Large Language Models (LLMs) demonstrate exceptional capabilities in a multitude of NLP tasks. However, the efficacy of such models to languages other than English is often limited. Prior works have shown that encoder-only models such as BERT or XLM-RoBERTa show impressive cross lingual transfer of their capabilities from English to other languages. In this work, we propose a pretraining strategy that uses active forgetting to achieve similar cross lingual transfer in decoder-only LLMs. We show that LLMs pretrained with active forgetting are highly effective when adapting to new and unseen languages. Through extensive experimentation, we find that LLMs pretrained with active forgetting are able to learn better multilingual representations which translates to better performance in many downstream tasks.
This paper introduces Moonshine, a family of speech recognition models optimized for live transcription and voice command processing. Moonshine is based on an encoder-decoder transformer architecture and employs Rotary Position Embedding (RoPE) instead of traditional absolute position embeddings. The model is trained on speech segments of various lengths, but without using zero-padding, leading to greater efficiency for the encoder during inference time. When benchmarked against OpenAI's Whisper tiny-en, Moonshine Tiny demonstrates a 5x reduction in compute requirements for transcribing a 10-second speech segment while incurring no increase in word error rates across standard evaluation datasets. These results highlight Moonshine's potential for real-time and resource-constrained applications.
We present a neural network-enhanced column generation (CG) approach for a parallel machine scheduling problem. The proposed approach utilizes an encoder-decoder attention model, namely the transformer and pointer architectures, to develop job sequences with negative reduced cost and thus generate columns to add to the master problem. By training the neural network offline and using it in inference mode to predict negative reduced costs columns, we achieve significant computational time savings compared to dynamic programming (DP). Since the exact DP procedure is used to verify that no further columns with negative reduced cost can be identified at termination, the optimality guarantee of the original CG procedure is preserved. For small to medium-sized instances, our approach achieves an average 45% reduction in computation time compared to solving the subproblems with DP. Furthermore, the model generalizes not only to unseen, larger problem instances from the same probability distribution but also to instances from different probability distributions than those presented at training time. For large-sized instances, the proposed approach achieves an 80% improvement in the objective value in under 500 seconds, demonstrating both its scalability and efficiency.
With the rapid development of deep learning, CNN-based U-shaped networks have succeeded in medical image segmentation and are widely applied for various tasks. However, their limitations in capturing global features hinder their performance in complex segmentation tasks. The rise of Vision Transformer (ViT) has effectively compensated for this deficiency of CNNs and promoted the application of ViT-based U-networks in medical image segmentation. However, the high computational demands of ViT make it unsuitable for many medical devices and mobile platforms with limited resources, restricting its deployment on resource-constrained and edge devices. To address this, we propose EViT-UNet, an efficient ViT-based segmentation network that reduces computational complexity while maintaining accuracy, making it ideal for resource-constrained medical devices. EViT-UNet is built on a U-shaped architecture, comprising an encoder, decoder, bottleneck layer, and skip connections, combining convolutional operations with self-attention mechanisms to optimize efficiency. Experimental results demonstrate that EViT-UNet achieves high accuracy in medical image segmentation while significantly reducing computational complexity.
The digital image manipulation and advancements in Generative AI, such as Deepfake, has raised significant concerns regarding the authenticity of images shared on social media. Traditional image forensic techniques, while helpful, are often passive and insufficient against sophisticated tampering methods. This paper introduces the Secure Learned Image Codec (SLIC), a novel active approach to ensuring image authenticity through watermark embedding in the compressed domain. SLIC leverages neural network-based compression to embed watermarks as adversarial perturbations in the latent space, creating images that degrade in quality upon re-compression if tampered with. This degradation acts as a defense mechanism against unauthorized modifications. Our method involves fine-tuning a neural encoder/decoder to balance watermark invisibility with robustness, ensuring minimal quality loss for non-watermarked images. Experimental results demonstrate SLIC's effectiveness in generating visible artifacts in tampered images, thereby preventing their redistribution. This work represents a significant step toward developing secure image codecs that can be widely adopted to safeguard digital image integrity.
Novel-view synthesis aims to generate novel views of a scene from multiple input images or videos, and recent advancements like 3D Gaussian splatting (3DGS) have achieved notable success in producing photorealistic renderings with efficient pipelines. However, generating high-quality novel views under challenging settings, such as sparse input views, remains difficult due to insufficient information in under-sampled areas, often resulting in noticeable artifacts. This paper presents 3DGS-Enhancer, a novel pipeline for enhancing the representation quality of 3DGS representations. We leverage 2D video diffusion priors to address the challenging 3D view consistency problem, reformulating it as achieving temporal consistency within a video generation process. 3DGS-Enhancer restores view-consistent latent features of rendered novel views and integrates them with the input views through a spatial-temporal decoder. The enhanced views are then used to fine-tune the initial 3DGS model, significantly improving its rendering performance. Extensive experiments on large-scale datasets of unbounded scenes demonstrate that 3DGS-Enhancer yields superior reconstruction performance and high-fidelity rendering results compared to state-of-the-art methods. The project webpage is https://xiliu8006.github.io/3DGS-Enhancer-project .
