Image captioning and cross-modal retrieval are examples of tasks that involve the joint analysis of visual and linguistic information. In connection to remote sensing imagery, these tasks can help non-expert users in extracting relevant Earth observation information for a variety of applications. Still, despite some previous efforts, the development and application of vision and language models to the remote sensing domain have been hindered by the relatively small size of the available datasets and models used in previous studies. In this work, we propose RS-CapRet, a Vision and Language method for remote sensing tasks, in particular image captioning and text-image retrieval. We specifically propose to use a highly capable large decoder language model together with image encoders adapted to remote sensing imagery through contrastive language-image pre-training. To bridge together the image encoder and language decoder, we propose training simple linear layers with examples from combining different remote sensing image captioning datasets, keeping the other parameters frozen. RS-CapRet can then generate descriptions for remote sensing images and retrieve images from textual descriptions, achieving SOTA or competitive performance with existing methods. Qualitative results illustrate that RS-CapRet can effectively leverage the pre-trained large language model to describe remote sensing images, retrieve them based on different types of queries, and also show the ability to process interleaved sequences of images and text in a dialogue manner.
In autonomous driving, 3D lane detection using monocular cameras is an important task for various downstream planning and control tasks. Recent CNN and Transformer approaches usually apply a two-stage scheme in the model design. The first stage transforms the image feature from a front image into a bird's-eye-view (BEV) representation. Subsequently, a sub-network processes the BEV feature map to generate the 3D detection results. However, these approaches heavily rely on a challenging image feature transformation module from a perspective view to a BEV representation. In our work, we present CurveFormer++, a single-stage Transformer-based method that does not require the image feature view transform module and directly infers 3D lane detection results from the perspective image features. Specifically, our approach models the 3D detection task as a curve propagation problem, where each lane is represented by a curve query with a dynamic and ordered anchor point set. By employing a Transformer decoder, the model can iteratively refine the 3D lane detection results. A curve cross-attention module is introduced in the Transformer decoder to calculate similarities between image features and curve queries of lanes. To handle varying lane lengths, we employ context sampling and anchor point restriction techniques to compute more relevant image features for a curve query. Furthermore, we apply a temporal fusion module that incorporates selected informative sparse curve queries and their corresponding anchor point sets to leverage historical lane information. In the experiments, we evaluate our approach for the 3D lane detection task on two publicly available real-world datasets. The results demonstrate that our method provides outstanding performance compared with both CNN and Transformer based methods. We also conduct ablation studies to analyze the impact of each component in our approach.
This early-stage research work aims to improve online human-robot imitation by translating sequences of joint positions from the domain of human motions to a domain of motions achievable by a given robot, thus constrained by its embodiment. Leveraging the generalization capabilities of deep learning methods, we address this problem by proposing an encoder-decoder neural network model performing domain-to-domain translation. In order to train such a model, one could use pairs of associated robot and human motions. Though, such paired data is extremely rare in practice, and tedious to collect. Therefore, we turn towards deep learning methods for unpaired domain-to-domain translation, that we adapt in order to perform human-robot imitation.
We present GenEFT: an effective theory framework for shedding light on the statics and dynamics of neural network generalization, and illustrate it with graph learning examples. We first investigate the generalization phase transition as data size increases, comparing experimental results with information-theory-based approximations. We find generalization in a Goldilocks zone where the decoder is neither too weak nor too powerful. We then introduce an effective theory for the dynamics of representation learning, where latent-space representations are modeled as interacting particles (repons), and find that it explains our experimentally observed phase transition between generalization and overfitting as encoder and decoder learning rates are scanned. This highlights the power of physics-inspired effective theories for bridging the gap between theoretical predictions and practice in machine learning.
In this paper, we propose a new decoding method called Permute-and-Flip (PF) decoder. It enjoys stability properties similar to the standard sampling decoder, but is provably up to 2x better in its quality-stability tradeoff than sampling and never worse than any other decoder. We also design a cryptographic watermarking scheme analogous to Aaronson (2023)'s Gumbel watermark, but naturally tailored for PF decoder. The watermarking scheme does not change the distribution to sample, while allowing arbitrarily low false positive rate and high recall whenever the generated text has high entropy. Our experiments show that the PF decoder (and its watermarked counterpart) significantly outperform(s) naive sampling (and its Gumbel watermarked counterpart) in terms of perplexity, while retaining the same stability (and detectability), hence making it a promising new approach for LLM decoding. The code is available at https://github.com/XuandongZhao/pf-decoding
Longitudinal data are important in numerous fields, such as healthcare, sociology and seismology, but real-world datasets present notable challenges for practitioners because they can be high-dimensional, contain structured missingness patterns, and measurement time points can be governed by an unknown stochastic process. While various solutions have been suggested, the majority of them have been designed to account for only one of these challenges. In this work, we propose a flexible and efficient latent-variable model that is capable of addressing all these limitations. Our approach utilizes Gaussian processes to capture temporal correlations between samples and their associated missingness masks as well as to model the underlying point process. We construct our model as a variational autoencoder together with deep neural network parameterised encoder and decoder models, and develop a scalable amortised variational inference approach for efficient model training. We demonstrate competitive performance using both simulated and real datasets.
