Lyric-to-melody generation is a highly challenging task in the field of AI music generation. Due to the difficulty of learning strict yet weak correlations between lyrics and melodies, previous methods have suffered from weak controllability, low-quality and poorly structured generation. To address these challenges, we propose CSL-L2M, a controllable song-level lyric-to-melody generation method based on an in-attention Transformer decoder with fine-grained lyric and musical controls, which is able to generate full-song melodies matched with the given lyrics and user-specified musical attributes. Specifically, we first introduce REMI-Aligned, a novel music representation that incorporates strict syllable- and sentence-level alignments between lyrics and melodies, facilitating precise alignment modeling. Subsequently, sentence-level semantic lyric embeddings independently extracted from a sentence-wise Transformer encoder are combined with word-level part-of-speech embeddings and syllable-level tone embeddings as fine-grained controls to enhance the controllability of lyrics over melody generation. Then we introduce human-labeled musical tags, sentence-level statistical musical attributes, and learned musical features extracted from a pre-trained VQ-VAE as coarse-grained, fine-grained and high-fidelity controls, respectively, to the generation process, thereby enabling user control over melody generation. Finally, an in-attention Transformer decoder technique is leveraged to exert fine-grained control over the full-song melody generation with the aforementioned lyric and musical conditions. Experimental results demonstrate that our proposed CSL-L2M outperforms the state-of-the-art models, generating melodies with higher quality, better controllability and enhanced structure. Demos and source code are available at https://lichaiustc.github.io/CSL-L2M/.
Blind estimation of intersymbol interference channels based on the Baum-Welch (BW) algorithm, a specific implementation of the expectation-maximization (EM) algorithm for training hidden Markov models, is robust and does not require labeled data. However, it is known for its extensive computation cost, slow convergence, and frequently converges to a local maximum. In this paper, we modified the trellis structure of the BW algorithm by associating the channel parameters with two consecutive states. This modification enables us to reduce the number of required states by half while maintaining the same performance. Moreover, to improve the convergence rate and the estimation performance, we construct a joint turbo-BW-equalization system by exploiting the extrinsic information produced by the turbo decoder to refine the BW-based estimator at each EM iteration. Our experiments demonstrate that the joint system achieves convergence in 10 EM iterations, which is 8 iterations less than a separate system design for a signal-to-noise ratio (SNR) of 4dB. Additionally, the joint system provides improved estimation accuracy with a mean square error (MSE) of $10^{-4}$ for an SNR of 6dB. We also identify scenarios where a joint design is not preferable, especially when the channel is noisy (e.g., SNR=2dB) and the decoder cannot provide reliable extrinsic information for a BW-based estimator.
Interpretability methods seek to understand language model representations, yet the outputs of most such methods -- circuits, vectors, scalars -- are not immediately human-interpretable. In response, we introduce LatentQA, the task of answering open-ended questions about model activations in natural language. Towards solving LatentQA, we propose Latent Interpretation Tuning (LIT), which finetunes a decoder LLM on a dataset of activations and associated question-answer pairs, similar to how visual instruction tuning trains on question-answer pairs associated with images. We use the decoder for diverse reading applications, such as extracting relational knowledge from representations or uncovering system prompts governing model behavior. Our decoder also specifies a differentiable loss that we use to control models, such as debiasing models on stereotyped sentences and controlling the sentiment of generations. Finally, we extend LatentQA to reveal harmful model capabilities, such as generating recipes for bioweapons and code for hacking.
Rapid advancements in video diffusion models have enabled the creation of realistic videos, raising concerns about unauthorized use and driving the demand for techniques to protect model ownership. Existing watermarking methods, while effective for image diffusion models, do not account for temporal consistency, leading to degraded video quality and reduced robustness against video distortions. To address this issue, we introduce LVMark, a novel watermarking method for video diffusion models. We propose a new watermark decoder tailored for generated videos by learning the consistency between adjacent frames. It ensures accurate message decoding, even under malicious attacks, by combining the low-frequency components of the 3D wavelet domain with the RGB features of the video. Additionally, our approach minimizes video quality degradation by embedding watermark messages in layers with minimal impact on visual appearance using an importance-based weight modulation strategy. We optimize both the watermark decoder and the latent decoder of diffusion model, effectively balancing the trade-off between visual quality and bit accuracy. Our experiments show that our method embeds invisible watermarks into video diffusion models, ensuring robust decoding accuracy with 512-bit capacity, even under video distortions.
