Online Handwritten Text Recognition (OLHTR) has gained considerable attention for its diverse range of applications. Current approaches usually treat OLHTR as a sequence recognition task, employing either a single trajectory or image encoder, or multi-stream encoders, combined with a CTC or attention-based recognition decoder. However, these approaches face several drawbacks: 1) single encoders typically focus on either local trajectories or visual regions, lacking the ability to dynamically capture relevant global features in challenging cases; 2) multi-stream encoders, while more comprehensive, suffer from complex structures and increased inference costs. To tackle this, we propose a Collaborative learning-based OLHTR framework, called Col-OLHTR, that learns multimodal features during training while maintaining a single-stream inference process. Col-OLHTR consists of a trajectory encoder, a Point-to-Spatial Alignment (P2SA) module, and an attention-based decoder. The P2SA module is designed to learn image-level spatial features through trajectory-encoded features and 2D rotary position embeddings. During training, an additional image-stream encoder-decoder is collaboratively trained to provide supervision for P2SA features. At inference, the extra streams are discarded, and only the P2SA module is used and merged before the decoder, simplifying the process while preserving high performance. Extensive experimental results on several OLHTR benchmarks demonstrate the state-of-the-art (SOTA) performance, proving the effectiveness and robustness of our design.
In image processing, solving inverse problems is the task of finding plausible reconstructions of an image that was corrupted by some (usually known) degradation operator. Commonly, this process is done using a generative image model that can guide the reconstruction towards solutions that appear natural. The success of diffusion models over the last few years has made them a leading candidate for this task. However, the sequential nature of diffusion models makes this conditional sampling process challenging. Furthermore, since diffusion models are often defined in the latent space of an autoencoder, the encoder-decoder transformations introduce additional difficulties. To address these challenges, we suggest a novel sampling method based on sequential Monte Carlo (SMC) in the latent space of diffusion models. We name our method LD-SMC. We define a generative model for the data using additional auxiliary observations and perform posterior inference with SMC sampling based on a backward diffusion process. Empirical evaluations on ImageNet and FFHQ show the benefits of LD-SMC over competing methods in various inverse problem tasks and especially in challenging inpainting tasks.
Recently, large language model (LLM) based text-to-speech (TTS) systems have gradually become the mainstream in the industry due to their high naturalness and powerful zero-shot voice cloning capabilities.Here, we introduce the IndexTTS system, which is mainly based on the XTTS and Tortoise model. We add some novel improvements. Specifically, in Chinese scenarios, we adopt a hybrid modeling method that combines characters and pinyin, making the pronunciations of polyphonic characters and long-tail characters controllable. We also performed a comparative analysis of the Vector Quantization (VQ) with Finite-Scalar Quantization (FSQ) for codebook utilization of acoustic speech tokens. To further enhance the effect and stability of voice cloning, we introduce a conformer-based speech conditional encoder and replace the speechcode decoder with BigVGAN2. Compared with XTTS, it has achieved significant improvements in naturalness, content consistency, and zero-shot voice cloning. As for the popular TTS systems in the open-source, such as Fish-Speech, CosyVoice2, FireRedTTS and F5-TTS, IndexTTS has a relatively simple training process, more controllable usage, and faster inference speed. Moreover, its performance surpasses that of these systems. Our demos are available at https://index-tts.github.io.
While autoregressive speech token generation models produce speech with remarkable variety and naturalness, their inherent lack of controllability often results in issues such as hallucinations and undesired vocalizations that do not conform to conditioning inputs. We introduce Koel-TTS, a suite of enhanced encoder-decoder Transformer TTS models that address these challenges by incorporating preference alignment techniques guided by automatic speech recognition and speaker verification models. Additionally, we incorporate classifier-free guidance to further improve synthesis adherence to the transcript and reference speaker audio. Our experiments demonstrate that these optimizations significantly enhance target speaker similarity, intelligibility, and naturalness of synthesized speech. Notably, Koel-TTS directly maps text and context audio to acoustic tokens, and on the aforementioned metrics, outperforms state-of-the-art TTS models, despite being trained on a significantly smaller dataset. Audio samples and demos are available on our website.