We study the fundamental optimization principles of self-attention, the defining mechanism of transformers, by analyzing the implicit bias of gradient-based optimizers in training a self-attention layer with a linear decoder in binary classification. Building on prior studies in linear logistic regression, recent findings demonstrate that the key-query matrix $W_t$ from gradient-descent (GD) converges in direction towards $W_{mm}$, which maximizes the margin between optimal and non-optimal tokens across sequences. However, this convergence is local, dependent on initial conditions, only holds asymptotically as the number of iterations increases, and leaves questions about the potential benefits of adaptive step-size rules unaddressed. To bridge this gap, we first establish scenarios for which convergence is provably \emph{global}. We then analyze two adaptive step-size strategies: normalized GD and Polyak step-size, demonstrating \emph{finite-time} convergence rates for $W_t$ to $W_{mm}$, and quantifying the sparsification rate of the attention map. These findings not only show that these strategies can accelerate parameter convergence over standard GD in a non-convex setting but also deepen the understanding of the implicit bias in self-attention, linking it more closely to the phenomena observed in linear logistic regression despite its intricate non-convex nature.
In recent years, the detection of infrared small targets using deep learning methods has garnered substantial attention due to notable advancements. To improve the detection capability of small targets, these methods commonly maintain a pathway that preserves high-resolution features of sparse and tiny targets. However, it can result in redundant and expensive computations. To tackle this challenge, we propose SpirDet, a novel approach for efficient detection of infrared small targets. Specifically, to cope with the computational redundancy issue, we employ a new dual-branch sparse decoder to restore the feature map. Firstly, the fast branch directly predicts a sparse map indicating potential small target locations (occupying only 0.5\% area of the map). Secondly, the slow branch conducts fine-grained adjustments at the positions indicated by the sparse map. Additionally, we design an lightweight DO-RepEncoder based on reparameterization with the Downsampling Orthogonality, which can effectively reduce memory consumption and inference latency. Extensive experiments show that the proposed SpirDet significantly outperforms state-of-the-art models while achieving faster inference speed and fewer parameters. For example, on the IRSTD-1K dataset, SpirDet improves $MIoU$ by 4.7 and has a $7\times$ $FPS$ acceleration compared to the previous state-of-the-art model. The code will be open to the public.
We examine pulse amplitude modulation (PAM) for intensity modulation and direct detection systems. Using a straight-forward, mixed noise model, we optimize the constellations with an autoencoder-based neural network (NN), an improve required signal-to-noise ratio of 4 dB for amplified spontaneous emission (ASE)-limited PAM4 and PAM8, without increasing system complexity. Performance can also be improved in O-band wavelength division multiplexing system with semiconductor optical amplifier amplification and chromatic dispersion. We show via simulation that for such a system operating at 53 Gbaud, we can extend the reach of PAM4 by 10-25 km with an optimized constellation and a NN decoder. We present an experimental validation of 4 dB improvement of an ASE-limited PAM4 at 60 Gbaud using an optimized constellation and a NN decoder.
Decoder-only large language models (LLMs) have recently demonstrated impressive capabilities in text generation and reasoning. Nonetheless, they have limited applications in simultaneous machine translation (SiMT), currently dominated by encoder-decoder transformers. This study demonstrates that, after fine-tuning on a small dataset comprising causally aligned source and target sentence pairs, a pre-trained open-source LLM can control input segmentation directly by generating a special "wait" token. This obviates the need for a separate policy and enables the LLM to perform English-German and English-Russian SiMT tasks with BLEU scores that are comparable to those of specific state-of-the-art baselines. We also evaluated closed-source models such as GPT-4, which displayed encouraging results in performing the SiMT task without prior training (zero-shot), indicating a promising avenue for enhancing future SiMT systems.
Autoencoders are a prominent model in many empirical branches of machine learning and lossy data compression. However, basic theoretical questions remain unanswered even in a shallow two-layer setting. In particular, to what degree does a shallow autoencoder capture the structure of the underlying data distribution? For the prototypical case of the 1-bit compression of sparse Gaussian data, we prove that gradient descent converges to a solution that completely disregards the sparse structure of the input. Namely, the performance of the algorithm is the same as if it was compressing a Gaussian source - with no sparsity. For general data distributions, we give evidence of a phase transition phenomenon in the shape of the gradient descent minimizer, as a function of the data sparsity: below the critical sparsity level, the minimizer is a rotation taken uniformly at random (just like in the compression of non-sparse data); above the critical sparsity, the minimizer is the identity (up to a permutation). Finally, by exploiting a connection with approximate message passing algorithms, we show how to improve upon Gaussian performance for the compression of sparse data: adding a denoising function to a shallow architecture already reduces the loss provably, and a suitable multi-layer decoder leads to a further improvement. We validate our findings on image datasets, such as CIFAR-10 and MNIST.