In this work, we explore new perspectives on cross-view completion learning by drawing an analogy to self-supervised correspondence learning. Through our analysis, we demonstrate that the cross-attention map within cross-view completion models captures correspondence more effectively than other correlations derived from encoder or decoder features. We verify the effectiveness of the cross-attention map by evaluating on both zero-shot matching and learning-based geometric matching and multi-frame depth estimation. Project page is available at https://cvlab-kaist.github.io/ZeroCo/.
It is widely acknowledged that the epileptic foci can be pinpointed by source localizing interictal epileptic discharges (IEDs) via Magnetoencephalography (MEG). However, manual detection of IEDs, which appear as spikes in MEG data, is extremely labor intensive and requires considerable professional expertise, limiting the broader adoption of MEG technology. Numerous studies have focused on automatic detection of MEG spikes to overcome this challenge, but these efforts often validate their models on synthetic datasets with balanced positive and negative samples. In contrast, clinical MEG data is highly imbalanced, raising doubts on the real-world efficacy of these models. To address this issue, we introduce LV-CadeNet, a Long View feature Convolution-Attention fusion Encoder-Decoder Network, designed for automatic MEG spike detection in real-world clinical scenarios. Beyond addressing the disparity between training data distribution and clinical test data through semi-supervised learning, our approach also mimics human specialists by constructing long view morphological input data. Moreover, we propose an advanced convolution-attention module to extract temporal and spatial features from the input data. LV-CadeNet significantly improves the accuracy of MEG spike detection, boosting it from 42.31\% to 54.88\% on a novel clinical dataset sourced from Sanbo Brain Hospital Capital Medical University. This dataset, characterized by a highly imbalanced distribution of positive and negative samples, accurately represents real-world clinical scenarios.
In this paper, we introduce ProtoOcc, a novel 3D occupancy prediction model designed to predict the occupancy states and semantic classes of 3D voxels through a deep semantic understanding of scenes. ProtoOcc consists of two main components: the Dual Branch Encoder (DBE) and the Prototype Query Decoder (PQD). The DBE produces a new 3D voxel representation by combining 3D voxel and BEV representations across multiple scales through a dual branch structure. This design enhances both performance and computational efficiency by providing a large receptive field for the BEV representation while maintaining a smaller receptive field for the voxel representation. The PQD introduces Prototype Queries to accelerate the decoding process. Scene-Adaptive Prototypes are derived from the 3D voxel features of input sample, while Scene-Agnostic Prototypes are computed by applying Scene-Adaptive Prototypes to an Exponential Moving Average during the training phase. By using these prototype-based queries for decoding, we can directly predict 3D occupancy in a single step, eliminating the need for iterative Transformer decoding. Additionally, we propose the Robust Prototype Learning, which injects noise into prototype generation process and trains the model to denoise during the training phase. ProtoOcc achieves state-of-the-art performance with 45.02% mIoU on the Occ3D-nuScenes benchmark. For single-frame method, it reaches 39.56% mIoU with an inference speed of 12.83 FPS on an NVIDIA RTX 3090. Our code can be found at https://github.com/SPA-junghokim/ProtoOcc.
Recent work has empirically shown that Vision-Language Models (VLMs) struggle to fully understand the compositional properties of the human language, usually modeling an image caption as a "bag of words". As a result, they perform poorly on compositional tasks, which require a deeper understanding of the different entities of a sentence (subject, verb, etc.) jointly with their mutual relationships in order to be solved. In this paper, we model the dependency relations among textual and visual tokens using a Causal Graphical Model (CGM), built using a dependency parser, and we train a decoder conditioned by the VLM visual encoder. Differently from standard autoregressive or parallel predictions, our decoder's generative process is partially-ordered following the CGM structure. This structure encourages the decoder to learn only the main causal dependencies in a sentence discarding spurious correlations. Using extensive experiments on five compositional benchmarks, we show that our method significantly outperforms all the state-of-the-art compositional approaches by a large margin, and it also improves over methods trained using much larger datasets.
