Model-based reinforcement learning agents utilizing transformers have shown improved sample efficiency due to their ability to model extended context, resulting in more accurate world models. However, for complex reasoning and planning tasks, these methods primarily rely on continuous representations. This complicates modeling of discrete properties of the real world such as disjoint object classes between which interpolation is not plausible. In this work, we introduce discrete abstract representations for transformer-based learning (DART), a sample-efficient method utilizing discrete representations for modeling both the world and learning behavior. We incorporate a transformer-decoder for auto-regressive world modeling and a transformer-encoder for learning behavior by attending to task-relevant cues in the discrete representation of the world model. For handling partial observability, we aggregate information from past time steps as memory tokens. DART outperforms previous state-of-the-art methods that do not use look-ahead search on the Atari 100k sample efficiency benchmark with a median human-normalized score of 0.790 and beats humans in 9 out of 26 games. We release our code at https://pranaval.github.io/DART/.
Capturing complex temporal patterns and relationships within multivariate data streams is a difficult task. We propose the Temporal Kolmogorov-Arnold Transformer (TKAT), a novel attention-based architecture designed to address this task using Temporal Kolmogorov-Arnold Networks (TKANs). Inspired by the Temporal Fusion Transformer (TFT), TKAT emerges as a powerful encoder-decoder model tailored to handle tasks in which the observed part of the features is more important than the a priori known part. This new architecture combined the theoretical foundation of the Kolmogorov-Arnold representation with the power of transformers. TKAT aims to simplify the complex dependencies inherent in time series, making them more "interpretable". The use of transformer architecture in this framework allows us to capture long-range dependencies through self-attention mechanisms.
Pre-trained language models demonstrate general intelligence and common sense, but long inputs quickly become a bottleneck for memorizing information at inference time. We resurface a simple method, Memorizing Transformers (Wu et al., 2022), that gives the model access to a bank of pre-computed memories. We show that it is possible to fix many of the shortcomings of the original method, such as the need for fine-tuning, by critically assessing how positional encodings should be updated for the keys and values retrieved. This intuitive method uses the model's own key/query system to select and attend to the most relevant memories at each generation step, rather than using external embeddings. We demonstrate the importance of external information being retrieved in a majority of decoder layers, contrary to previous work. We open source a new counterfactual long-range retrieval benchmark, and show that Extended Mind Transformers outperform today's state of the art by 6% on average.
In this paper, we address the challenging problem of long-term 3D human motion generation. Specifically, we aim to generate a long sequence of smoothly connected actions from a stream of multiple sentences (i.e., paragraph). Previous long-term motion generating approaches were mostly based on recurrent methods, using previously generated motion chunks as input for the next step. However, this approach has two drawbacks: 1) it relies on sequential datasets, which are expensive; 2) these methods yield unrealistic gaps between motions generated at each step. To address these issues, we introduce simple yet effective T2LM, a continuous long-term generation framework that can be trained without sequential data. T2LM comprises two components: a 1D-convolutional VQVAE, trained to compress motion to sequences of latent vectors, and a Transformer-based Text Encoder that predicts a latent sequence given an input text. At inference, a sequence of sentences is translated into a continuous stream of latent vectors. This is then decoded into a motion by the VQVAE decoder; the use of 1D convolutions with a local temporal receptive field avoids temporal inconsistencies between training and generated sequences. This simple constraint on the VQ-VAE allows it to be trained with short sequences only and produces smoother transitions. T2LM outperforms prior long-term generation models while overcoming the constraint of requiring sequential data; it is also competitive with SOTA single-action generation models.