Modern systems for automatic speech recognition, including the RNN-Transducer and Attention-based Encoder-Decoder (AED), are designed so that the encoder is not required to alter the time-position of information from the audio sequence into the embedding; alignment to the final text output is processed during decoding. We discover that the transformer-based encoder adopted in recent years is actually capable of performing the alignment internally during the forward pass, prior to decoding. This new phenomenon enables a simpler and more efficient model, the "Aligner-Encoder". To train it, we discard the dynamic programming of RNN-T in favor of the frame-wise cross-entropy loss of AED, while the decoder employs the lighter text-only recurrence of RNN-T without learned cross-attention -- it simply scans embedding frames in order from the beginning, producing one token each until predicting the end-of-message. We conduct experiments demonstrating performance remarkably close to the state of the art, including a special inference configuration enabling long-form recognition. In a representative comparison, we measure the total inference time for our model to be 2x faster than RNN-T and 16x faster than AED. Lastly, we find that the audio-text alignment is clearly visible in the self-attention weights of a certain layer, which could be said to perform "self-transduction".
Motivated by emerging decentralized applications, the \emph{game of coding} framework has been recently introduced to address scenarios where the adversary's control over coded symbols surpasses the fundamental limits of traditional coding theory. Still, the reward mechanism available in decentralized systems, motivates the adversary to act rationally. While the decoder, as the data collector (DC), has an acceptance and rejection mechanism, followed by an estimation module, the adversary aims to maximize its utility, as an increasing function of (1) the chance of acceptance (to increase the reward), and (2) estimation error. On the other hand, the decoder also adjusts its acceptance rule to maximize its own utility, as (1) an increasing function of the chance of acceptance (to keep the system functional), (2) decreasing function of the estimation error. Prior works within this framework rely on the assumption that the game is complete, that is, both the DC and the adversary are fully aware of each other's utility functions. However, in practice, the decoder is often unaware of the utility of the adversary. To address this limitation, we develop an algorithm enabling the DC to commit to a strategy that achieves within the vicinity of the equilibrium, without knowledge of the adversary's utility function. Our approach builds on an observation that at the equilibrium, the relationship between the probability of acceptance and the mean squared error (MSE) follows a predetermined curve independent of the specific utility functions of the players. By exploiting this invariant relationship, the DC can iteratively refine its strategy based on observable parameters, converging to a near-optimal solution. We provide theoretical guarantees on sample complexity and accuracy of the proposed scheme.
MedSAM, a medical foundation model derived from the SAM architecture, has demonstrated notable success across diverse medical domains. However, its clinical application faces two major challenges: the dependency on labor-intensive manual prompt generation, which imposes a significant burden on clinicians, and the absence of semantic labeling in the generated segmentation masks for organs or lesions, limiting its practicality for non-expert users. To address these limitations, we propose AutoMedSAM, an end-to-end framework derived from SAM, designed to enhance usability and segmentation performance. AutoMedSAM retains MedSAM's image encoder and mask decoder structure while introducing a novel diffusion-based class prompt encoder. The diffusion-based encoder employs a dual-decoder structure to collaboratively generate prompt embeddings guided by sparse and dense prompt definitions. These embeddings enhance the model's ability to understand and process clinical imagery autonomously. With this encoder, AutoMedSAM leverages class prompts to embed semantic information into the model's predictions, transforming MedSAM's semi-automated pipeline into a fully automated workflow. Furthermore, AutoMedSAM employs an uncertainty-aware joint optimization strategy during training to effectively inherit MedSAM's pre-trained knowledge while improving generalization by integrating multiple loss functions. Experimental results across diverse datasets demonstrate that AutoMedSAM achieves superior performance while broadening its applicability to both clinical settings and non-expert users. Code is available at https://github.com/HP-ML/AutoPromptMedSAM.git.