We introduce SeaDAG, a semi-autoregressive diffusion model for conditional generation of Directed Acyclic Graphs (DAGs). Considering their inherent layer-wise structure, we simulate layer-wise autoregressive generation by designing different denoising speed for different layers. Unlike conventional autoregressive generation that lacks a global graph structure view, our method maintains a complete graph structure at each diffusion step, enabling operations such as property control that require the full graph structure. Leveraging this capability, we evaluate the DAG properties during training by employing a graph property decoder. We explicitly train the model to learn graph conditioning with a condition loss, which enhances the diffusion model's capacity to generate graphs that are both realistic and aligned with specified properties. We evaluate our method on two representative conditional DAG generation tasks: (1) circuit generation from truth tables, where precise DAG structures are crucial for realizing circuit functionality, and (2) molecule generation based on quantum properties. Our approach demonstrates promising results, generating high-quality and realistic DAGs that closely align with given conditions.
We develop a novel data-driven framework as an alternative to dynamic flux balance analysis, bypassing the demand for deep domain knowledge and manual efforts to formulate the optimization problem. The proposed framework is end-to-end, which trains a structured neural ODE process (SNODEP) model to estimate flux and balance samples using gene-expression time-series data. SNODEP is designed to circumvent the limitations of the standard neural ODE process model, including restricting the latent and decoder sampling distributions to be normal and lacking structure between context points for calculating the latent, thus more suitable for modeling the underlying dynamics of a metabolic system. Through comprehensive experiments ($156$ in total), we demonstrate that SNODEP not only predicts the unseen time points of real-world gene-expression data and the flux and balance estimates well but can even generalize to more challenging unseen knockout configurations and irregular data sampling scenarios, all essential for metabolic pathway analysis. We hope our work can serve as a catalyst for building more scalable and powerful models for genome-scale metabolic analysis. Our code is available at: \url{https://github.com/TrustMLRG/SNODEP}.
Deep learning models have achieved remarkable progress in precipitation prediction. However, they still face significant challenges in accurately capturing spatial details of radar images, particularly in regions of high precipitation intensity. This limitation results in reduced spatial localization accuracy when predicting radar echo images across varying precipitation intensities. To address this challenge, we propose an innovative precipitation prediction approach termed the Multi-Task Latent Diffusion Model (MTLDM). The core idea of MTLDM lies in the recognition that precipitation radar images represent a combination of multiple components, each corresponding to different precipitation intensities. Thus, we adopt a divide-and-conquer strategy, decomposing radar images into several sub-images based on their precipitation intensities and individually modeling these components. During the prediction stage, MTLDM integrates these sub-image representations by utilizing a trained latent-space rainfall diffusion model, followed by decoding through a multi-task decoder to produce the final precipitation prediction. Experimental evaluations conducted on the MRMS dataset demonstrate that the proposed MTLDM method surpasses state-of-the-art techniques, achieving a Critical Success Index (CSI) improvement of 13-26%.
Spiking neural networks (SNNs) are neural networks that enable energy-efficient signal processing due to their event-based nature. This paper proposes a novel decoding algorithm for low-density parity-check (LDPC) codes that integrates SNNs into belief propagation (BP) decoding by approximating the check node update equations using SNNs. For the (273,191) and (1023,781) finite-geometry LDPC code, the proposed decoder outperforms sum-product decoder at high signal-to-noise ratios (SNRs). The decoder achieves a similar bit error rate to normalized sum-product decoding with successive relaxation. Furthermore, the novel decoding operates without requiring knowledge of the SNR, making it robust to SNR mismatch.
Users engage with financial services companies through multiple channels, often interacting with mobile applications, web platforms, call centers, and physical locations to service their accounts. The resulting interactions are recorded at heterogeneous temporal resolutions across these domains. This multi-channel data can be combined and encoded to create a comprehensive representation of the customer's journey for accurate intent prediction. This demands sequential learning solutions. NMT transformers achieve state-of-the-art sequential representation learning by encoding context and decoding for the next best action to represent long-range dependencies. However, three major challenges exist while combining multi-domain sequences within an encoder-decoder transformers architecture for intent prediction applications: a) aligning sequences with different sampling rates b) learning temporal dynamics across multi-variate, multi-domain sequences c) combining dynamic and static sequences. We propose an encoder-decoder transformer model to address these challenges for contextual and sequential intent prediction in financial servicing applications. Our experiments show significant improvement over the existing tabular method.
