It remains a significant challenge to compress images at ultra-low bitrate while achieving both semantic consistency and high perceptual quality. We propose a novel image compression framework, Semantically Disentangled Image Compression (SEDIC) in this paper. Our proposed SEDIC leverages large multimodal models (LMMs) to disentangle the image into several essential semantic information, including an extremely compressed reference image, overall and object-level text descriptions, and the semantic masks. A multi-stage semantic decoder is designed to progressively restore the transmitted reference image object-by-object, ultimately producing high-quality and perceptually consistent reconstructions. In each decoding stage, a pre-trained controllable diffusion model is utilized to restore the object details on the reference image conditioned by the text descriptions and semantic masks. Experimental results demonstrate that SEDIC significantly outperforms state-of-the-art approaches, achieving superior perceptual quality and semantic consistency at ultra-low bitrates ($\le$ 0.05 bpp). Our code is available at https://github.com/yang-xidian/SEDIC.
Despite recent progress, video generative models still struggle to animate human actions from static images, particularly when handling uncommon actions whose training data are limited. In this paper, we investigate the task of learning to animate human actions from a small number of videos -- 16 or fewer -- which is highly valuable in real-world applications like video and movie production. Few-shot learning of generalizable motion patterns while ensuring smooth transitions from the initial reference image is exceedingly challenging. We propose FLASH (Few-shot Learning to Animate and Steer Humans), which improves motion generalization by aligning motion features and inter-frame correspondence relations between videos that share the same motion but have different appearances. This approach minimizes overfitting to visual appearances in the limited training data and enhances the generalization of learned motion patterns. Additionally, FLASH extends the decoder with additional layers to compensate lost details in the latent space, fostering smooth transitions from the initial reference image. Experiments demonstrate that FLASH effectively animates images with unseen human or scene appearances into specified actions while maintaining smooth transitions from the reference image.
Pre-training and representation learning have been playing an increasingly important role in modern speech processing. Nevertheless, different applications have been relying on different foundation models, since predominant pre-training techniques are either designed for discriminative tasks or generative tasks. In this work, we make the first attempt at building a unified pre-training framework for both types of tasks in speech. We show that with the appropriate design choices for pre-training, one can jointly learn a representation encoder and generative audio decoder that can be applied to both types of tasks. We propose UniWav, an encoder-decoder framework designed to unify pre-training representation learning and generative tasks. On speech recognition, text-to-speech, and speech tokenization, UniWav achieves comparable performance to different existing foundation models, each trained on a specific task. Our findings suggest that a single general-purpose foundation model for speech can be built to replace different foundation models, reducing the overhead and cost of pre-training.
Most existing works in image caption synthesis use computation heavy deep neural networks and generates image descriptions in English language. This often restricts this important assistive tool for widespread use across language and accessibility barriers. This work presents AC-Lite, a computationally efficient model for image captioning in low-resource Assamese language. AC-Lite reduces computational requirements by replacing computation-heavy deep network components with lightweight alternatives. The AC-Lite model is designed through extensive ablation experiments with different image feature extractor networks and language decoders. A combination of ShuffleNetv2x1.5 with GRU based language decoder along with bilinear attention is found to provide the best performance with minimum compute. AC-Lite was observed to achieve an 82.3 CIDEr score on the COCO-AC dataset with 2.45 GFLOPs and 22.87M parameters.
In recent years, the rise of DNA data storage technology has brought significant attention to the challenge of correcting insertion, deletion, and substitution (IDS) errors. Among various coding methods for IDS correction, Varshamov-Tenengolts (VT) codes, primarily designed for single-error correction, have emerged as a central research focus. While existing decoding methods achieve high accuracy in correcting a single error, they often fail to correct multiple IDS errors. In this work, we observe that VT codes retain some capability for addressing multiple errors by introducing a transformer-based VT decoder (TVTD) along with symbol- and statistic-based codeword embedding. Experimental results demonstrate that the proposed TVTD achieves perfect correction of a single error. Furthermore, when decoding multiple errors across various codeword lengths, the bit error rate and frame error rate are significantly improved compared to existing hard decision and soft-in soft-out algorithms. Additionally, through model architecture optimization, the proposed method reduces time consumption by an order of magnitude compared to other soft decoders.
Modern automatic speech recognition (ASR) models, such as OpenAI's Whisper, rely on deep encoder-decoder architectures, and their encoders are a critical bottleneck for efficient deployment due to high computational intensity. We introduce LiteASR, a low-rank compression scheme for ASR encoders that significantly reduces inference costs while maintaining transcription accuracy. Our approach leverages the strong low-rank properties observed in intermediate activations: by applying principal component analysis (PCA) with a small calibration dataset, we approximate linear transformations with a chain of low-rank matrix multiplications, and further optimize self-attention to work in the reduced dimension. Evaluation results show that our method can compress Whisper large-v3's encoder size by over 50%, matching Whisper medium's size with better transcription accuracy, thereby establishing a new Pareto-optimal frontier of efficiency and performance. The code of LiteASR is available at https://github.com/efeslab/LiteASR.
