Decoder-only Transformers often struggle with complex reasoning tasks, particularly arithmetic reasoning requiring multiple sequential operations. In this work, we identify representation collapse in the model's intermediate layers as a key factor limiting their reasoning capabilities. To address this, we propose Sequential Variance-Covariance Regularization (Seq-VCR), which enhances the entropy of intermediate representations and prevents collapse. Combined with dummy pause tokens as substitutes for chain-of-thought (CoT) tokens, our method significantly improves performance in arithmetic reasoning problems. In the challenging $5 \times 5$ integer multiplication task, our approach achieves $99.5\%$ exact match accuracy, outperforming models of the same size (which yield $0\%$ accuracy) and GPT-4 with five-shot CoT prompting ($44\%$). We also demonstrate superior results on arithmetic expression and longest increasing subsequence (LIS) datasets. Our findings highlight the importance of preventing intermediate layer representation collapse to enhance the reasoning capabilities of Transformers and show that Seq-VCR offers an effective solution without requiring explicit CoT supervision.
The adoption of increasingly complex deep models has fueled an urgent need for insight into how these models make predictions. Counterfactual explanations form a powerful tool for providing actionable explanations to practitioners. Previously, counterfactual explanation methods have been designed by traversing the latent space of generative models. Yet, these latent spaces are usually greatly simplified, with most of the data distribution complexity contained in the decoder rather than the latent embedding. Thus, traversing the latent space naively without taking the nonlinear decoder into account can lead to unnatural counterfactual trajectories. We introduce counterfactual explanations obtained using a Riemannian metric pulled back via the decoder and the classifier under scrutiny. This metric encodes information about the complex geometric structure of the data and the learned representation, enabling us to obtain robust counterfactual trajectories with high fidelity, as demonstrated by our experiments in real-world tabular datasets.
Monocular metric depth estimation (MMDE) is a crucial task to solve for indoor scene reconstruction on edge devices. Despite this importance, existing models are sensitive to factors such as boundary frequency of objects in the scene and scene complexity, failing to fully capture many indoor scenes. In this work, we propose to close this gap through the task of monocular metric depth refinement (MMDR) by leveraging state-of-the-art MMDE models. MultiDepth proposes a solution by taking samples of the image along with the initial depth map prediction made by a pre-trained MMDE model. Compared to existing iterative depth refinement techniques, MultiDepth does not employ normal map prediction as part of its architecture, effectively lowering the model size and computation overhead while outputting impactful changes from refining iterations. MultiDepth implements a lightweight encoder-decoder architecture for the refinement network, processing multiple samples from the given image, including segmentation masking. We evaluate MultiDepth on four datasets and compare them to state-of-the-art methods to demonstrate its effective refinement with minimal overhead, displaying accuracy improvement upward of 45%.
Recent video semantic segmentation (VSS) methods have demonstrated promising results in well-lit environments. However, their performance significantly drops in low-light scenarios due to limited visibility and reduced contextual details. In addition, unfavorable low-light conditions make it harder to incorporate temporal consistency across video frames and thus, lead to video flickering effects. Compared with conventional cameras, event cameras can capture motion dynamics, filter out temporal-redundant information, and are robust to lighting conditions. To this end, we propose EVSNet, a lightweight framework that leverages event modality to guide the learning of a unified illumination-invariant representation. Specifically, we leverage a Motion Extraction Module to extract short-term and long-term temporal motions from event modality and a Motion Fusion Module to integrate image features and motion features adaptively. Furthermore, we use a Temporal Decoder to exploit video contexts and generate segmentation predictions. Such designs in EVSNet result in a lightweight architecture while achieving SOTA performance. Experimental results on 3 large-scale datasets demonstrate our proposed EVSNet outperforms SOTA methods with up to 11x higher parameter efficiency.