Navigation is a fundamental cognitive skill extensively studied in neuroscientific experiments and has lately gained substantial interest in artificial intelligence research. Recreating the task solved by rodents in the well-established Morris Water Maze (MWM) experiment, this work applies a transformer-based architecture using deep reinforcement learning -- an approach previously unexplored in this context -- to navigate a 2D version of the maze. Specifically, the agent leverages a decoder-only transformer architecture serving as a deep Q-network performing effective decision making in the partially observable environment. We demonstrate that the proposed architecture enables the agent to efficiently learn spatial navigation strategies, overcoming challenges associated with a limited field of vision, corresponding to the visual information available to a rodent in the MWM. Demonstrating the potential of transformer-based models for enhancing navigation performance in partially observable environments, this work suggests promising avenues for future research in artificial agents whose behavior resembles that of biological agents. Finally, the flexibility of the transformer architecture in supporting varying input sequence lengths opens opportunities for gaining increased understanding of the artificial agent's inner representation of the environment.
This paper presents a novel system for reconstructing high-resolution GPS trajectory data from truncated or synthetic low-resolution inputs, addressing the critical challenge of balancing data utility with privacy preservation in mobility applications. The system integrates transformer-based encoder-decoder models with graph convolutional networks (GCNs) to effectively capture both the temporal dependencies of trajectory data and the spatial relationships in road networks. By combining these techniques, the system is able to recover fine-grained trajectory details that are lost through data truncation or rounding, a common practice to protect user privacy. We evaluate the system on the Beijing trajectory dataset, demonstrating its superior performance over traditional map-matching algorithms and LSTM-based synthetic data generation methods. The proposed model achieves an average Fr\'echet distance of 0.198 km, significantly outperforming map-matching algorithms (0.632 km) and synthetic trajectory models (0.498 km). The results show that the system is not only capable of accurately reconstructing real-world trajectories but also generalizes effectively to synthetic data. These findings suggest that the system can be deployed in urban mobility applications, providing both high accuracy and robust privacy protection.
Numerically solving a large parametric nonlinear dynamical system is challenging due to its high complexity and the high computational costs. In recent years, machine-learning-aided surrogates are being actively researched. However, many methods fail in accurately generalizing in the entire time interval $[0, T]$, when the training data is available only in a training time interval $[0, T_0]$, with $T_0<T$. To improve the extrapolation capabilities of the surrogate models in the entire time domain, we propose a new deep learning framework, where kernel dynamic mode decomposition (KDMD) is employed to evolve the dynamics of the latent space generated by the encoder part of a convolutional autoencoder (CAE). After adding the KDMD-decoder-extrapolated data into the original data set, we train the CAE along with a feed-forward deep neural network using the augmented data. The trained network can predict future states outside the training time interval at any out-of-training parameter samples. The proposed method is tested on two numerical examples: a FitzHugh-Nagumo model and a model of incompressible flow past a cylinder. Numerical results show accurate and fast prediction performance in both the time and the parameter domain.
Parameter-efficient fine-tuning (PEFT) methods typically assume that Large Language Models (LLMs) are trained on data from a single device or client. However, real-world scenarios often require fine-tuning these models on private data distributed across multiple devices. Federated Learning (FL) offers an appealing solution by preserving user privacy, as sensitive data remains on local devices during training. Nonetheless, integrating PEFT methods into FL introduces two main challenges: communication overhead and data heterogeneity. In this paper, we introduce FedTT and FedTT+, methods for adapting LLMs by integrating tensorized adapters into client-side models' encoder/decoder blocks. FedTT is versatile and can be applied to both cross-silo FL and large-scale cross-device FL. FedTT+, an extension of FedTT tailored for cross-silo FL, enhances robustness against data heterogeneity by adaptively freezing portions of tensor factors, further reducing the number of trainable parameters. Experiments on BERT and LLaMA models demonstrate that our proposed methods successfully address data heterogeneity challenges and perform on par or even better than existing federated PEFT approaches while achieving up to 10$\times$ reduction in communication cost.
The pervasiveness of proprietary language models has raised privacy concerns for users' sensitive data, emphasizing the need for private inference (PI), where inference is performed directly on encrypted inputs. However, current PI methods face prohibitively higher communication and latency overheads, primarily due to nonlinear operations. In this paper, we present a comprehensive analysis to understand the role of nonlinearities in transformer-based decoder-only language models. We introduce AERO, a four-step architectural optimization framework that refines the existing LLM architecture for efficient PI by systematically removing nonlinearities such as LayerNorm and GELU and reducing FLOPs counts. For the first time, we propose a Softmax-only architecture with significantly fewer FLOPs tailored for efficient PI. Furthermore, we devise a novel entropy regularization technique to improve the performance of Softmax-only models. AERO achieves up to 4.23$\times$ communication and 1.94$\times$ latency reduction. We validate the effectiveness of AERO by benchmarking it against the state-of-the-art.