Conventional 2D pose estimation models are constrained by their design to specific object categories. This limits their applicability to predefined objects. To overcome these limitations, category-agnostic pose estimation (CAPE) emerged as a solution. CAPE aims to facilitate keypoint localization for diverse object categories using a unified model, which can generalize from minimal annotated support images. Recent CAPE works have produced object poses based on arbitrary keypoint definitions annotated on a user-provided support image. Our work departs from conventional CAPE methods, which require a support image, by adopting a text-based approach instead of the support image. Specifically, we use a pose-graph, where nodes represent keypoints that are described with text. This representation takes advantage of the abstraction of text descriptions and the structure imposed by the graph. Our approach effectively breaks symmetry, preserves structure, and improves occlusion handling. We validate our novel approach using the MP-100 benchmark, a comprehensive dataset spanning over 100 categories and 18,000 images. Under a 1-shot setting, our solution achieves a notable performance boost of 1.07\%, establishing a new state-of-the-art for CAPE. Additionally, we enrich the dataset by providing text description annotations, further enhancing its utility for future research.
Current video-language models (VLMs) rely extensively on instance-level alignment between video and language modalities, which presents two major limitations: (1) visual reasoning disobeys the natural perception that humans do in first-person perspective, leading to a lack of reasoning interpretation; and (2) learning is limited in capturing inherent fine-grained relationships between two modalities. In this paper, we take an inspiration from human perception and explore a compositional approach for egocentric video representation. We introduce HENASY (Hierarchical ENtities ASsemblY), which includes a spatiotemporal token grouping mechanism to explicitly assemble dynamically evolving scene entities through time and model their relationship for video representation. By leveraging compositional structure understanding, HENASY possesses strong interpretability via visual grounding with free-form text queries. We further explore a suite of multi-grained contrastive losses to facilitate entity-centric understandings. This comprises three alignment types: video-narration, noun-entity, verb-entities alignments. Our method demonstrates strong interpretability in both quantitative and qualitative experiments; while maintaining competitive performances on five downstream tasks via zero-shot transfer or as video/text representation, including video/text retrieval, action recognition, multi-choice query, natural language query, and moments query.
Automatic prediction of fluorescently labeled organelles from label-free transmitted light input images is an important, yet difficult task. The traditional way to obtain fluorescence images is related to performing biochemical labeling which is time-consuming and costly. Therefore, an automatic algorithm to perform the task based on the label-free transmitted light microscopy could be strongly beneficial. The importance of the task motivated researchers from the France-BioImaging to organize the LightMyCells challenge where the goal is to propose an algorithm that automatically predicts the fluorescently labeled nucleus, mitochondria, tubulin, and actin, based on the input consisting of bright field, phase contrast, or differential interference contrast microscopic images. In this work, we present the contribution of the AGHSSO team based on a carefully prepared and trained encoder-decoder deep neural network that achieves a considerable score in the challenge, being placed among the best-performing teams.
Accurately predicting future pedestrian trajectories is crucial across various domains. Due to the uncertainty in future pedestrian trajectories, it is important to learn complex spatio-temporal representations in multi-agent scenarios. To address this, we propose a novel Cross-Correction Framework (CCF) to learn spatio-temporal representations of pedestrian trajectories better. Our framework consists of two trajectory prediction models, known as subnets, which share the same architecture and are trained with both cross-correction loss and trajectory prediction loss. Cross-correction leverages the learning from both subnets and enables them to refine their underlying representations of trajectories through a mutual correction mechanism. Specifically, we use the cross-correction loss to learn how to correct each other through an inter-subnet interaction. To induce diverse learning among the subnets, we use the transformed observed trajectories produced by a neural network as input to one subnet and the original observed trajectories as input to the other subnet. We utilize transformer-based encoder-decoder architecture for each subnet to capture motion and social interaction among pedestrians. The encoder of the transformer captures motion patterns in trajectories, while the decoder focuses on pedestrian interactions with neighbors. Each subnet performs the primary task of predicting future trajectories (a regression task) along with the secondary task of classifying the predicted trajectories (a classification task). Extensive experiments on real-world benchmark datasets such as ETH-UCY and SDD demonstrate the efficacy of our proposed framework, CCF, in precisely predicting pedestrian future trajectories. We also conducted several ablation experiments to demonstrate the effectiveness of various modules and loss functions used in our approach.