We propose a novel family of language models, Latent-Thought Language Models (LTMs), which incorporate explicit latent thought vectors that follow an explicit prior model in latent space. These latent thought vectors guide the autoregressive generation of ground tokens through a Transformer decoder. Training employs a dual-rate optimization process within the classical variational Bayes framework: fast learning of local variational parameters for the posterior distribution of latent vectors, and slow learning of global decoder parameters. Empirical studies reveal that LTMs possess additional scaling dimensions beyond traditional LLMs, yielding a structured design space. Higher sample efficiency can be achieved by increasing training compute per token, with further gains possible by trading model size for more inference steps. Designed based on these scaling properties, LTMs demonstrate superior sample and parameter efficiency compared to conventional autoregressive models and discrete diffusion models. They significantly outperform these counterparts in validation perplexity and zero-shot language modeling. Additionally, LTMs exhibit emergent few-shot in-context reasoning capabilities that scale with model and latent size, and achieve competitive performance in conditional and unconditional text generation.
Our approach to training 3D vision-language understanding models is to train a feedforward model that makes predictions in 3D, but never requires 3D labels and is supervised only in 2D, using 2D losses and differentiable rendering. The approach is new for vision-language understanding. By treating the reconstruction as a ``latent variable'', we can render the outputs without placing unnecessary constraints on the network architecture (e.g. can be used with decoder-only models). For training, only need images and camera pose, and 2D labels. We show that we can even remove the need for 2D labels by using pseudo-labels from pretrained 2D models. We demonstrate this to pretrain a network, and we finetune it for 3D vision-language understanding tasks. We show this approach outperforms baselines/sota for 3D vision-language grounding, and also outperforms other 3D pretraining techniques. Project page: https://liftgs.github.io.
Significant advancements have been achieved in the realm of understanding poses and interactions of two hands manipulating an object. The emergence of augmented reality (AR) and virtual reality (VR) technologies has heightened the demand for real-time performance in these applications. However, current state-of-the-art models often exhibit promising results at the expense of substantial computational overhead. In this paper, we present a query-optimized real-time Transformer (QORT-Former), the first Transformer-based real-time framework for 3D pose estimation of two hands and an object. We first limit the number of queries and decoders to meet the efficiency requirement. Given limited number of queries and decoders, we propose to optimize queries which are taken as input to the Transformer decoder, to secure better accuracy: (1) we propose to divide queries into three types (a left hand query, a right hand query and an object query) and enhance query features (2) by using the contact information between hands and an object and (3) by using three-step update of enhanced image and query features with respect to one another. With proposed methods, we achieved real-time pose estimation performance using just 108 queries and 1 decoder (53.5 FPS on an RTX 3090TI GPU). Surpassing state-of-the-art results on the H2O dataset by 17.6% (left hand), 22.8% (right hand), and 27.2% (object), as well as on the FPHA dataset by 5.3% (right hand) and 10.4% (object), our method excels in accuracy. Additionally, it sets the state-of-the-art in interaction recognition, maintaining real-time efficiency with an off-the-shelf action recognition module.
Learning representations that are useful for unknown downstream tasks is a fundamental challenge in representation learning. Prominent approaches in this domain include contrastive learning, self-supervised masking, and denoising auto-encoders. In this paper, we introduce a novel method, termed contrastive Mutual Information Machine (cMIM), which aims to enhance the utility of learned representations for downstream tasks. cMIM integrates a new contrastive learning loss with the Mutual Information Machine (MIM) learning framework, a probabilistic auto-encoder that maximizes the mutual information between inputs and latent representations while clustering the latent codes. Despite MIM's potential, initial experiments indicated that the representations learned by MIM were less effective for discriminative downstream tasks compared to state-of-the-art (SOTA) models. The proposed cMIM method directly addresses this limitation. The main contributions of this work are twofold: (1) We propose a novel contrastive extension to MIM for learning discriminative representations which eliminates the need for data augmentation and is robust to variations in the number of negative examples (i.e., batch size). (2) We introduce a generic method for extracting informative embeddings from encoder-decoder models, which significantly improves performance in discriminative downstream tasks without requiring additional training. This method is applicable to any pre-trained encoder-decoder model. By presenting cMIM, we aim to offer a unified generative model that is effective for both generative and discriminative tasks. Our results demonstrate that the learned representations are valuable for downstream tasks while maintaining the generative capabilities of MIM.