This paper presents a new voice conversion model capable of transforming both speaking and singing voices. It addresses key challenges in current systems, such as conveying emotions, managing pronunciation and accent changes, and reproducing non-verbal sounds. One of the model's standout features is its ability to perform accent conversion on hybrid voice samples that encompass both speech and singing, allowing it to change the speaker's accent while preserving the original content and prosody. The proposed model uses an encoder-decoder architecture: the encoder is based on HuBERT to process the speech's acoustic and linguistic content, while the HiFi-GAN decoder audio matches the target speaker's voice. The model incorporates fundamental frequency (f0) features and singer embeddings to enhance performance while ensuring the pitch & tone accuracy and vocal identity are preserved during transformation. This approach improves how naturally and flexibly voice style can be transformed, showing strong potential for applications in voice dubbing, content creation, and technologies like Text-to-Speech (TTS) and Interactive Voice Response (IVR) systems.
Extraction of building footprint polygons from remotely sensed data is essential for several urban understanding tasks such as reconstruction, navigation, and mapping. Despite significant progress in the area, extracting accurate polygonal building footprints remains an open problem. In this paper, we introduce Pix2Poly, an attention-based end-to-end trainable and differentiable deep neural network capable of directly generating explicit high-quality building footprints in a ring graph format. Pix2Poly employs a generative encoder-decoder transformer to produce a sequence of graph vertex tokens whose connectivity information is learned by an optimal matching network. Compared to previous graph learning methods, ours is a truly end-to-end trainable approach that extracts high-quality building footprints and road networks without requiring complicated, computationally intensive raster loss functions and intricate training pipelines. Upon evaluating Pix2Poly on several complex and challenging datasets, we report that Pix2Poly outperforms state-of-the-art methods in several vector shape quality metrics while being an entirely explicit method. Our code is available at https://github.com/yeshwanth95/Pix2Poly.
Open-vocabulary panoptic segmentation aims to segment and classify everything in diverse scenes across an unbounded vocabulary. Existing methods typically employ two-stage or single-stage framework. The two-stage framework involves cropping the image multiple times using masks generated by a mask generator, followed by feature extraction, while the single-stage framework relies on a heavyweight mask decoder to make up for the lack of spatial position information through self-attention and cross-attention in multiple stacked Transformer blocks. Both methods incur substantial computational overhead, thereby hindering the efficiency of model inference. To fill the gap in efficiency, we propose EOV-Seg, a novel single-stage, shared, efficient, and spatialaware framework designed for open-vocabulary panoptic segmentation. Specifically, EOV-Seg innovates in two aspects. First, a Vocabulary-Aware Selection (VAS) module is proposed to improve the semantic comprehension of visual aggregated features and alleviate the feature interaction burden on the mask decoder. Second, we introduce a Two-way Dynamic Embedding Experts (TDEE), which efficiently utilizes the spatial awareness capabilities of ViT-based CLIP backbone. To the best of our knowledge, EOV-Seg is the first open-vocabulary panoptic segmentation framework towards efficiency, which runs faster and achieves competitive performance compared with state-of-the-art methods. Specifically, with COCO training only, EOV-Seg achieves 24.5 PQ, 32.1 mIoU, and 11.6 FPS on the ADE20K dataset and the inference time of EOV-Seg is 4-19 times faster than state-of-theart methods. Especially, equipped with ResNet50 backbone, EOV-Seg runs 23.8 FPS with only 71M parameters on a single RTX 3090 GPU. Code is available at https://github.com/nhw649/EOV-Seg.
