Nowcasting, the short-term prediction of weather, is essential for making timely and weather-dependent decisions. Specifically, precipitation nowcasting aims to predict precipitation at a local level within a 6-hour time frame. This task can be framed as a spatial-temporal sequence forecasting problem, where deep learning methods have been particularly effective. However, despite advancements in self-supervised learning, most successful methods for nowcasting remain fully supervised. Self-supervised learning is advantageous for pretraining models to learn representations without requiring extensive labeled data. In this work, we leverage the benefits of self-supervised learning and integrate it with spatial-temporal learning to develop a novel model, SpaT-SparK. SpaT-SparK comprises a CNN-based encoder-decoder structure pretrained with a masked image modeling (MIM) task and a translation network that captures temporal relationships among past and future precipitation maps in downstream tasks. We conducted experiments on the NL-50 dataset to evaluate the performance of SpaT-SparK. The results demonstrate that SpaT-SparK outperforms existing baseline supervised models, such as SmaAt-UNet, providing more accurate nowcasting predictions.
Connected component (CC) is a proper text shape representation that aligns with human reading intuition. However, CC-based text detection methods have recently faced a developmental bottleneck that their time-consuming post-processing is difficult to eliminate. To address this issue, we introduce an explicit relational reasoning network (ERRNet) to elegantly model the component relationships without post-processing. Concretely, we first represent each text instance as multiple ordered text components, and then treat these components as objects in sequential movement. In this way, scene text detection can be innovatively viewed as a tracking problem. From this perspective, we design an end-to-end tracking decoder to achieve a CC-based method dispensing with post-processing entirely. Additionally, we observe that there is an inconsistency between classification confidence and localization quality, so we propose a Polygon Monte-Carlo method to quickly and accurately evaluate the localization quality. Based on this, we introduce a position-supervised classification loss to guide the task-aligned learning of ERRNet. Experiments on challenging benchmarks demonstrate the effectiveness of our ERRNet. It consistently achieves state-of-the-art accuracy while holding highly competitive inference speed.
Self-supervised video hashing (SSVH) is a practical task in video indexing and retrieval. Although Transformers are predominant in SSVH for their impressive temporal modeling capabilities, they often suffer from computational and memory inefficiencies. Drawing inspiration from Mamba, an advanced state-space model, we explore its potential in SSVH to achieve a better balance between efficacy and efficiency. We introduce S5VH, a Mamba-based video hashing model with an improved self-supervised learning paradigm. Specifically, we design bidirectional Mamba layers for both the encoder and decoder, which are effective and efficient in capturing temporal relationships thanks to the data-dependent selective scanning mechanism with linear complexity. In our learning strategy, we transform global semantics in the feature space into semantically consistent and discriminative hash centers, followed by a center alignment loss as a global learning signal. Our self-local-global (SLG) paradigm significantly improves learning efficiency, leading to faster and better convergence. Extensive experiments demonstrate S5VH's improvements over state-of-the-art methods, superior transferability, and scalable advantages in inference efficiency. Code is available at https://github.com/gimpong/AAAI25-S5VH.
Gaussian Splatting has enabled real-time 3D human avatars with unprecedented levels of visual quality. While previous methods require a desktop GPU for real-time inference of a single avatar, we aim to squeeze multiple Gaussian avatars onto a portable virtual reality headset with real-time drivable inference. We begin by training a previous work, Animatable Gaussians, on a high quality dataset captured with 512 cameras. The Gaussians are animated by controlling base set of Gaussians with linear blend skinning (LBS) motion and then further adjusting the Gaussians with a neural network decoder to correct their appearance. When deploying the model on a Meta Quest 3 VR headset, we find two major computational bottlenecks: the decoder and the rendering. To accelerate the decoder, we train the Gaussians in UV-space instead of pixel-space, and we distill the decoder to a single neural network layer. Further, we discover that neighborhoods of Gaussians can share a single corrective from the decoder, which provides an additional speedup. To accelerate the rendering, we develop a custom pipeline in Vulkan that runs on the mobile GPU. Putting it all together, we run 3 Gaussian avatars concurrently at 72 FPS on a VR headset. Demo videos are at https://forresti.github.io/squeezeme.