Understanding the transformer architecture and its workings is essential for machine learning (ML) engineers. However, truly understanding the transformer architecture can be demanding, even if you have a solid background in machine learning or deep learning. The main working horse is attention, which yields to the transformer encoder-decoder structure. However, putting attention aside leaves several programming components that are easy to implement but whose role for the whole is unclear. These components are 'tokenization', 'embedding' ('un-embedding'), 'masking', 'positional encoding', and 'padding'. The focus of this work is on understanding them. To keep things simple, the understanding is built incrementally by adding components one by one, and after each step investigating what is doable and what is undoable with the current model. Simple sequences of zeros (0) and ones (1) are used to study the workings of each step.
We introduce CoT-ICL Lab, a framework and methodology to generate synthetic tokenized datasets and systematically study chain-of-thought (CoT) in-context learning (ICL) in language models. CoT-ICL Lab allows fine grained control over the complexity of in-context examples by decoupling (1) the causal structure involved in chain token generation from (2) the underlying token processing functions. We train decoder-only transformers (up to 700M parameters) on these datasets and show that CoT accelerates the accuracy transition to higher values across model sizes. In particular, we find that model depth is crucial for leveraging CoT with limited in-context examples, while more examples help shallow models match deeper model performance. Additionally, limiting the diversity of token processing functions throughout training improves causal structure learning via ICL. We also interpret these transitions by analyzing transformer embeddings and attention maps. Overall, CoT-ICL Lab serves as a simple yet powerful testbed for theoretical and empirical insights into ICL and CoT in language models.
Large Language Models (LLMs) have recently shown remarkable advancement in various NLP tasks. As such, a popular trend has emerged lately where NLP researchers extract word/sentence/document embeddings from these large decoder-only models and use them for various inference tasks with promising results. However, it is still unclear whether the performance improvement of LLM-induced embeddings is merely because of scale or whether underlying embeddings they produce significantly differ from classical encoding models like Word2Vec, GloVe, Sentence-BERT (SBERT) or Universal Sentence Encoder (USE). This is the central question we investigate in the paper by systematically comparing classical decontextualized and contextualized word embeddings with the same for LLM-induced embeddings. Our results show that LLMs cluster semantically related words more tightly and perform better on analogy tasks in decontextualized settings. However, in contextualized settings, classical models like SimCSE often outperform LLMs in sentence-level similarity assessment tasks, highlighting their continued relevance for fine-grained semantics.
Targeted protein degradation (TPD) induced by small molecules has emerged as a rapidly evolving modality in drug discovery, targeting proteins traditionally considered "undruggable". Proteolysis-targeting chimeras (PROTACs) and molecular glue degraders (MGDs) are the primary small molecules that induce TPD. Both types of molecules form a ternary complex linking an E3 ligase with a target protein, a crucial step for drug discovery. While significant advances have been made in binary structure prediction for proteins and small molecules, ternary structure prediction remains challenging due to obscure interaction mechanisms and insufficient training data. Traditional methods relying on manually assigned rules perform poorly and are computationally demanding due to extensive random sampling. In this work, we introduce DeepTernary, a novel deep learning-based approach that directly predicts ternary structures in an end-to-end manner using an encoder-decoder architecture. DeepTernary leverages an SE(3)-equivariant graph neural network (GNN) with both intra-graph and ternary inter-graph attention mechanisms to capture intricate ternary interactions from our collected high-quality training dataset, TernaryDB. The proposed query-based Pocket Points Decoder extracts the 3D structure of the final binding ternary complex from learned ternary embeddings, demonstrating state-of-the-art accuracy and speed in existing PROTAC benchmarks without prior knowledge from known PROTACs. It also achieves notable accuracy on the more challenging MGD benchmark under the blind docking protocol. Remarkably, our experiments reveal that the buried surface area calculated from predicted structures correlates with experimentally obtained degradation potency-related metrics. Consequently, DeepTernary shows potential in effectively assisting and accelerating the development of TPDs for previously undruggable targets.
Quantum error correction is crucial for large-scale quantum computing, but the absence of efficient decoders for new codes like quantum low-density parity-check (QLDPC) codes has hindered progress. Here we introduce a universal decoder based on linear attention sequence modeling and graph neural network that operates directly on any stabilizer code's graph structure. Our numerical experiments demonstrate that this decoder outperforms specialized algorithms in both accuracy and speed across diverse stabilizer codes, including surface codes, color codes, and QLDPC codes. The decoder maintains linear time scaling with syndrome measurements and requires no structural modifications between different codes. For the Bivariate Bicycle code with distance 12, our approach achieves a 39.4% lower logical error rate than previous best decoders while requiring only ~1% of the decoding time. These results provide a practical, universal solution for quantum error correction, eliminating the need for code-specific decoders.