Medical image segmentation plays an important role in clinical decision making, treatment planning, and disease tracking. However, it still faces two major challenges. On the one hand, there is often a ``soft boundary'' between foreground and background in medical images, with poor illumination and low contrast further reducing the distinguishability of foreground and background within the image. On the other hand, co-occurrence phenomena are widespread in medical images, and learning these features is misleading to the model's judgment. To address these challenges, we propose a general framework called Contrast-Driven Medical Image Segmentation (ConDSeg). First, we develop a contrastive training strategy called Consistency Reinforcement. It is designed to improve the encoder's robustness in various illumination and contrast scenarios, enabling the model to extract high-quality features even in adverse environments. Second, we introduce a Semantic Information Decoupling module, which is able to decouple features from the encoder into foreground, background, and uncertainty regions, gradually acquiring the ability to reduce uncertainty during training. The Contrast-Driven Feature Aggregation module then contrasts the foreground and background features to guide multi-level feature fusion and key feature enhancement, further distinguishing the entities to be segmented. We also propose a Size-Aware Decoder to solve the scale singularity of the decoder. It accurately locate entities of different sizes in the image, thus avoiding erroneous learning of co-occurrence features. Extensive experiments on five medical image datasets across three scenarios demonstrate the state-of-the-art performance of our method, proving its advanced nature and general applicability to various medical image segmentation scenarios. Our released code is available at \url{https://github.com/Mengqi-Lei/ConDSeg}.
Prevalent lossy image compression schemes can be divided into: 1) explicit image compression (EIC), including traditional standards and neural end-to-end algorithms; 2) implicit image compression (IIC) based on implicit neural representations (INR). The former is encountering impasses of either leveling off bitrate reduction at a cost of tremendous complexity while the latter suffers from excessive smoothing quality as well as lengthy decoder models. In this paper, we propose an innovative paradigm, which we dub \textbf{Unicorn} (\textbf{U}nified \textbf{N}eural \textbf{I}mage \textbf{C}ompression with \textbf{O}ne \textbf{N}number \textbf{R}econstruction). By conceptualizing the images as index-image pairs and learning the inherent distribution of pairs in a subtle neural network model, Unicorn can reconstruct a visually pleasing image from a randomly generated noise with only one index number. The neural model serves as the unified decoder of images while the noises and indexes corresponds to explicit representations. As a proof of concept, we propose an effective and efficient prototype of Unicorn based on latent diffusion models with tailored model designs. Quantitive and qualitative experimental results demonstrate that our prototype achieves significant bitrates reduction compared with EIC and IIC algorithms. More impressively, benefitting from the unified decoder, our compression ratio escalates as the quantity of images increases. We envision that more advanced model designs will endow Unicorn with greater potential in image compression. We will release our codes in \url{https://github.com/uniqzheng/Unicorn-Laduree}.
The aim of multispectral image fusion is to combine object or scene features of images with different spectral characteristics to increase the perceptual quality. In this paper, we present a novel learning-based solution to image fusion problem focusing on infrared and visible spectrum images. The proposed solution utilizes only convolution and pooling layers together with a loss function using no-reference quality metrics. The analysis is performed qualitatively and quantitatively on various datasets. The results show better performance than state-of-the-art methods. Also, the size of our network enables real-time performance on embedded devices. Project codes can be found at \url{https://github.com/ferhatcan/pyFusionSR}.
In this work, we present Multiformer, a novel approach to depth-aware video panoptic segmentation (DVPS) based on the mask transformer paradigm. Our method learns object representations that are shared across segmentation, monocular depth estimation, and object tracking subtasks. In contrast to recent unified approaches that progressively refine a common object representation, we propose a hybrid method using task-specific branches within each decoder block, ultimately fusing them into a shared representation at the block interfaces. Extensive experiments on the Cityscapes-DVPS and SemKITTI-DVPS datasets demonstrate that Multiformer achieves state-of-the-art performance across all DVPS metrics, outperforming previous methods by substantial margins. With a ResNet-50 backbone, Multiformer surpasses the previous best result by 3.0 DVPQ points while also improving depth estimation accuracy. Using a Swin-B backbone, Multiformer further improves performance by 4.0 DVPQ points. Multiformer also provides valuable insights into the design of multi-task decoder architectures.