Prompt caching in large language models (LLMs) results in data-dependent timing variations: cached prompts are processed faster than non-cached prompts. These timing differences introduce the risk of side-channel timing attacks. For example, if the cache is shared across users, an attacker could identify cached prompts from fast API response times to learn information about other users' prompts. Because prompt caching may cause privacy leakage, transparency around the caching policies of API providers is important. To this end, we develop and conduct statistical audits to detect prompt caching in real-world LLM API providers. We detect global cache sharing across users in seven API providers, including OpenAI, resulting in potential privacy leakage about users' prompts. Timing variations due to prompt caching can also result in leakage of information about model architecture. Namely, we find evidence that OpenAI's embedding model is a decoder-only Transformer, which was previously not publicly known.
Large Language Models (LLMs) are discovered to suffer from accurately retrieving key information. To address this, we propose Mask-Enhanced Autoregressive Prediction (MEAP), a simple yet effective training paradigm that seamlessly integrates Masked Language Modeling (MLM) into Next-Token Prediction (NTP) to enhance the latter's in-context retrieval capabilities. Specifically, MEAP first randomly masks a small fraction of input tokens and then directly performs the standard next-token prediction autoregressive using a decoder-only Transformer. MEAP eliminates the need for bidirectional attention or encoder-decoder architectures for MLM, incurring no additional computational overhead during pre-training or inference. Intensive experiments demonstrate that MEAP substantially outperforms NTP on key information retrieval and long-context reasoning tasks, while performing on par or better on commonsense reasoning tasks. The benefits of MEAP also extend to supervised fine-tuning, where it shows remarkable advantages in lost-in-the-middle scenarios, outperforming NTP by 11.77 percentage points. Our analysis indicates that MEAP's effectiveness arises from its ability to promote more distinguishable attention scores by concentrating on a reduced set of non-masked tokens. This mechanism improves the model's focus on task-relevant signals while mitigating the influence of peripheral context. These findings position MEAP as a promising training paradigm for large language models.
Recent advances in large language models (LLMs) have shown remarkable performance across diverse tasks. However, these models are typically deployed with fixed weights, which limits their ability to adapt dynamically to the variability inherent in real-world data during inference. This paper introduces ChamaleonLLM, a novel framework that enables inference-time adaptation of LLMs by leveraging batch-aware clustering and on-the-fly generation of low-rank updates. Unlike traditional fine-tuning approaches such as Low-Rank Adaptation (LoRA) or methods that rely on a fixed set of pre-learned uniforms (changeable masks), our method dynamically generates adaptive modifications to the decoder weights based on the aggregated statistics of clustered batches. By intelligently grouping similar inputs and computing context-aware low-rank updates via a hyper-network, ChamaleonLLM achieves significant performance gains, outperforming conventional LoRA methods while eliminating the overhead of maintaining multiple expert models. Our experiments highlight the potential of our approach to serve as a versatile and highly adaptive solution for language model inference. ChamaleonLLM is open-sourced to ensure the reproducibility of our experiments: https://anonymous.4open.science/r/ChamaleonLLM/
This paper introduces a novel approach to efficiently feeding knowledge to language models (LLMs) during prediction by integrating retrieval and generation processes within a unified framework. While the Retrieval-Augmented Generation (RAG) model addresses gaps in LLMs' training data and knowledge limits, it is hindered by token limit restrictions and dependency on the retrieval system's accuracy. Our proposed architecture incorporates in-context vectors (ICV) to overcome these challenges. ICV recasts in-context learning by using latent embeddings of LLMs to create a vector that captures essential task information. This vector is then used to shift the latent states of the LLM, enhancing the generation process without adding demonstration examples to the prompt. ICV directly integrates information into the model, enabling it to process this information more effectively. Our extensive experimental evaluation demonstrates that ICV outperforms standard in-context learning and fine-tuning across question-answering, information retrieval, and other tasks. This approach mitigates the limitations of current RAG models and offers a more robust solution for handling extensive and diverse datasets. Despite leveraging a fraction of the parameters, our ICV-enhanced model achieves competitive performance against models like LLaMA-3, Gemma, and Phi-3, significantly reducing computational costs and memory requirements. ICV reduces prompt length, is easy to control, surpasses token limitations, and is computationally efficient compared to fine-tuning.
