilities of decoder-only Transformers in the context of the boolean satisfiability (SAT) problem. First, we prove by construction that decoder-only Transformers can decide 3-SAT, in a non-uniform model of computation, using backtracking and deduction via Chain-of-Thought (CoT). %We prove its correctness by showing trace equivalence to the well-known DPLL SAT-solving algorithm. Second, we implement our construction as a PyTorch model with a tool (PARAT) that we designed to empirically demonstrate its correctness and investigate its properties. Third, rather than \textit{programming} a transformer to reason, we evaluate empirically whether it can be \textit{trained} to do so by learning directly from algorithmic traces (``reasoning paths'') from our theoretical construction. The trained models demonstrate strong out-of-distribution generalization on problem sizes seen during training but has limited length generalization, which is consistent with the implications of our theoretical result
We introduce a neural architecture finetuned for the task of scenario context generation: The relevant location and time of an event or entity mentioned in text. Contextualizing information extraction helps to scope the validity of automated finings when aggregating them as knowledge graphs. Our approach uses a high-quality curated dataset of time and location annotations in a corpus of epidemiology papers to train an encoder-decoder architecture. We also explored the use of data augmentation techniques during training. Our findings suggest that a relatively small fine-tuned encoder-decoder model performs better than out-of-the-box LLMs and semantic role labeling parsers to accurate predict the relevant scenario information of a particular entity or event.
Selective attention enables humans to efficiently process visual stimuli by enhancing important elements and filtering out irrelevant information. Locating visual attention is fundamental in neuroscience with potential applications in brain-computer interfaces. Conventional paradigms often use synthetic stimuli or static images, but visual stimuli in real life contain smooth and highly irregular dynamics. We show that these irregular dynamics can be decoded from electroencephalography (EEG) signals for selective visual attention decoding. To this end, we propose a free-viewing paradigm in which participants attend to one of two superimposed videos, each showing a center-aligned person performing a stage act. Superimposing ensures that the relative differences in the neural responses are not driven by differences in object locations. A stimulus-informed decoder is trained to extract EEG components correlated with the motion patterns of the attended object, and can detect the attended object in unseen data with significantly above-chance accuracy. This shows that the EEG responses to naturalistic motion are modulated by selective attention. Eye movements are also found to be correlated to the motion patterns in the attended video, despite the spatial overlap with the distractor. We further show that these eye movements do not dominantly drive the EEG-based decoding and that complementary information exists in EEG and gaze data. Moreover, our results indicate that EEG may also capture neural responses to unattended objects. To our knowledge, this study is the first to explore EEG-based selective visual attention decoding on natural videos, opening new possibilities for experiment design.
In this work, we explore the decoding of mental imagery from subjects using their fMRI measurements. In order to achieve this decoding, we first created a mapping between a subject's fMRI signals elicited by the videos the subjects watched. This mapping associates the high dimensional fMRI activation states with visual imagery. Next, we prompted the subjects textually, primarily with emotion labels which had no direct reference to visual objects. Then to decode visual imagery that may have been in a person's mind's eye, we align a latent representation of these fMRI measurements with a corresponding video-fMRI based on textual labels given to the videos themselves. This alignment has the effect of overlapping the video fMRI embedding with the text-prompted fMRI embedding, thus allowing us to use our fMRI-to-video mapping to decode. Additionally, we enhance an existing fMRI dataset, initially consisting of data from five subjects, by including recordings from three more subjects gathered by our team. We demonstrate the efficacy of our model on this augmented dataset both in accurately creating a mapping, as well as in plausibly decoding mental imagery.
Digital pathology has made significant advances in tumor diagnosis and segmentation, but image variability due to differences in organs, tissue preparation, and acquisition - known as domain shift - limits the effectiveness of current algorithms. The COSAS (Cross-Organ and Cross-Scanner Adenocarcinoma Segmentation) challenge addresses this issue by improving the resilience of segmentation algorithms to domain shift, with Task 2 focusing on adenocarcinoma segmentation using a diverse dataset from six scanners, pushing the boundaries of clinical diagnostics. Our approach employs unsupervised learning through stain separation within a multi-task learning framework using a multi-decoder autoencoder. This model isolates stain matrix and stain density, allowing it to handle color variation and improve generalization across scanners. We further enhanced the robustness of the model with a mixture of stain augmentation techniques and used a U-net architecture for segmentation. The novelty of our method lies in the use of stain separation within a multi-task learning framework, which effectively disentangles histological structures from color variations. This approach shows promise for improving segmentation accuracy and generalization across different histopathological stains, paving the way for more reliable diagnostic tools in digital pathology.