PET imaging is an invaluable tool in clinical settings as it captures the functional activity of both healthy anatomy and cancerous lesions. Developing automatic lesion segmentation methods for PET images is crucial since manual lesion segmentation is laborious and prone to inter- and intra-observer variability. We propose PET-Disentangler, a 3D disentanglement method that uses a 3D UNet-like encoder-decoder architecture to disentangle disease and normal healthy anatomical features with losses for segmentation, reconstruction, and healthy component plausibility. A critic network is used to encourage the healthy latent features to match the distribution of healthy samples and thus encourages these features to not contain any lesion-related features. Our quantitative results show that PET-Disentangler is less prone to incorrectly declaring healthy and high tracer uptake regions as cancerous lesions, since such uptake pattern would be assigned to the disentangled healthy component.
Understanding and predicting human actions has been a long-standing challenge and is a crucial measure of perception in robotics AI. While significant progress has been made in anticipating the future actions of individual agents, prior work has largely overlooked a key aspect of real-world human activity -- interactions. To address this gap in human-like forecasting within multi-agent environments, we present the Hierarchical Memory-Aware Transformer (HiMemFormer), a transformer-based model for online multi-agent action anticipation. HiMemFormer integrates and distributes global memory that captures joint historical information across all agents through a transformer framework, with a hierarchical local memory decoder that interprets agent-specific features based on these global representations using a coarse-to-fine strategy. In contrast to previous approaches, HiMemFormer uniquely hierarchically applies the global context with agent-specific preferences to avoid noisy or redundant information in multi-agent action anticipation. Extensive experiments on various multi-agent scenarios demonstrate the significant performance of HiMemFormer, compared with other state-of-the-art methods.
We introduce a novel framework, Online Relational Inference (ORI), designed to efficiently identify hidden interaction graphs in evolving multi-agent interacting systems using streaming data. Unlike traditional offline methods that rely on a fixed training set, ORI employs online backpropagation, updating the model with each new data point, thereby allowing it to adapt to changing environments in real-time. A key innovation is the use of an adjacency matrix as a trainable parameter, optimized through a new adaptive learning rate technique called AdaRelation, which adjusts based on the historical sensitivity of the decoder to changes in the interaction graph. Additionally, a data augmentation method named Trajectory Mirror (TM) is introduced to improve generalization by exposing the model to varied trajectory patterns. Experimental results on both synthetic datasets and real-world data (CMU MoCap for human motion) demonstrate that ORI significantly improves the accuracy and adaptability of relational inference in dynamic settings compared to existing methods. This approach is model-agnostic, enabling seamless integration with various neural relational inference (NRI) architectures, and offers a robust solution for real-time applications in complex, evolving systems.
Inverse tone mapping, a technique to restore a high dynamic range (HDR) image from a single low dynamic range (LDR) image, exhibits wide versatility since it may be easily applied to any camera device. Besides, the recent advancement in deep learning has produced great performance improvement in the field of inverse tone mapping. However, it remains a difficult task to accurately restore a wide-range HDR image from a single LDR image. A recent study attempts a spatially adaptive exposure value (EV) condition generated from luminance values to create a pseudo-multi-exposure stack. However, by adopting only luminance values as input, the conditioning method cannot precisely reflect the input image information when generating the EV condition, resulting in the loss of color expression. Moreover, there are some concerns regarding how to apply the EV condition to the image feature. Thus, the key idea of this study is to directly adopt image features in generating EV conditions that are adaptive to both color and brightness. To do this, we design a condition generation network with an encoder-decoder structure and propose a novel multi-exposure stack generation network, which bidirectionally synthesizes the image features and EV-conditioned features. Additionally, to better preserve the feature information in the synthesis of the features, we propose a spatially-adaptive feature transformation block. Our proposed method exhibits outstanding results in restoring the multi-exposure stacks for HDR image synthesis. Furthermore, our method achieves state-of-the-art performance compared to existing methods in multi-exposure stack generation and stack-based HDR restoration.