Fake news poses a significant threat to public opinion and social stability in modern society. This study presents a comparative evaluation of BERT-like encoder-only models and autoregressive decoder-only large language models (LLMs) for fake news detection. We introduce a dataset of news articles labeled with GPT-4 assistance (an AI-labeling method) and verified by human experts to ensure reliability. Both BERT-like encoder-only models and LLMs were fine-tuned on this dataset. Additionally, we developed an instruction-tuned LLM approach with majority voting during inference for label generation. Our analysis reveals that BERT-like models generally outperform LLMs in classification tasks, while LLMs demonstrate superior robustness against text perturbations. Compared to weak labels (distant supervision) data, the results show that AI labels with human supervision achieve better classification results. This study highlights the effectiveness of combining AI-based annotation with human oversight and demonstrates the performance of different families of machine learning models for fake news detection
Due to the superior ability of global dependency, transformer and its variants have become the primary choice in Masked Time-series Modeling (MTM) towards time-series classification task. In this paper, we experimentally analyze that existing transformer-based MTM methods encounter with two under-explored issues when dealing with time series data: (1) they encode features by performing long-dependency ensemble averaging, which easily results in rank collapse and feature homogenization as the layer goes deeper; (2) they exhibit distinct priorities in fitting different frequency components contained in the time-series, inevitably leading to spectrum energy imbalance of encoded feature. To tackle these issues, we propose an auxiliary content-aware balanced decoder (CBD) to optimize the encoding quality in the spectrum space within masked modeling scheme. Specifically, the CBD iterates on a series of fundamental blocks, and thanks to two tailored units, each block could progressively refine the masked representation via adjusting the interaction pattern based on local content variations of time-series and learning to recalibrate the energy distribution across different frequency components. Moreover, a dual-constraint loss is devised to enhance the mutual optimization of vanilla decoder and our CBD. Extensive experimental results on ten time-series classification datasets show that our method nearly surpasses a bunch of baselines. Meanwhile, a series of explanatory results are showcased to sufficiently demystify the behaviors of our method.
In the quest to overcome language barriers, encoder-decoder models like NLLB have expanded machine translation to rare languages, with some models (e.g., NLLB 1.3B) even trainable on a single GPU. While general-purpose LLMs perform well in translation, open LLMs prove highly competitive when fine-tuned for specific tasks involving unknown corpora. We introduce LYRA (Language verY Rare for All), a novel approach that combines open LLM fine-tuning, retrieval-augmented generation (RAG), and transfer learning from related high-resource languages. This study is exclusively focused on single-GPU training to facilitate ease of adoption. Our study focuses on two-way translation between French and Mon\'egasque, a rare language unsupported by existing translation tools due to limited corpus availability. Our results demonstrate LYRA's effectiveness, frequently surpassing and consistently matching state-of-the-art encoder-decoder models in rare language translation.
Computer-Aided Design (CAD) models are typically constructed by sequentially drawing parametric sketches and applying CAD operations to obtain a 3D model. The problem of 3D CAD reverse engineering consists of reconstructing the sketch and CAD operation sequences from 3D representations such as point clouds. In this paper, we address this challenge through novel contributions across three levels: CAD sequence representation, network design, and dataset. In particular, we represent CAD sketch-extrude sequences as Python code. The proposed CAD-Recode translates a point cloud into Python code that, when executed, reconstructs the CAD model. Taking advantage of the exposure of pre-trained Large Language Models (LLMs) to Python code, we leverage a relatively small LLM as a decoder for CAD-Recode and combine it with a lightweight point cloud projector. CAD-Recode is trained solely on a proposed synthetic dataset of one million diverse CAD sequences. CAD-Recode significantly outperforms existing methods across three datasets while requiring fewer input points. Notably, it achieves 10 times lower mean Chamfer distance than state-of-the-art methods on DeepCAD and Fusion360 datasets. Furthermore, we show that our CAD Python code output is interpretable by off-the-shelf LLMs, enabling CAD editing and CAD-specific question answering from point clouds.
Prompts play a critical role in unleashing the power of language and vision foundation models for specific tasks. For the first time, we introduce prompting into depth foundation models, creating a new paradigm for metric depth estimation termed Prompt Depth Anything. Specifically, we use a low-cost LiDAR as the prompt to guide the Depth Anything model for accurate metric depth output, achieving up to 4K resolution. Our approach centers on a concise prompt fusion design that integrates the LiDAR at multiple scales within the depth decoder. To address training challenges posed by limited datasets containing both LiDAR depth and precise GT depth, we propose a scalable data pipeline that includes synthetic data LiDAR simulation and real data pseudo GT depth generation. Our approach sets new state-of-the-arts on the ARKitScenes and ScanNet++ datasets and benefits downstream applications, including 3D reconstruction and generalized robotic grasping.
Encoder-only transformer models such as BERT offer a great performance-size tradeoff for retrieval and classification tasks with respect to larger decoder-only models. Despite being the workhorse of numerous production pipelines, there have been limited Pareto improvements to BERT since its release. In this paper, we introduce ModernBERT, bringing modern model optimizations to encoder-only models and representing a major Pareto improvement over older encoders. Trained on 2 trillion tokens with a native 8192 sequence length, ModernBERT models exhibit state-of-the-art results on a large pool of evaluations encompassing diverse classification tasks and both single and multi-vector retrieval on different domains (including code). In addition to strong downstream performance, ModernBERT is also the most speed and memory efficient encoder and is designed for inference on common GPUs.