We present a novel approach to end-to-end automatic speech recognition (ASR) that utilizes pre-trained masked language models (LMs) to facilitate the extraction of linguistic information. The proposed models, BERT-CTC and BECTRA, are specifically designed to effectively integrate pre-trained LMs (e.g., BERT) into end-to-end ASR models. BERT-CTC adapts BERT for connectionist temporal classification (CTC) by addressing the constraint of the conditional independence assumption between output tokens. This enables explicit conditioning of BERT's contextualized embeddings in the ASR process, seamlessly merging audio and linguistic information through an iterative refinement algorithm. BECTRA extends BERT-CTC to the transducer framework and trains the decoder network using a vocabulary suitable for ASR training. This aims to bridge the gap between the text processed in end-to-end ASR and BERT, as these models have distinct vocabularies with varying text formats and styles, such as the presence of punctuation. Experimental results on various ASR tasks demonstrate that the proposed models improve over both the CTC and transducer-based baselines, owing to the incorporation of BERT knowledge. Moreover, our in-depth analysis and investigation verify the effectiveness of the proposed formulations and architectural designs.
This work proposes a frame-wise online/streaming end-to-end neural diarization (EEND) method, which detects speaker activities in a frame-in-frame-out fashion. The proposed model mainly consists of a causal embedding encoder and an online attractor decoder. Speakers are modeled in the self-attention-based decoder along both the time and speaker dimensions, and frame-wise speaker attractors are automatically generated and updated for new speakers and existing speakers, respectively. Retention mechanism is employed and especially adapted for long-form diarization with a linear temporal complexity. A multi-step progressive training strategy is proposed for gradually learning from easy tasks to hard tasks in terms of the number of speakers and audio length. Finally, the proposed model (referred to as long-form streaming EEND, LS-EEND) is able to perform streaming diarization for a high (up to 8) and flexible number speakers and very long (say one hour) audio recordings. Experiments on various simulated and real-world datasets show that: 1) when not using oracle speech activity information, the proposed model achieves new state-of-the-art online diarization error rate on all datasets, including CALLHOME (12.11%), DIHARD II (27.58%), DIHARD III (19.61%), and AMI (20.76%); 2) Due to the frame-in-frame-out processing fashion and the linear temporal complexity, the proposed model achieves several times lower real-time-factor than comparison online diarization models.
Addressing the incompleteness problem in knowledge graph remains a significant challenge. Current knowledge graph completion methods have their limitations. For example, ComDensE is prone to overfitting and suffers from the degradation with the increase of network depth while InteractE has the limitations in feature interaction and interpretability. To this end, we propose a new method called dynamic sparse embedding (DSparsE) for knowledge graph completion. The proposed model embeds the input entity-relation pairs by a shallow encoder composed of a dynamic layer and a relation-aware layer. Subsequently, the concatenated output of the dynamic layer and relation-aware layer is passed through a projection layer and a deep decoder with residual connection structure. This model ensures the network robustness and maintains the capability of feature extraction. Furthermore, the conventional dense layers are replaced by randomly initialized sparse connection layers in the proposed method, which can mitigate the model overfitting. Finally, comprehensive experiments are conducted on the datasets of FB15k-237, WN18RR and YAGO3-10. It was demonstrated that the proposed method achieves the state-of-the-art performance in terms of Hits@1 compared to the existing baseline approaches. An ablation study is performed to examine the effects of the dynamic layer and relation-aware layer, where the combined model achieves the best performance.
Large language models (LLMs) have demonstrated immense utility across various industries. However, as LLMs advance, the risk of harmful outputs increases due to incorrect or malicious instruction prompts. While current methods effectively address jailbreak risks, they share common limitations: 1) Judging harmful responses from the prefill-level lacks utilization of the model's decoding outputs, leading to relatively lower effectiveness and robustness. 2) Rejecting potentially harmful responses based on a single evaluation can significantly impair the model's helpfulness.This paper examines the LLMs' capability to recognize harmful outputs, revealing and quantifying their proficiency in assessing the danger of previous tokens. Motivated by pilot experiment results, we design a robust defense mechanism at the decoding level. Our novel decoder-oriented, step-by-step defense architecture corrects harmful queries directly rather than rejecting them outright. We introduce speculative decoding to enhance usability and facilitate deployment to boost secure decoding speed. Extensive experiments demonstrate that our approach improves model security without compromising reasoning speed. Notably, our method leverages the model's ability to discern hazardous information, maintaining its helpfulness compared to existing methods.
Multi-level sentence simplification generates simplified sentences with varying language proficiency levels. We propose Label Confidence Weighted Learning (LCWL), a novel approach that incorporates a label confidence weighting scheme in the training loss of the encoder-decoder model, setting it apart from existing confidence-weighting methods primarily designed for classification. Experimentation on English grade-level simplification dataset shows that LCWL outperforms state-of-the-art unsupervised baselines. Fine-tuning the LCWL model on in-domain data and combining with Symmetric Cross Entropy (SCE) consistently delivers better simplifications compared to strong supervised methods. Our results highlight the effectiveness of label confidence weighting techniques for text simplification tasks with encoder-decoder architectures.