Current vision systems typically assign fixed-length representations to images, regardless of the information content. This contrasts with human intelligence - and even large language models - which allocate varying representational capacities based on entropy, context and familiarity. Inspired by this, we propose an approach to learn variable-length token representations for 2D images. Our encoder-decoder architecture recursively processes 2D image tokens, distilling them into 1D latent tokens over multiple iterations of recurrent rollouts. Each iteration refines the 2D tokens, updates the existing 1D latent tokens, and adaptively increases representational capacity by adding new tokens. This enables compression of images into a variable number of tokens, ranging from 32 to 256. We validate our tokenizer using reconstruction loss and FID metrics, demonstrating that token count aligns with image entropy, familiarity and downstream task requirements. Recurrent token processing with increasing representational capacity in each iteration shows signs of token specialization, revealing potential for object / part discovery.
Trajectory prediction module in an autonomous driving system is crucial for the decision-making and safety of the autonomous agent car and its surroundings. This work presents a novel scheme called AiGem (Agent-Interaction Graph Embedding) to predict traffic vehicle trajectories around the autonomous car. AiGem tackles this problem in four steps. First, AiGem formulates the historical traffic interaction with the autonomous agent as a graph in two steps: (1) at each time step of the history frames, agent-interactions are captured using spatial edges between the agents (nodes of the graph), and then, (2) connects the spatial graphs in chronological order using temporal edges. Then, AiGem applies a depthwise graph encoder network on the spatial-temporal graph to generate graph embedding, i.e., embedding of all the nodes in the graph. Next, a sequential Gated Recurrent Unit decoder network uses the embedding of the current timestamp to get the decoded states. Finally, an output network comprising a Multilayer Perceptron is used to predict the trajectories utilizing the decoded states as its inputs. Results show that AiGem outperforms the state-of-the-art deep learning algorithms for longer prediction horizons.
This paper proposes a novel neural audio codec, named APCodec+, which is an improved version of APCodec. The APCodec+ takes the audio amplitude and phase spectra as the coding object, and employs an adversarial training strategy. Innovatively, we propose a two-stage joint-individual training paradigm for APCodec+. In the joint training stage, the encoder, quantizer, decoder and discriminator are jointly trained with complete spectral loss, quantization loss, and adversarial loss. In the individual training stage, the encoder and quantizer fix their parameters and provide high-quality training data for the decoder and discriminator. The decoder and discriminator are individually trained from scratch without the quantization loss. The purpose of introducing individual training is to reduce the learning difficulty of the decoder, thereby further improving the fidelity of the decoded audio. Experimental results confirm that our proposed APCodec+ at low bitrates achieves comparable performance with baseline codecs at higher bitrates, thanks to the proposed staged training paradigm.
We present a novel approach to paraphrase generation that enables precise control and fine-tuning of 40 linguistic attributes for English. Our model is an encoder-decoder architecture that takes as input a source sentence and desired linguistic attributes, and produces paraphrases of the source that satisfy the desired attributes. To guarantee high-quality outputs at inference time, our method is equipped with a quality control mechanism that gradually adjusts the embedding of linguistic attributes to find the nearest and most attainable configuration of desired attributes for paraphrase generation. We evaluate the effectiveness of our method by comparing it to recent controllable generation models. Experimental results demonstrate that the proposed model outperforms baselines in generating paraphrases that satisfy desired linguistic attributes.
Spoken language models (SLMs) have gained increasing attention with advancements in text-based, decoder-only language models. SLMs process text and speech, enabling simultaneous speech understanding and generation. This paper presents Double-Codebook Speaker-invariant Clustering (DC-Spin), which aims to improve speech tokenization by bridging audio signals and SLM tokens. DC-Spin extracts speaker-invariant tokens rich in phonetic information and resilient to input variations, enhancing zero-shot SLM tasks and speech resynthesis. We propose a chunk-wise approach to enable streamable DC-Spin without retraining and degradation. Comparisons of tokenization methods (self-supervised and neural audio codecs), model scalability, and downstream task proxies show that tokens easily modeled by an n-gram LM or aligned with phonemes offer strong performance, providing insights for designing speech tokenizers for SLMs.