In forensic investigations of machine learning models, techniques that determine a model's data domain play an essential role, with prior work relying on large-scale corpora like ImageNet to approximate the target model's domain. Although such methods are effective in finding broad domains, they often struggle in identifying finer-grained classes within those domains. In this paper, we introduce an enhanced approach to determine not just the general data domain (e.g., human face) but also its specific attributes (e.g., wearing glasses). Our approach uses an image embedding model as the encoder and a generative model as the decoder. Beginning with a coarse-grained description, the decoder generates a set of images, which are then presented to the unknown target model. Successful classifications by the model guide the encoder to refine the description, which in turn, are used to produce a more specific set of images in the subsequent iteration. This iterative refinement narrows down the exact class of interest. A key strength of our approach lies in leveraging the expansive dataset, LAION-5B, on which the generative model Stable Diffusion is trained. This enlarges our search space beyond traditional corpora, such as ImageNet. Empirical results showcase our method's performance in identifying specific attributes of a model's input domain, paving the way for more detailed forensic analyses of deep learning models.
In recent advancements in medical image analysis, Convolutional Neural Networks (CNN) and Vision Transformers (ViT) have set significant benchmarks. While the former excels in capturing local features through its convolution operations, the latter achieves remarkable global context understanding by leveraging self-attention mechanisms. However, both architectures exhibit limitations in efficiently modeling long-range dependencies within medical images, which is a critical aspect for precise segmentation. Inspired by the Mamba architecture, known for its proficiency in handling long sequences and global contextual information with enhanced computational efficiency as a State Space Model (SSM), we propose Mamba-UNet, a novel architecture that synergizes the U-Net in medical image segmentation with Mamba's capability. Mamba-UNet adopts a pure Visual Mamba (VMamba)-based encoder-decoder structure, infused with skip connections to preserve spatial information across different scales of the network. This design facilitates a comprehensive feature learning process, capturing intricate details and broader semantic contexts within medical images. We introduce a novel integration mechanism within the VMamba blocks to ensure seamless connectivity and information flow between the encoder and decoder paths, enhancing the segmentation performance. We conducted experiments on publicly available ACDC MRI Cardiac segmentation dataset, and Synapse CT Abdomen segmentation dataset. The results show that Mamba-UNet outperforms several types of UNet in medical image segmentation under the same hyper-parameter setting. The source code and baseline implementations are available.
Comprehending natural language instructions is a charming property for 3D indoor scene synthesis systems. Existing methods directly model object joint distributions and express object relations implicitly within a scene, thereby hindering the controllability of generation. We introduce InstructScene, a novel generative framework that integrates a semantic graph prior and a layout decoder to improve controllability and fidelity for 3D scene synthesis. The proposed semantic graph prior jointly learns scene appearances and layout distributions, exhibiting versatility across various downstream tasks in a zero-shot manner. To facilitate the benchmarking for text-driven 3D scene synthesis, we curate a high-quality dataset of scene-instruction pairs with large language and multimodal models. Extensive experimental results reveal that the proposed method surpasses existing state-of-the-art approaches by a large margin. Thorough ablation studies confirm the efficacy of crucial design components. Project page: https://chenguolin.github.io/projects/InstructScene.
Chinese grammatical error correction (CGEC) faces serious overcorrection challenges when employing autoregressive generative models such as sequence-to-sequence (Seq2Seq) models and decoder-only large language models (LLMs). While previous methods aim to address overcorrection in Seq2Seq models, they are difficult to adapt to decoder-only LLMs. In this paper, we propose an alignment-enhanced corrector for the overcorrection problem that applies to both Seq2Seq models and decoder-only LLMs. Our method first trains a correction model to generate an initial correction of the source sentence. Then, we combine the source sentence with the initial correction and feed it through an alignment model for another round of correction, aiming to enforce the alignment model to focus on potential overcorrection. Moreover, to enhance the model's ability to identify nuances, we further explore the reverse alignment of the source sentence and the initial correction. Finally, we transfer the alignment knowledge from two alignment models to the correction model, instructing it on how to avoid overcorrection. Experimental results on three CGEC datasets demonstrate the effectiveness of our approach in alleviating overcorrection and improving overall performance. Our code has been made publicly available.