3D instance segmentation aims to predict a set of object instances in a scene and represent them as binary foreground masks with corresponding semantic labels. Currently, transformer-based methods are gaining increasing attention due to their elegant pipelines, reduced manual selection of geometric properties, and superior performance. However, transformer-based methods fail to simultaneously maintain strong position and content information during query initialization. Additionally, due to supervision at each decoder layer, there exists a phenomenon of object disappearance with the deepening of layers. To overcome these hurdles, we introduce Beyond the Final Layer: Hierarchical Query Fusion Transformer with Agent-Interpolation Initialization for 3D Instance Segmentation (BFL). Specifically, an Agent-Interpolation Initialization Module is designed to generate resilient queries capable of achieving a balance between foreground coverage and content learning. Additionally, a Hierarchical Query Fusion Decoder is designed to retain low overlap queries, mitigating the decrease in recall with the deepening of layers. Extensive experiments on ScanNetV2, ScanNet200, ScanNet++ and S3DIS datasets demonstrate the superior performance of BFL.
This paper presents LP-DETR (Layer-wise Progressive DETR), a novel approach that enhances DETR-based object detection through multi-scale relation modeling. Our method introduces learnable spatial relationships between object queries through a relation-aware self-attention mechanism, which adaptively learns to balance different scales of relations (local, medium and global) across decoder layers. This progressive design enables the model to effectively capture evolving spatial dependencies throughout the detection pipeline. Extensive experiments on COCO 2017 dataset demonstrate that our method improves both convergence speed and detection accuracy compared to standard self-attention module. The proposed method achieves competitive results, reaching 52.3\% AP with 12 epochs and 52.5\% AP with 24 epochs using ResNet-50 backbone, and further improving to 58.0\% AP with Swin-L backbone. Furthermore, our analysis reveals an interesting pattern: the model naturally learns to prioritize local spatial relations in early decoder layers while gradually shifting attention to broader contexts in deeper layers, providing valuable insights for future research in object detection.
Groundbreaking advancements in text-to-image generation have recently been achieved with the emergence of diffusion models. These models exhibit a remarkable ability to generate highly artistic and intricately detailed images based on textual prompts. However, obtaining desired generation outcomes often necessitates repetitive trials of manipulating text prompts just like casting spells on a magic mirror, and the reason behind that is the limited capability of semantic understanding inherent in current image generation models. Specifically, existing diffusion models encode the text prompt input with a pre-trained encoder structure, which is usually trained on a limited number of image-caption pairs. The state-of-the-art large language models (LLMs) based on the decoder-only structure have shown a powerful semantic understanding capability as their architectures are more suitable for training on very large-scale unlabeled data. In this work, we propose to enhance text-to-image diffusion models by borrowing the strength of semantic understanding from large language models, and devise a simple yet effective adapter to allow the diffusion models to be compatible with the decoder-only structure. Meanwhile, we also provide a supporting theoretical analysis with various architectures (e.g., encoder-only, encoder-decoder, and decoder-only), and conduct extensive empirical evaluations to verify its effectiveness. The experimental results show that the enhanced models with our adapter module are superior to the stat-of-the-art models in terms of text-to-image generation quality and reliability.
Multi-Object Tracking (MOT) is a critical problem in computer vision, essential for understanding how objects move and interact in videos. This field faces significant challenges such as occlusions and complex environmental dynamics, impacting model accuracy and efficiency. While traditional approaches have relied on Convolutional Neural Networks (CNNs), introducing transformers has brought substantial advancements. This work introduces OneTrack-M, a transformer-based MOT model designed to enhance tracking computational efficiency and accuracy. Our approach simplifies the typical transformer-based architecture by eliminating the need for a decoder model for object detection and tracking. Instead, the encoder alone serves as the backbone for temporal data interpretation, significantly reducing processing time and increasing inference speed. Additionally, we employ innovative data pre-processing and multitask training techniques to address occlusion and diverse objective challenges within a single set of weights. Experimental results demonstrate that OneTrack-M achieves at least 25% faster inference times compared to state-of-the-art models in the literature while maintaining or improving tracking accuracy metrics. These improvements highlight the potential of the proposed solution for real-time applications such as autonomous vehicles, surveillance systems, and robotics, where rapid responses are crucial for system effectiveness.