Remote sensing image change caption (RSICC) aims to provide natural language descriptions for bi-temporal remote sensing images. Since Change Caption (CC) task requires both spatial and temporal features, previous works follow an encoder-fusion-decoder architecture. They use an image encoder to extract spatial features and the fusion module to integrate spatial features and extract temporal features, which leads to increasingly complex manual design of the fusion module. In this paper, we introduce a novel video model-based paradigm without design of the fusion module and propose a Mask-enhanced Video model for Change Caption (MV-CC). Specifically, we use the off-the-shelf video encoder to simultaneously extract the temporal and spatial features of bi-temporal images. Furthermore, the types of changes in the CC are set based on specific task requirements, and to enable the model to better focus on the regions of interest, we employ masks obtained from the Change Detection (CD) method to explicitly guide the CC model. Experimental results demonstrate that our proposed method can obtain better performance compared with other state-of-the-art RSICC methods. The code is available at https://github.com/liuruixun/MV-CC.
With the advancement of deepfake generation techniques, the importance of deepfake detection in protecting multimedia content integrity has become increasingly obvious. Recently, temporal inconsistency clues have been explored to improve the generalizability of deepfake video detection. According to our observation, the temporal artifacts of forged videos in terms of motion information usually exhibits quite distinct inconsistency patterns along horizontal and vertical directions, which could be leveraged to improve the generalizability of detectors. In this paper, a transformer-based framework for Diffusion Learning of Inconsistency Pattern (DIP) is proposed, which exploits directional inconsistencies for deepfake video detection. Specifically, DIP begins with a spatiotemporal encoder to represent spatiotemporal information. A directional inconsistency decoder is adopted accordingly, where direction-aware attention and inconsistency diffusion are incorporated to explore potential inconsistency patterns and jointly learn the inherent relationships. In addition, the SpatioTemporal Invariant Loss (STI Loss) is introduced to contrast spatiotemporally augmented sample pairs and prevent the model from overfitting nonessential forgery artifacts. Extensive experiments on several public datasets demonstrate that our method could effectively identify directional forgery clues and achieve state-of-the-art performance.
This paper presents OmniJARVIS, a novel Vision-Language-Action (VLA) model for open-world instruction-following agents in Minecraft. Compared to prior works that either emit textual goals to separate controllers or produce the control command directly, OmniJARVIS seeks a different path to ensure both strong reasoning and efficient decision-making capabilities via unified tokenization of multimodal interaction data. First, we introduce a self-supervised approach to learn a behavior encoder that produces discretized tokens for behavior trajectories $\tau = \{o_0, a_0, \dots\}$ and an imitation learning policy decoder conditioned on these tokens. These additional behavior tokens will be augmented to the vocabulary of pretrained Multimodal Language Models. With this encoder, we then pack long-term multimodal interactions involving task instructions, memories, thoughts, observations, textual responses, behavior trajectories, etc into unified token sequences and model them with autoregressive transformers. Thanks to the semantically meaningful behavior tokens, the resulting VLA model, OmniJARVIS, can reason (by producing chain-of-thoughts), plan, answer questions, and act (by producing behavior tokens for the imitation learning policy decoder). OmniJARVIS demonstrates excellent performances on a comprehensive collection of atomic, programmatic, and open-ended tasks in open-world Minecraft. Our analysis further unveils the crucial design principles in interaction data formation, unified tokenization, and its scaling potentials. The dataset, models, and code will be released at https://craftjarvis.org/OmniJARVIS.