In many problem settings that require spatio-temporal forecasting, the values in the time-series not only exhibit spatio-temporal correlations but are also influenced by spatial diffusion across locations. One such example is forecasting the concentration of fine particulate matter (PM2.5) in the atmosphere which is influenced by many complex factors, the most important ones being diffusion due to meteorological factors as well as transport across vast distances over a period of time. We present a novel Spatio-Temporal Graph Neural Network architecture, that specifically captures these dependencies to forecast the PM2.5 concentration. Our model is based on an encoder-decoder architecture where the encoder and decoder parts leverage gated recurrent units (GRU) augmented with a graph neural network (TransformerConv) to account for spatial diffusion. Our model can also be seen as a generalization of various existing models for time-series or spatio-temporal forecasting. We demonstrate the model's effectiveness on two real-world PM2.5 datasets: (1) data collected by us using a recently deployed network of low-cost PM$_{2.5}$ sensors from 511 locations spanning the entirety of the Indian state of Bihar over a period of one year, and (2) another publicly available dataset that covers severely polluted regions from China for a period of 4 years. Our experimental results show our model's impressive ability to account for both spatial as well as temporal dependencies precisely.
Chain-of-thought (CoT) decoding enables language models to improve reasoning performance at the cost of high generation latency in decoding. Recent proposals have explored variants of contemplation tokens, a term we introduce that refers to special tokens used during inference to allow for extra computation. Prior work has considered fixed-length sequences drawn from a discrete set of embeddings as contemplation tokens. Here we propose Compressed Chain-of-Thought (CCoT), a framework to generate contentful and continuous contemplation tokens of variable sequence length. The generated contemplation tokens are compressed representations of explicit reasoning chains, and our method can be applied to off-the-shelf decoder language models. Through experiments, we illustrate how CCoT enables additional reasoning over dense contentful representations to achieve corresponding improvements in accuracy. Moreover, the reasoning improvements can be adaptively modified on demand by controlling the number of contemplation tokens generated.
Transformers have emerged as the leading architecture in deep learning, proving to be versatile and highly effective across diverse domains beyond language and image processing. However, their impressive performance often incurs high computational costs due to their substantial model size. This paper focuses on compressing decoder-only transformer-based autoregressive models through structural weight pruning to improve the model efficiency while preserving performance for both language and image generation tasks. Specifically, we propose a training-free pruning method that calculates a numerical score with Newton's method for the Attention and MLP modules, respectively. Besides, we further propose another compensation algorithm to recover the pruned model for better performance. To verify the effectiveness of our method, we provide both theoretical support and extensive experiments. Our experiments show that our method achieves state-of-the-art performance with reduced memory usage and faster generation speeds on GPUs.
Tasks in which rewards depend upon past information not available in the current observation set can only be solved by agents that are equipped with short-term memory. Usual choices for memory modules include trainable recurrent hidden layers, often with gated memory. Reservoir computing presents an alternative, in which a recurrent layer is not trained, but rather has a set of fixed, sparse recurrent weights. The weights are scaled to produce stable dynamical behavior such that the reservoir state contains a high-dimensional, nonlinear impulse response function of the inputs. An output decoder network can then be used to map the compressive history represented by the reservoir's state to any outputs, including agent actions or predictions. In this study, we find that reservoir computing greatly simplifies and speeds up reinforcement learning on memory tasks by (1) eliminating the need for backpropagation of gradients through time, (2) presenting all recent history simultaneously to the downstream network, and (3) performing many useful and generic nonlinear computations upstream from the trained modules. In particular, these findings offer significant benefit to meta-learning that depends primarily on efficient and highly general memory systems.
Recently, developing unified medical image segmentation models gains increasing attention, especially with the advent of the Segment Anything Model (SAM). SAM has shown promising binary segmentation performance in natural domains, however, transferring it to the medical domain remains challenging, as medical images often possess substantial inter-category overlaps. To address this, we propose the SEmantic-Guided SAM (SEG-SAM), a unified medical segmentation model that incorporates semantic medical knowledge to enhance medical segmentation performance. First, to avoid the potential conflict between binary and semantic predictions, we introduce a semantic-aware decoder independent of SAM's original decoder, specialized for both semantic segmentation on the prompted object and classification on unprompted objects in images. To further enhance the model's semantic understanding, we solicit key characteristics of medical categories from large language models and incorporate them into SEG-SAM through a text-to-vision semantic module, adaptively transferring the language information into the visual segmentation task. In the end, we introduce the cross-mask spatial alignment strategy to encourage greater overlap between the predicted masks from SEG-SAM's two decoders, thereby benefiting both predictions. Extensive experiments demonstrate that SEG-SAM outperforms state-of-the-art SAM-based methods in unified binary medical segmentation and task-specific methods in semantic medical segmentation, showcasing promising results and potential for broader medical applications.