We focus on improving the visual understanding capability for boosting the vision-language models. We propose \textbf{Arcana}, a multiModal language model, which introduces two crucial techniques. First, we present Multimodal LoRA (MM-LoRA), a module designed to enhance the decoder. Unlike traditional language-driven decoders, MM-LoRA consists of two parallel LoRAs -- one for vision and one for language -- each with its own parameters. This disentangled parameters design allows for more specialized learning in each modality and better integration of multimodal information. Second, we introduce the Query Ladder adapter (QLadder) to improve the visual encoder. QLadder employs a learnable ``\textit{ladder}'' structure to deeply aggregates the intermediate representations from the frozen pretrained visual encoder (e.g., CLIP image encoder). This enables the model to learn new and informative visual features, as well as remaining the powerful capabilities of the pretrained visual encoder. These techniques collectively enhance Arcana's visual perception power, enabling it to leverage improved visual information for more accurate and contextually relevant outputs across various multimodal scenarios. Extensive experiments and ablation studies demonstrate the effectiveness and generalization capability of our Arcana. The code and re-annotated data are available at \url{https://arcana-project-page.github.io}.
Large language models have shown exceptional capabilities in a wide range of tasks, such as text generation and video generation, among others. However, due to their massive parameter count, these models often require substantial storage space, imposing significant constraints on the machines deploying LLMs. To overcome this limitation, one research direction proposes to compress the models using integer replacements for floating-point numbers, in a process known as Quantization. Some recent studies suggest quantizing the key and value cache (KV Cache) of LLMs, and designing quantization techniques that treat the key and value matrices equivalently. This work delves deeper into the asymmetric structural roles of KV Cache, a phenomenon where the transformer's output loss is more sensitive to the quantization of key matrices. We conduct a systematic examination of the attention output error resulting from key and value quantization. The phenomenon inspires us to propose an asymmetric quantization strategy. Our approach allows for 1-bit quantization of the KV cache by implementing distinct configurations for key and value matrices. We carry out experiments across a variety of datasets, demonstrating that our proposed model allows for the quantization of up to 75% decoder layers with 1 bit, while simultaneously maintaining performance levels comparable to those of the models with floating parameters.
In this work, we introduce UniGS, a novel 3D Gaussian reconstruction and novel view synthesis model that predicts a high-fidelity representation of 3D Gaussians from arbitrary number of posed sparse-view images. Previous methods often regress 3D Gaussians locally on a per-pixel basis for each view and then transfer them to world space and merge them through point concatenation. In contrast, Our approach involves modeling unitary 3D Gaussians in world space and updating them layer by layer. To leverage information from multi-view inputs for updating the unitary 3D Gaussians, we develop a DETR (DEtection TRansformer)-like framework, which treats 3D Gaussians as queries and updates their parameters by performing multi-view cross-attention (MVDFA) across multiple input images, which are treated as keys and values. This approach effectively avoids `ghosting' issue and allocates more 3D Gaussians to complex regions. Moreover, since the number of 3D Gaussians used as decoder queries is independent of the number of input views, our method allows arbitrary number of multi-view images as input without causing memory explosion or requiring retraining. Extensive experiments validate the advantages of our approach, showcasing superior performance over existing methods quantitatively (improving PSNR by 4.2 dB when trained on Objaverse and tested on the GSO benchmark) and qualitatively. The code will be released at https://github.com/jwubz123/UNIG.
We analyze contextual representations in neural autoregressive language models, emphasizing long-range contexts that span several thousand tokens. Our methodology employs a perturbation setup and the metric \emph{Anisotropy-Calibrated Cosine Similarity}, to capture the degree of contextualization of long-range patterns from the perspective of representation geometry. We begin the analysis with a case study on standard decoder-only Transformers, demonstrating that similar perplexity can exhibit markedly different downstream task performance, which can be explained by the difference in contextualization of long-range content. Next, we extend the analysis to other models, covering recent novel architectural designs and various training configurations. The representation-level results illustrate a reduced capacity for high-complexity (i.e., less compressible) sequences across architectures, and that fully recurrent models rely heavily on local context, whereas hybrid models more effectively encode the entire sequence structure. Finally, preliminary analysis of model size and training configurations on the encoding of long-range context suggest potential directions for improving existing language models.