Neural codecs, comprising an encoder, quantizer, and decoder, enable signal transmission at exceptionally low bitrates. Training these systems requires techniques like the straight-through estimator, soft-to-hard annealing, or statistical quantizer emulation to allow a non-zero gradient across the quantizer. Evaluating the effect of quantization in neural codecs, like the influence of gradient passing techniques on the whole system, is often costly and time-consuming due to training demands and the lack of affordable and reliable metrics. This paper proposes an efficient evaluation framework for neural codecs using simulated data with a defined number of bits and low-complexity neural encoders/decoders to emulate the non-linear behavior in larger networks. Our system is highly efficient in terms of training time and computational and hardware requirements, allowing us to uncover distinct behaviors in neural codecs. We propose a modification to stabilize training with the straight-through estimator based on our findings. We validate our findings against an internal neural audio codec and against the state-of-the-art descript-audio-codec.
We introduce Hibiki, a decoder-only model for simultaneous speech translation. Hibiki leverages a multistream language model to synchronously process source and target speech, and jointly produces text and audio tokens to perform speech-to-text and speech-to-speech translation. We furthermore address the fundamental challenge of simultaneous interpretation, which unlike its consecutive counterpart, where one waits for the end of the source utterance to start translating, adapts its flow to accumulate just enough context to produce a correct translation in real-time, chunk by chunk. To do so, we introduce a weakly-supervised method that leverages the perplexity of an off-the-shelf text translation system to identify optimal delays on a per-word basis and create aligned synthetic data. After supervised training, Hibiki performs adaptive, simultaneous speech translation with vanilla temperature sampling. On a French-English simultaneous speech translation task, Hibiki demonstrates state-of-the-art performance in translation quality, speaker fidelity and naturalness. Moreover, the simplicity of its inference process makes it compatible with batched translation and even real-time on-device deployment. We provide examples as well as models and inference code.
Consumer-grade depth sensors provide low-resolution depth maps; however, a high-resolution RGB camera is usually mounted on the same device and acquires a high-resolution image of the same scene. While deep learning and guided filtering methods gave decent results, recent works have highlighted the superiority of using RGB images for Depth Super Resolution. This paper proposes CoReGAN, a generative data fusion model that employs contrastive learning to regularize the extracted features of 2 independent encoders and 1 decoder for Guided Depth Super Resolution, demonstrating state-of-the-art results.
Guessing random additive noise decoding (GRAND) is a code-agnostic decoding method that iteratively guesses the noise pattern affecting the received codeword. The number of noise sequences to test depends on the noise realization. Thus, GRAND exhibits random runtime which results in nondeterministic throughput. However, real-time systems must process the incoming data at a fixed rate, necessitating a fixed-throughput decoder in order to avoid losing data. We propose a first-in first-out (FIFO) scheduling architecture that enables a fixed throughput while improving the block error rate (BLER) compared to the common approach of imposing a maximum runtime constraint per received codeword. Moreover, we demonstrate that the average throughput metric of GRAND-based hardware implementations typically provided in the literature can be misleading as one needs to operate at approximately one order of magnitude lower throughput to achieve the BLER of an unconstrained decoder.
Similarity between training examples is used to curate pretraining datasets for language models by many methods -- for diversification and to select examples similar to high-quality data. However, similarity is typically measured with off-the-shelf embedding models that are generic or trained for tasks such as retrieval. This paper introduces a framework to analyze the suitability of embedding models specifically for data curation in the language model pretraining setting. We quantify the correlation between similarity in the embedding space to similarity in pretraining loss between different training examples, and how diversifying in the embedding space affects pretraining quality. We analyze a variety of embedding models in our framework, with experiments using the Pile dataset for pretraining a 1.7B parameter decoder-only language model. We find that the embedding models we consider are all useful for pretraining data curation. Moreover, a simple approach of averaging per-token embeddings proves to be surprisingly competitive with more sophisticated embedding models -- likely because the latter are not designed specifically for pretraining data curation. Indeed, we believe our analysis and evaluation framework can serve as a foundation for the design of embedding models that specifically reason about similarity in pretraining datasets.