3D multiple object tracking (MOT) plays a crucial role in autonomous driving perception. Recent end-to-end query-based trackers simultaneously detect and track objects, which have shown promising potential for the 3D MOT task. However, existing methods are still in the early stages of development and lack systematic improvements, failing to track objects in certain complex scenarios, like occlusions and the small size of target object's situations. In this paper, we first summarize the current end-to-end 3D MOT framework by decomposing it into three constituent parts: query initialization, query propagation, and query matching. Then we propose corresponding improvements, which lead to a strong yet simple tracker: S2-Track. Specifically, for query initialization, we present 2D-Prompted Query Initialization, which leverages predicted 2D object and depth information to prompt an initial estimate of the object's 3D location. For query propagation, we introduce an Uncertainty-aware Probabilistic Decoder to capture the uncertainty of complex environment in object prediction with probabilistic attention. For query matching, we propose a Hierarchical Query Denoising strategy to enhance training robustness and convergence. As a result, our S2-Track achieves state-of-the-art performance on nuScenes benchmark, i.e., 66.3% AMOTA on test split, surpassing the previous best end-to-end solution by a significant margin of 8.9% AMOTA. We achieve 1st place on the nuScenes tracking task leaderboard.
Nutrition estimation is crucial for effective dietary management and overall health and well-being. Existing methods often struggle with sub-optimal accuracy and can be time-consuming. In this paper, we propose NuNet, a transformer-based network designed for nutrition estimation that utilizes both RGB and depth information from food images. We have designed and implemented a multi-scale encoder and decoder, along with two types of feature fusion modules, specialized for estimating five nutritional factors. These modules effectively balance the efficiency and effectiveness of feature extraction with flexible usage of our customized attention mechanisms and fusion strategies. Our experimental study shows that NuNet outperforms its variants and existing solutions significantly for nutrition estimation. It achieves an error rate of 15.65%, the lowest known to us, largely due to our multi-scale architecture and fusion modules. This research holds practical values for dietary management with huge potential for transnational research and deployment and could inspire other applications involving multiple data types with varying degrees of importance.
Speech in-painting is the task of regenerating missing audio contents using reliable context information. Despite various recent studies in multi-modal perception of audio in-painting, there is still a need for an effective infusion of visual and auditory information in speech in-painting. In this paper, we introduce a novel sequence-to-sequence model that leverages the visual information to in-paint audio signals via an encoder-decoder architecture. The encoder plays the role of a lip-reader for facial recordings and the decoder takes both encoder outputs as well as the distorted audio spectrograms to restore the original speech. Our model outperforms an audio-only speech in-painting model and has comparable results with a recent multi-modal speech in-painter in terms of speech quality and intelligibility metrics for distortions of 300 ms to 1500 ms duration, which proves the effectiveness of the introduced multi-modality in speech in-painting.
In recent years, a plethora of diverse methods have been proposed for 3D pose estimation. Among these, self-attention mechanisms and graph convolutions have both been proven to be effective and practical methods. Recognizing the strengths of those two techniques, we have developed a novel Semantic Graph Attention Network which can benefit from the ability of self-attention to capture global context, while also utilizing the graph convolutions to handle the local connectivity and structural constraints of the skeleton. We also design a Body Part Decoder that assists in extracting and refining the information related to specific segments of the body. Furthermore, our approach incorporates Distance Information, enhancing our model's capability to comprehend and accurately predict spatial relationships. Finally, we introduce a Geometry Loss who makes a critical constraint on the structural skeleton of the body, ensuring that the model's predictions adhere to the natural limits of human posture. The experimental results validate the effectiveness of our approach, demonstrating that every element within the system is essential for improving pose estimation outcomes. With comparison to state-of-the-art, the proposed work not only meets but exceeds the existing benchmarks.