We present EfficientViT-SAM, a new family of accelerated segment anything models. We retain SAM's lightweight prompt encoder and mask decoder while replacing the heavy image encoder with EfficientViT. For the training, we begin with the knowledge distillation from the SAM-ViT-H image encoder to EfficientViT. Subsequently, we conduct end-to-end training on the SA-1B dataset. Benefiting from EfficientViT's efficiency and capacity, EfficientViT-SAM delivers 48.9x measured TensorRT speedup on A100 GPU over SAM-ViT-H without sacrificing performance. Our code and pre-trained models are released at https://github.com/mit-han-lab/efficientvit.
The decoder-only Transformer architecture with causal masking and relative position encoding (RPE) has become the de facto choice in language modeling. Despite its exceptional performance across various tasks, we have identified two limitations: First, it requires all attention scores to be non-zero and sum up to 1, even if the current embedding has sufficient self-contained information. This compels the model to assign disproportional excessive attention to specific tokens. Second, RPE-based Transformers are not universal approximators due to their limited capacity at encoding absolute positional information, which limits their application in position-critical tasks. In this work, we propose StableMask: a parameter-free method to address both limitations by refining the causal mask. It introduces pseudo-attention values to balance attention distributions and encodes absolute positional information via a progressively decreasing mask ratio. StableMask's effectiveness is validated both theoretically and empirically, showing significant enhancements in language models with parameter sizes ranging from 71M to 1.4B across diverse datasets and encoding methods. We further show that it naturally supports (1) efficient extrapolation without special tricks such as StreamingLLM and (2) easy integration with existing attention optimization techniques.
Large Language Models (LLMs) are one of the most promising technologies for the next era of speech generation systems, due to their scalability and in-context learning capabilities. Nevertheless, they suffer from multiple stability issues at inference time, such as hallucinations, content skipping or speech repetitions. In this work, we introduce a new self-supervised Voice Conversion (VC) architecture which can be used to learn to encode transitory features, such as content, separately from stationary ones, such as speaker ID or recording conditions, creating speaker-disentangled representations. Using speaker-disentangled codes to train LLMs for text-to-speech (TTS) allows the LLM to generate the content and the style of the speech only from the text, similarly to humans, while the speaker identity is provided by the decoder of the VC model. Results show that LLMs trained over speaker-disentangled self-supervised representations provide an improvement of 4.7pp in speaker similarity over SOTA entangled representations, and a word error rate (WER) 5.4pp lower. Furthermore, they achieve higher naturalness than human recordings of the LibriTTS test-other dataset. Finally, we show that using explicit reference embedding negatively impacts intelligibility (stability), with WER increasing by 14pp compared to the model that only uses text to infer the style.
Network traffic refers to the amount of data being sent and received over the internet or any system that connects computers. Analyzing and understanding network traffic is vital for improving network security and management. However, the analysis of network traffic is challenging due to the diverse nature of data packets, which often feature heterogeneous headers and encrypted payloads lacking semantics. To capture the latent semantics of traffic, a few studies have adopted pre-training techniques based on the Transformer encoder or decoder to learn the representations from massive traffic data. However, these methods typically excel in traffic understanding (classification) or traffic generation tasks. To address this issue, we develop Lens, a foundation model for network traffic that leverages the T5 architecture to learn the pre-trained representations from large-scale unlabeled data. Harnessing the strength of the encoder-decoder framework, which captures the global information while preserving the generative ability, our model can better learn the representations from raw data. To further enhance pre-training effectiveness, we design a novel loss that combines three distinct tasks: Masked Span Prediction (MSP), Packet Order Prediction (POP), and Homologous Traffic Prediction (HTP). Evaluation results across various benchmark datasets demonstrate that the proposed Lens outperforms the baselines in most downstream tasks related to both traffic understanding and generation. Notably, it also requires much less labeled data for fine-tuning compared to current methods.
Scattering and attenuation of light in no-homogeneous imaging media or inconsistent light intensity will cause insufficient contrast and color distortion in the collected images, which limits the developments such as vision-driven smart urban, autonomous vehicles, and intelligent robots. In this paper, we propose an all-in-one scene recovery network via multi-knowledge integration (termed AoSRNet) to improve the visibility of imaging devices in typical low-visibility imaging scenes (e.g., haze, sand dust, and low light). It combines gamma correction (GC) and optimized linear stretching (OLS) to create the detail enhancement module (DEM) and color restoration module (CRM). Additionally, we suggest a multi-receptive field extraction module (MEM) to attenuate the loss of image texture details caused by GC nonlinear and OLS linear transformations. Finally, we refine the coarse features generated by DEM, CRM, and MEM through Encoder-Decoder to generate the final restored image. Comprehensive experimental results demonstrate the effectiveness and stability of AoSRNet compared to other state-of-the-art methods. The source code is available at \url{https://github.com/LouisYuxuLu/AoSRNet}.