Self-supervised learning has become a popular and effective approach for enhancing time series forecasting, enabling models to learn universal representations from unlabeled data. However, effectively capturing both the global sequence dependence and local detail features within time series data remains challenging. To address this, we propose a novel generative self-supervised method called TimeDART, denoting Diffusion Auto-regressive Transformer for Time series forecasting. In TimeDART, we treat time series patches as basic modeling units. Specifically, we employ an self-attention based Transformer encoder to model the dependencies of inter-patches. Additionally, we introduce diffusion and denoising mechanisms to capture the detail locality features of intra-patch. Notably, we design a cross-attention-based denoising decoder that allows for adjustable optimization difficulty in the self-supervised task, facilitating more effective self-supervised pre-training. Furthermore, the entire model is optimized in an auto-regressive manner to obtain transferable representations. Extensive experiments demonstrate that TimeDART achieves state-of-the-art fine-tuning performance compared to the most advanced competitive methods in forecasting tasks. Our code is publicly available at https://github.com/Melmaphother/TimeDART.
Cross-domain Recommendation systems leverage multi-domain user interactions to improve performance, especially in sparse data or new user scenarios. However, CDR faces challenges such as effectively capturing user preferences and avoiding negative transfer. To address these issues, we propose the Multi-view Disentangled and Adaptive Preference Learning (MDAP) framework. Our MDAP framework uses a multiview encoder to capture diverse user preferences. The framework includes a gated decoder that adaptively combines embeddings from different views to generate a comprehensive user representation. By disentangling representations and allowing adaptive feature selection, our model enhances adaptability and effectiveness. Extensive experiments on benchmark datasets demonstrate that our method significantly outperforms state-of-the-art CDR and single-domain models, providing more accurate recommendations and deeper insights into user behavior across different domains.
VRAM requirements for transformer models scale quadratically with context length due to the self-attention mechanism. In this paper we modify the decoder-only transformer, replacing self-attention with InAttention, which scales linearly with context length during inference by having tokens attend only to initial states. Benchmarking shows that InAttention significantly reduces VRAM usage during inference, enabling handling of long sequences on consumer GPUs. We corroborate that fine-tuning extends context length efficiently, improving performance on long sequences without high training costs. InAttention offers a scalable solution for long-range dependencies in transformer models, paving the way for further optimization.
The global shortage of healthcare workers has demanded the development of smart healthcare assistants, which can help monitor and alert healthcare workers when necessary. We examine the healthcare knowledge of existing Large Vision Language Models (LVLMs) via the Visual Question Answering (VQA) task in hospital settings through expert annotated open-ended questions. We introduce the Emergency Room Visual Question Answering (ERVQA) dataset, consisting of <image, question, answer> triplets covering diverse emergency room scenarios, a seminal benchmark for LVLMs. By developing a detailed error taxonomy and analyzing answer trends, we reveal the nuanced nature of the task. We benchmark state-of-the-art open-source and closed LVLMs using traditional and adapted VQA metrics: Entailment Score and CLIPScore Confidence. Analyzing errors across models, we infer trends based on properties like decoder type, model size, and in-context examples. Our findings suggest the ERVQA dataset presents a highly complex task, highlighting the need for specialized, domain-specific solutions.
The text-to-video (T2V) generation models, offering convenient visual creation, have recently garnered increasing attention. Despite their substantial potential, the generated videos may present artifacts, including structural implausibility, temporal inconsistency, and a lack of motion, often resulting in near-static video. In this work, we have identified a correlation between the disparity of temporal attention maps across different blocks and the occurrence of temporal inconsistencies. Additionally, we have observed that the energy contained within the temporal attention maps is directly related to the magnitude of motion amplitude in the generated videos. Based on these observations, we present BroadWay, a training-free method to improve the quality of text-to-video generation without introducing additional parameters, augmenting memory or sampling time. Specifically, BroadWay is composed of two principal components: 1) Temporal Self-Guidance improves the structural plausibility and temporal consistency of generated videos by reducing the disparity between the temporal attention maps across various decoder blocks. 2) Fourier-based Motion Enhancement enhances the magnitude and richness of motion by amplifying the energy of the map. Extensive experiments demonstrate that BroadWay significantly improves the quality of text-to-video generation with negligible additional cost.