This paper focuses on transferring control policies between robot manipulators with different morphology. While reinforcement learning (RL) methods have shown successful results in robot manipulation tasks, transferring a trained policy from simulation to a real robot or deploying it on a robot with different states, actions, or kinematics is challenging. To achieve cross-embodiment policy transfer, our key insight is to project the state and action spaces of the source and target robots to a common latent space representation. We first introduce encoders and decoders to associate the states and actions of the source robot with a latent space. The encoders, decoders, and a latent space control policy are trained simultaneously using loss functions measuring task performance, latent dynamics consistency, and encoder-decoder ability to reconstruct the original states and actions. To transfer the learned control policy, we only need to train target encoders and decoders that align a new target domain to the latent space. We use generative adversarial training with cycle consistency and latent dynamics losses without access to the task reward or reward tuning in the target domain. We demonstrate sim-to-sim and sim-to-real manipulation policy transfer with source and target robots of different states, actions, and embodiments. The source code is available at \url{https://github.com/ExistentialRobotics/cross_embodiment_transfer}.
Spectrum prediction is considered as a key technology to assist spectrum decision. Despite the great efforts that have been put on the construction of spectrum prediction, achieving accurate spectrum prediction emphasizes the need for more advanced solutions. In this paper, we propose a new multichannel multi-step spectrum prediction method using Transformer and stacked bidirectional LSTM (Bi- LSTM), named TSB. Specifically, we use multi-head attention and stacked Bi-LSTM to build a new Transformer based on encoder-decoder architecture. The self-attention mechanism composed of multiple layers of multi-head attention can continuously attend to all positions of the multichannel spectrum sequences. The stacked Bi-LSTM can learn these focused coding features by multi-head attention layer by layer. The advantage of this fusion mode is that it can deeply capture the long-term dependence of multichannel spectrum data. We have conducted extensive experiments on a dataset generated by a real simulation platform. The results show that the proposed algorithm performs better than the baselines.
Singing voice conversion is to convert the source singing voice into the target singing voice except for the content. Currently, flow-based models can complete the task of voice conversion, but they struggle to effectively extract latent variables in the more rhythmically rich and emotionally expressive task of singing voice conversion, while also facing issues with low efficiency in speech processing. In this paper, we propose a high-fidelity flow-based model based on multi-decoupling feature constraints called RASVC, which enhances the capture of vocal details by integrating multiple latent attribute encoders. We also use Multi-stream inverse short-time Fourier transform(MS-iSTFT) to enhance the speed of speech processing by skipping some complicated decoder processing steps. We compare the synthesized singing voice with other models from multiple dimensions, and our proposed model is highly consistent with the current state-of-the-art, with the demo which is available at \url{https://lazycat1119.github.io/RASVC-demo/}.
Estimating 3D full-body avatars from AR/VR devices is essential for creating immersive experiences in AR/VR applications. This task is challenging due to the limited input from Head Mounted Devices, which capture only sparse observations from the head and hands. Predicting the full-body avatars, particularly the lower body, from these sparse observations presents significant difficulties. In this paper, we are inspired by the inherent property of the kinematic tree defined in the Skinned Multi-Person Linear (SMPL) model, where the upper body and lower body share only one common ancestor node, bringing the potential of decoupled reconstruction. We propose a stratified approach to decouple the conventional full-body avatar reconstruction pipeline into two stages, with the reconstruction of the upper body first and a subsequent reconstruction of the lower body conditioned on the previous stage. To implement this straightforward idea, we leverage the latent diffusion model as a powerful probabilistic generator, and train it to follow the latent distribution of decoupled motions explored by a VQ-VAE encoder-decoder model. Extensive experiments on AMASS mocap dataset demonstrate our state-of-the-art performance in the reconstruction of full-body motions.