Referring multi-object tracking (RMOT) is an emerging cross-modal task that aims to localize an arbitrary number of targets based on a language expression and continuously track them in a video. This intricate task involves reasoning on multi-modal data and precise target localization with temporal association. However, prior studies overlook the imbalanced data distribution between newborn targets and existing targets due to the nature of the task. In addition, they only indirectly fuse multi-modal features, struggling to deliver clear guidance on newborn target detection. To solve the above issues, we conduct a collaborative matching strategy to alleviate the impact of the imbalance, boosting the ability to detect newborn targets while maintaining tracking performance. In the encoder, we integrate and enhance the cross-modal and multi-scale fusion, overcoming the bottlenecks in previous work, where limited multi-modal information is shared and interacted between feature maps. In the decoder, we also develop a referring-infused adaptation that provides explicit referring guidance through the query tokens. The experiments showcase the superior performance of our model (+3.42%) compared to prior works, demonstrating the effectiveness of our designs.
This paper addresses key challenges in object-centric representation learning of video. While existing approaches struggle with complex scenes, we propose a novel weakly-supervised framework that emphasises geometric understanding and leverages pre-trained vision models to enhance object discovery. Our method introduces an efficient slot decoder specifically designed for object-centric learning, enabling effective representation of multi-object scenes without requiring explicit depth information. Results on synthetic video benchmarks with increasing complexity in terms of objects and their movement, object occlusion and camera motion demonstrate that our approach achieves comparable performance to supervised methods while maintaining computational efficiency. This advances the field towards more practical applications in complex real-world scenarios.
Recent advances in Artificial Intelligence Generated Content (AIGC) have garnered significant interest, accompanied by an increasing need to transmit and compress the vast number of AI-generated images (AIGIs). However, there is a noticeable deficiency in research focused on compression methods for AIGIs. To address this critical gap, we introduce a scalable cross-modal compression framework that incorporates multiple human-comprehensible modalities, designed to efficiently capture and relay essential visual information for AIGIs. In particular, our framework encodes images into a layered bitstream consisting of a semantic layer that delivers high-level semantic information through text prompts; a structural layer that captures spatial details using edge or skeleton maps; and a texture layer that preserves local textures via a colormap. Utilizing Stable Diffusion as the backend, the framework effectively leverages these multimodal priors for image generation, effectively functioning as a decoder when these priors are encoded. Qualitative and quantitative results show that our method proficiently restores both semantic and visual details, competing against baseline approaches at extremely low bitrates ( <0.02 bpp). Additionally, our framework facilitates downstream editing applications without requiring full decoding, thereby paving a new direction for future research in AIGI compression.
In radar-camera 3D object detection, the radar point clouds are sparse and noisy, which causes difficulties in fusing camera and radar modalities. To solve this, we introduce a novel query-based detection method named Radar-Camera Transformer (RCTrans). Specifically, we first design a Radar Dense Encoder to enrich the sparse valid radar tokens, and then concatenate them with the image tokens. By doing this, we can fully explore the 3D information of each interest region and reduce the interference of empty tokens during the fusing stage. We then design a Pruning Sequential Decoder to predict 3D boxes based on the obtained tokens and random initialized queries. To alleviate the effect of elevation ambiguity in radar point clouds, we gradually locate the position of the object via a sequential fusion structure. It helps to get more precise and flexible correspondences between tokens and queries. A pruning training strategy is adopted in the decoder, which can save much time during inference and inhibit queries from losing their distinctiveness. Extensive experiments on the large-scale nuScenes dataset prove the superiority of our method, and we also achieve new state-of-the-art radar-camera 3D detection results. Our implementation is available at https://github.com/liyih/RCTrans.
In this paper, we tackle the challenge of predicting stock movements in financial markets by introducing Higher Order Transformers, a novel architecture designed for processing multivariate time-series data. We extend the self-attention mechanism and the transformer architecture to a higher order, effectively capturing complex market dynamics across time and variables. To manage computational complexity, we propose a low-rank approximation of the potentially large attention tensor using tensor decomposition and employ kernel attention, reducing complexity to linear with respect to the data size. Additionally, we present an encoder-decoder model that integrates technical and fundamental analysis, utilizing multimodal signals from historical prices and related tweets. Our experiments on the Stocknet dataset demonstrate the effectiveness of our method, highlighting its potential for enhancing stock movement prediction in financial markets.