Masked autoencoders (MAE) have shown tremendous potential for self-supervised learning (SSL) in vision and beyond. However, point clouds from LiDARs used in automated driving are particularly challenging for MAEs since large areas of the 3D volume are empty. Consequently, existing work suffers from leaking occupancy information into the decoder and has significant computational complexity, thereby limiting the SSL pre-training to only 2D bird's eye view encoders in practice. In this work, we propose the novel neighborhood occupancy MAE (NOMAE) that overcomes the aforementioned challenges by employing masked occupancy reconstruction only in the neighborhood of non-masked voxels. We incorporate voxel masking and occupancy reconstruction at multiple scales with our proposed hierarchical mask generation technique to capture features of objects of different sizes in the point cloud. NOMAEs are extremely flexible and can be directly employed for SSL in existing 3D architectures. We perform extensive evaluations on the nuScenes and Waymo Open datasets for the downstream perception tasks of semantic segmentation and 3D object detection, comparing with both discriminative and generative SSL methods. The results demonstrate that NOMAE sets the new state-of-the-art on multiple benchmarks for multiple point cloud perception tasks.
Providing effective feedback for programming assignments in computer science education can be challenging: students solve problems by iteratively submitting code, executing it, and using limited feedback from the compiler or the auto-grader to debug. Analyzing student debugging behavior in this process may reveal important insights into their knowledge and inform better personalized support tools. In this work, we propose an encoder-decoder-based model that learns meaningful code-edit embeddings between consecutive student code submissions, to capture their debugging behavior. Our model leverages information on whether a student code submission passes each test case to fine-tune large language models (LLMs) to learn code editing representations. It enables personalized next-step code suggestions that maintain the student's coding style while improving test case correctness. Our model also enables us to analyze student code-editing patterns to uncover common student errors and debugging behaviors, using clustering techniques. Experimental results on a real-world student code submission dataset demonstrate that our model excels at code reconstruction and personalized code suggestion while revealing interesting patterns in student debugging behavior.
Masked language modelling (MLM) as a pretraining objective has been widely adopted in genomic sequence modelling. While pretrained models can successfully serve as encoders for various downstream tasks, the distribution shift between pretraining and inference detrimentally impacts performance, as the pretraining task is to map [MASK] tokens to predictions, yet the [MASK] is absent during downstream applications. This means the encoder does not prioritize its encodings of non-[MASK] tokens, and expends parameters and compute on work only relevant to the MLM task, despite this being irrelevant at deployment time. In this work, we propose a modified encoder-decoder architecture based on the masked autoencoder framework, designed to address this inefficiency within a BERT-based transformer. We empirically show that the resulting mismatch is particularly detrimental in genomic pipelines where models are often used for feature extraction without fine-tuning. We evaluate our approach on the BIOSCAN-5M dataset, comprising over 2 million unique DNA barcodes. We achieve substantial performance gains in both closed-world and open-world classification tasks when compared against causal models and bidirectional architectures pretrained with MLM tasks.
In this study, we propose MHEX+, a framework adaptable to any U-Net architecture. Built upon MHEX+, we introduce novel U-Net variants, EU-Nets, which enhance explainability and uncertainty estimation, addressing the limitations of traditional U-Net models while improving performance and stability. A key innovation is the Equivalent Convolutional Kernel, which unifies consecutive convolutional layers, boosting interpretability. For uncertainty estimation, we propose the collaboration gradient approach, measuring gradient consistency across decoder layers. Notably, EU-Nets achieve an average accuracy improvement of 1.389\% and a variance reduction of 0.83\% across all networks and datasets in our experiments, requiring fewer than 0.1M parameters.
Unmanned Aerial Vehicles (UAVs) in Wireless Power Transfer (WPT)-assisted Internet of Things (IoT) systems face the following challenges: limited resources and suboptimal trajectory planning. Reinforcement learning-based trajectory planning schemes face issues of low search efficiency and learning instability when optimizing large-scale systems. To address these issues, we present an Attention-based UAV Trajectory Optimization (AUTO) framework based on the graph transformer, which consists of an Attention Trajectory Optimization Model (ATOM) and a Trajectory lEarNing Method based on Actor-critic (TENMA). In ATOM, a graph encoder is used to calculate the self-attention characteristics of all IoTDs, and a trajectory decoder is developed to optimize the number and trajectories of UAVs. TENMA then trains the ATOM using an improved Actor-Critic method, in which the real reward of the system is applied as the baseline to reduce variances in the critic network. This method is suitable for high-quality and large-scale multi-UAV trajectory planning. Finally, we develop numerous experiments, including a hardware experiment in the field case, to verify the feasibility and efficiency of the AUTO framework.