The iterated learning model is an agent model which simulates the transmission of of language from generation to generation. It is used to study how the language adapts to pressures imposed by transmission. In each iteration, a language tutor exposes a na\"ive pupil to a limited training set of utterances, each pairing a random meaning with the signal that conveys it. Then the pupil becomes a tutor for a new na\"ive pupil in the next iteration. The transmission bottleneck ensures that tutors must generalize beyond the training set that they experienced. Repeated cycles of learning and generalization can result in a language that is expressive, compositional and stable. Previously, the agents in the iterated learning model mapped signals to meanings using an artificial neural network but relied on an unrealistic and computationally expensive process of obversion to map meanings to signals. Here, both maps are neural networks, trained separately through supervised learning and together through unsupervised learning in the form of an autoencoder. This avoids the computational burden entailed in obversion and introduces a mixture of supervised and unsupervised learning as observed during language learning in children. The new model demonstrates a linear relationship between the dimensionality of meaning-signal space and effective bottleneck size and suggests that internal reflection on potential utterances is important in language learning and evolution.
We present new results to model and understand the role of encoder-decoder design in machine learning (ML) from an information-theoretic angle. We use two main information concepts, information sufficiency (IS) and mutual information loss (MIL), to represent predictive structures in machine learning. Our first main result provides a functional expression that characterizes the class of probabilistic models consistent with an IS encoder-decoder latent predictive structure. This result formally justifies the encoder-decoder forward stages many modern ML architectures adopt to learn latent (compressed) representations for classification. To illustrate IS as a realistic and relevant model assumption, we revisit some known ML concepts and present some interesting new examples: invariant, robust, sparse, and digital models. Furthermore, our IS characterization allows us to tackle the fundamental question of how much performance (predictive expressiveness) could be lost, using the cross entropy risk, when a given encoder-decoder architecture is adopted in a learning setting. Here, our second main result shows that a mutual information loss quantifies the lack of expressiveness attributed to the choice of a (biased) encoder-decoder ML design. Finally, we address the problem of universal cross-entropy learning with an encoder-decoder design where necessary and sufficiency conditions are established to meet this requirement. In all these results, Shannon's information measures offer new interpretations and explanations for representation learning.
Work zone is one of the major causes of non-recurrent traffic congestion and road incidents. Despite the significance of its impact, studies on predicting the traffic impact of work zones remain scarce. In this paper, we propose a data integration pipeline that enhances the utilization of work zone and traffic data from diversified platforms, and introduce a novel deep learning model to predict the traffic speed and incident likelihood during planned work zone events. The proposed model transforms traffic patterns into 2D space-time images for both model input and output and employs an attention-based multi-context convolutional encoder-decoder architecture to capture the spatial-temporal dependencies between work zone events and traffic variations. Trained and validated on four years of archived work zone traffic data from Maryland, USA, the model demonstrates superior performance over baseline models in predicting traffic speed, incident likelihood, and inferred traffic attributes such as queue length and congestion timings (i.e., start time and duration). Specifically, the proposed model outperforms the baseline models by reducing the prediction error of traffic speed by 5% to 34%, queue length by 11% to 29%, congestion timing by 6% to 17%, and increasing the accuracy of incident predictions by 5% to 7%. Consequently, this model offers substantial promise for enhancing the planning and traffic management of work zones.
Temporal logic is a framework for representing and reasoning about propositions that evolve over time. It is commonly used for specifying requirements in various domains, including hardware and software systems, as well as robotics. Specification mining or formula generation involves extracting temporal logic formulae from system traces and has numerous applications, such as detecting bugs and improving interpretability. Although there has been a surge of deep learning-based methods for temporal logic satisfiability checking in recent years, the specification mining literature has been lagging behind in adopting deep learning methods despite their many advantages, such as scalability. In this paper, we introduce autoregressive models that can generate linear temporal logic formulae from traces, towards addressing the specification mining problem. We propose multiple architectures for this task: transformer encoder-decoder, decoder-only transformer, and Mamba, which is an emerging alternative to transformer models. Additionally, we devise a metric for quantifying the distinctiveness of the generated formulae and a straightforward algorithm to enforce the syntax constraints. Our experiments show that the proposed architectures yield promising results, generating correct and distinct formulae at a fraction of the compute cost needed for the combinatorial baseline.