Deep Neural Networks (DNNs) are well-known to act as over-parameterized deep image priors (DIP) that regularize various image inverse problems. Meanwhile, researchers also proposed extremely compact, under-parameterized image priors (e.g., deep decoder) that are strikingly competent for image restoration too, despite a loss of accuracy. These two extremes push us to think whether there exists a better solution in the middle: between over- and under-parameterized image priors, can one identify "intermediate" parameterized image priors that achieve better trade-offs between performance, efficiency, and even preserving strong transferability? Drawing inspirations from the lottery ticket hypothesis (LTH), we conjecture and study a novel "lottery image prior" (LIP) by exploiting DNN inherent sparsity, stated as: given an over-parameterized DNN-based image prior, it will contain a sparse subnetwork that can be trained in isolation, to match the original DNN's performance when being applied as a prior to various image inverse problems. Our results validate the superiority of LIPs: we can successfully locate the LIP subnetworks from over-parameterized DIPs at substantial sparsity ranges. Those LIP subnetworks significantly outperform deep decoders under comparably compact model sizes (by often fully preserving the effectiveness of their over-parameterized counterparts), and they also possess high transferability across different images as well as restoration task types. Besides, we also extend LIP to compressive sensing image reconstruction, where a pre-trained GAN generator is used as the prior (in contrast to untrained DIP or deep decoder), and confirm its validity in this setting too. To our best knowledge, this is the first time that LTH is demonstrated to be relevant in the context of inverse problems or image priors.
State-of-the-art Extreme Multi-Label Text Classification (XMTC) models rely heavily on multi-label attention layers to focus on key tokens in input text, but obtaining optimal attention weights is challenging and resource-intensive. To address this, we introduce PLANT -- Pretrained and Leveraged AtteNTion -- a novel transfer learning strategy for fine-tuning XMTC decoders. PLANT surpasses existing state-of-the-art methods across all metrics on mimicfull, mimicfifty, mimicfour, eurlex, and wikiten datasets. It particularly excels in few-shot scenarios, outperforming previous models specifically designed for few-shot scenarios by over 50 percentage points in F1 scores on mimicrare and by over 36 percentage points on mimicfew, demonstrating its superior capability in handling rare codes. PLANT also shows remarkable data efficiency in few-shot scenarios, achieving precision comparable to traditional models with significantly less data. These results are achieved through key technical innovations: leveraging a pretrained Learning-to-Rank model as the planted attention layer, integrating mutual-information gain to enhance attention, introducing an inattention mechanism, and implementing a stateful-decoder to maintain context. Comprehensive ablation studies validate the importance of these contributions in realizing the performance gains.
One-shot transfer of dexterous grasps to novel scenes with object and context variations has been a challenging problem. While distilled feature fields from large vision models have enabled semantic correspondences across 3D scenes, their features are point-based and restricted to object surfaces, limiting their capability of modeling complex semantic feature distributions for hand-object interactions. In this work, we propose the \textit{neural attention field} for representing semantic-aware dense feature fields in the 3D space by modeling inter-point relevance instead of individual point features. Core to it is a transformer decoder that computes the cross-attention between any 3D query point with all the scene points, and provides the query point feature with an attention-based aggregation. We further propose a self-supervised framework for training the transformer decoder from only a few 3D pointclouds without hand demonstrations. Post-training, the attention field can be applied to novel scenes for semantics-aware dexterous grasping from one-shot demonstration. Experiments show that our method provides better optimization landscapes by encouraging the end-effector to focus on task-relevant scene regions, resulting in significant improvements in success rates on real robots compared with the feature-field-based methods.
With the rapid development of Natural Language Processing (NLP) technology, the accuracy and efficiency of machine translation have become hot topics of research. This paper proposes a novel Seq2Seq model aimed at improving translation quality while reducing the storage space required by the model. The model employs a Bidirectional Long Short-Term Memory network (Bi-LSTM) as the encoder to capture the context information of the input sequence; the decoder incorporates an attention mechanism, enhancing the model's ability to focus on key information during the translation process. Compared to the current mainstream Transformer model, our model achieves superior performance on the WMT14 machine translation dataset while maintaining a smaller size. The study first introduces the design principles and innovative points of the model architecture, followed by a series of experiments to verify the effectiveness of the model. The experimental includes an assessment of the model's performance on different language pairs, as well as comparative analysis with traditional Seq2Seq models. The results show that while maintaining translation accuracy, our model significantly reduces the storage requirements, which is of great significance for translation applications in resource-constrained scenarios. our code are available at https://github.com/mindspore-lab/models/tree/master/research/arxiv_papers/miniformer. Thanks for the support provided by MindSpore Community.