Recent approaches attempt to adapt powerful interactive segmentation models, such as SAM, to interactive matting and fine-tune the models based on synthetic matting datasets. However, models trained on synthetic data fail to generalize to complex and occlusion scenes. We address this challenge by proposing a new matting dataset based on the COCO dataset, namely COCO-Matting. Specifically, the construction of our COCO-Matting includes accessory fusion and mask-to-matte, which selects real-world complex images from COCO and converts semantic segmentation masks to matting labels. The built COCO-Matting comprises an extensive collection of 38,251 human instance-level alpha mattes in complex natural scenarios. Furthermore, existing SAM-based matting methods extract intermediate features and masks from a frozen SAM and only train a lightweight matting decoder by end-to-end matting losses, which do not fully exploit the potential of the pre-trained SAM. Thus, we propose SEMat which revamps the network architecture and training objectives. For network architecture, the proposed feature-aligned transformer learns to extract fine-grained edge and transparency features. The proposed matte-aligned decoder aims to segment matting-specific objects and convert coarse masks into high-precision mattes. For training objectives, the proposed regularization and trimap loss aim to retain the prior from the pre-trained model and push the matting logits extracted from the mask decoder to contain trimap-based semantic information. Extensive experiments across seven diverse datasets demonstrate the superior performance of our method, proving its efficacy in interactive natural image matting. We open-source our code, models, and dataset at https://github.com/XiaRho/SEMat.
How well do text-only Large Language Models (LLMs) grasp the visual world? As LLMs are increasingly used in computer vision, addressing this question becomes both fundamental and pertinent. However, existing studies have primarily focused on limited scenarios, such as their ability to generate visual content or cluster multimodal data. To this end, we propose the Visual Text Representation Benchmark (ViTeRB) to isolate key properties that make language models well-aligned with the visual world. With this, we identify large-scale decoder-based LLMs as ideal candidates for representing text in vision-centric contexts, counter to the current practice of utilizing text encoders. Building on these findings, we propose ShareLock, an ultra-lightweight CLIP-like model. By leveraging precomputable frozen features from strong vision and language models, ShareLock achieves an impressive 51% accuracy on ImageNet despite utilizing just 563k image-caption pairs. Moreover, training requires only 1 GPU hour (or 10 hours including the precomputation of features) - orders of magnitude less than prior methods. Code will be released.
In this work we propose a novel joint training method for Visual Place Recognition (VPR), which simultaneously learns a global descriptor and a pair classifier for re-ranking. The pair classifier can predict whether a given pair of images are from the same place or not. The network only comprises Vision Transformer components for both the encoder and the pair classifier, and both components are trained using their respective class tokens. In existing VPR methods, typically the network is initialized using pre-trained weights from a generic image dataset such as ImageNet. In this work we propose an alternative pre-training strategy, by using Siamese Masked Image Modelling as a pre-training task. We propose a Place-aware image sampling procedure from a collection of large VPR datasets for pre-training our model, to learn visual features tuned specifically for VPR. By re-using the Mask Image Modelling encoder and decoder weights in the second stage of training, Pair-VPR can achieve state-of-the-art VPR performance across five benchmark datasets with a ViT-B encoder, along with further improvements in localization recall with larger encoders. The Pair-VPR website is: https://csiro-robotics.github.io/Pair-VPR.
Recent works have shown that prompting large language models with audio encodings can unlock speech recognition capabilities. However, existing techniques do not scale efficiently, especially while handling long form streaming audio inputs -- not only do they extrapolate poorly beyond the audio length seen during training, but they are also computationally inefficient due to the quadratic cost of attention. In this work, we introduce SpeechLLM-XL, a linear scaling decoder-only model for streaming speech recognition. We process audios in configurable chunks using limited attention window for reduced computation, and the text tokens for each audio chunk are generated auto-regressively until an EOS is predicted. During training, the transcript is segmented into chunks, using a CTC forced alignment estimated from encoder output. SpeechLLM-XL with 1.28 seconds chunk size achieves 2.7%/6.7% WER on LibriSpeech test clean/other, and it shows no quality degradation on long form utterances 10x longer than the training utterances.
We consider coverless steganography where a Large Language Model (LLM) drives an arithmetic coding decoder to generate stego-texts. An efficient method should embed secret message bits in as few language tokens as possible, while still keeping the stego-text natural and fluent. We show that on the individual token level, this problem is mathematically equivalent to maximizing the entropy of a replacement probability distribution of the next token generation, subject to a constraint on the KL divergence between the chosen probability distribution and the original distribution given by the LLM. A closed-form solution is provided for the optimization problem, which can be computed efficiently. Several important practical issues are also tackled: 1) An often-overlooked tokenization mismatch issue is resolved with a simple prompt selection approach, 2) The combination of the optimized distribution and the vocabulary truncation technique is considered, and 3) The combination of the optimized distribution with other sequence-level selection heuristics to further enhance the efficiency and reliability is studied.