Neurogliomas are among the most aggressive forms of cancer, presenting considerable challenges in both treatment and monitoring due to their unpredictable biological behavior. Magnetic resonance imaging (MRI) is currently the preferred method for diagnosing and monitoring gliomas. However, the lack of specific imaging techniques often compromises the accuracy of tumor segmentation during the imaging process. To address this issue, we introduce the XLSTM-HVED model. This model integrates a hetero-modal encoder-decoder framework with the Vision XLSTM module to reconstruct missing MRI modalities. By deeply fusing spatial and temporal features, it enhances tumor segmentation performance. The key innovation of our approach is the Self-Attention Variational Encoder (SAVE) module, which improves the integration of modal features. Additionally, it optimizes the interaction of features between segmentation and reconstruction tasks through the Squeeze-Fusion-Excitation Cross Awareness (SFECA) module. Our experiments using the BraTS 2024 dataset demonstrate that our model significantly outperforms existing advanced methods in handling cases where modalities are missing. Our source code is available at https://github.com/Quanato607/XLSTM-HVED.
Both encoder-only models (e.g., BERT, RoBERTa) and large language models (LLMs, e.g., Llama3) have been widely used for text classification tasks. However, there is a lack of systematic studies comparing the performance of encoder-based models and LLMs in text classification, particularly when fine-tuning is involved. This study employed a diverse range of models and methods, varying in size and architecture, and including both fine-tuned and pre-trained approaches. We first assessed the performances of these LLMs on the 20 Newsgroups (20NG) and MASSIVE datasets, comparing them to encoder-only RoBERTa models. Additionally, we explored the multi-task capabilities of both model types by combining multiple classification tasks, including intent detection and slot-filling, into a single model using data from both datasets. Our results indicate that fully fine-tuned Llama3-70B models outperform RoBERTa-large and other decoder LLMs across various classification tasks and datasets. Moreover, the consolidated multi-task fine-tuned LLMs matched the performance of dual-model setups in both tasks across both datasets. Overall, our study provides a comprehensive benchmark of encoder-only and LLM models on text classification tasks and demonstrates a method to combine two or more fully fine-tuned decoder LLMs for reduced latency and equivalent performance.
Understanding the general priniciples behind transformer models remains a complex endeavor. Experiments with probing and disentangling features using sparse autoencoders (SAE) suggest that these models might manage linear features embedded as directions in the residual stream. This paper explores the resemblance between decoder-only transformer architecture and vector symbolic architectures (VSA) and presents experiments indicating that GPT-2 uses mechanisms involving nearly orthogonal vector bundling and binding operations similar to VSA for computation and communication between layers. It further shows that these principles help explain a significant portion of the actual neural weights.
The advent of Large Language Models (LLM) has reformed the Automatic Speech Recognition (ASR). Prompting LLM with audio embeddings to generate transcriptions becomes the new state-of-the-art ASR. Despite LLMs being trained with an extensive amount of text corpora, high-quality domain-specific text data can still significantly enhance ASR performance on domain adaptation tasks. Although LLM-based ASR can naturally incorporate more text corpora by fine-tuning the LLM decoder, fine-tuning such ASR on text-only data without paired prompts may diminish the effectiveness of domain-specific knowledge. To mitigate this issue, we propose a two-step soft prompt fine-tuning strategy that enhances domain-specific text adaptation. Experimental results show that text adaptation with our proposed method achieved a relative up to 9% Word Error Rate (WER) reduction and up to 18% Entity Error Rate (EER) reduction on the target domain compared to the baseline ASR. Combining this with domain-specific Language Model (LM) fusion can further improve the EER by a relative 2-5%
This paper explores the utility of diffusion-based models for anomaly detection, focusing on their efficacy in identifying deviations in both compact and high-resolution datasets. Diffusion-based architectures, including Denoising Diffusion Probabilistic Models (DDPMs) and Diffusion Transformers (DiTs), are evaluated for their performance using reconstruction objectives. By leveraging the strengths of these models, this study benchmarks their performance against traditional anomaly detection methods such as Isolation Forests, One-Class SVMs, and COPOD. The results demonstrate the superior adaptability, scalability, and robustness of diffusion-based methods in handling complex real-world anomaly detection tasks. Key findings highlight the role of reconstruction error in enhancing detection accuracy and underscore the scalability of these models to high-dimensional datasets. Future directions include optimizing encoder-decoder architectures and exploring multi-modal datasets to further advance diffusion-based anomaly detection.