Recent advances in text-to-motion generation using dif-fusion and autoregressive models have shown promising re-sults. However, these models often suffer from a trade-off between real-time performance, high fidelity, and motion editability. To address this gap, we introduce MMM, a novel yet simple motion generation paradigm based on Masked Motion Model. MMM consists of two key components: (1) a motion tokenizer that transforms 3D human motion into a sequence of discrete tokens in latent space, and (2) a conditional masked motion transformer that learns to predict randomly masked motion tokens, conditioned on the pre-computed text tokens. By attending to motion and text to-kens in all directions, MMM explicitly captures inherent dependency among motion tokens and semantic mapping between motion and text tokens. During inference, this al-lows parallel and iterative decoding of multiple motion to-kens that are highly consistent with fine-grained text de-scriptions, therefore simultaneously achieving high-fidelity and high-speed motion generation. In addition, MMM has innate motion editability. By simply placing mask tokens in the place that needs editing, MMM automatically fills the gaps while guaranteeing smooth transitions between editing and non-editing parts. Extensive experiments on the HumanML3D and KIT-ML datasets demonstrate that MMM surpasses current leading methods in generating high-quality motion (evidenced by superior FID scores of 0.08 and 0.429), while offering advanced editing features such as body-part modification, motion in-betweening, and the synthesis of long motion sequences. In addition, MMM is two orders of magnitude faster on a single mid-range GPU than editable motion diffusion models. Our project page is available at https://exitudio.github.io/MMM-page/.

Most existing gait recognition methods are appearance-based, which rely on the silhouettes extracted from the video data of human walking activities. The less-investigated skeleton-based gait recognition methods directly learn the gait dynamics from 2D/3D human skeleton sequences, which are theoretically more robust solutions in the presence of appearance changes caused by clothes, hairstyles, and carrying objects. However, the performance of skeleton-based solutions is still largely behind the appearance-based ones. This paper aims to close such performance gap by proposing a novel network model, GaitMixer, to learn more discriminative gait representation from skeleton sequence data. In particular, GaitMixer follows a heterogeneous multi-axial mixer architecture, which exploits the spatial self-attention mixer followed by the temporal large-kernel convolution mixer to learn rich multi-frequency signals in the gait feature maps. Experiments on the widely used gait database, CASIA-B, demonstrate that GaitMixer outperforms the previous SOTA skeleton-based methods by a large margin while achieving a competitive performance compared with the representative appearance-based solutions. Code will be available at https://github.com/exitudio/gaitmixer

Most existing deep neural networks are static, which means they can only perform inference at a fixed complexity. But the resource budget can vary substantially across different devices. Even on a single device, the affordable budget can change with different scenarios, and repeatedly training networks for each required budget would be incredibly expensive. Therefore, in this work, we propose a general method called MutualNet to train a single network that can run at a diverse set of resource constraints. Our method trains a cohort of model configurations with various network widths and input resolutions. This mutual learning scheme not only allows the model to run at different width-resolution configurations but also transfers the unique knowledge among these configurations, helping the model to learn stronger representations overall. MutualNet is a general training methodology that can be applied to various network structures (e.g., 2D networks: MobileNets, ResNet, 3D networks: SlowFast, X3D) and various tasks (e.g., image classification, object detection, segmentation, and action recognition), and is demonstrated to achieve consistent improvements on a variety of datasets. Since we only train the model once, it also greatly reduces the training cost compared to independently training several models. Surprisingly, MutualNet can also be used to significantly boost the performance of a single network, if dynamic resource constraints are not a concern. In summary, MutualNet is a unified method for both static and adaptive, 2D and 3D networks. Code and pre-trained models are available at https://github.com/taoyang1122/MutualNet.

Federated learning (FL) is a promising strategy for performing privacy-preserving, distributed learning with a network of clients (i.e., edge devices). However, the data distribution among clients is often non-IID in nature, making efficient optimization difficult. To alleviate this issue, many FL algorithms focus on mitigating the effects of data heterogeneity across clients by introducing a variety of proximal terms, some incurring considerable compute and/or memory overheads, to restrain local updates with respect to the global model. Instead, we consider rethinking solutions to data heterogeneity in FL with a focus on local learning generality rather than proximal restriction. To this end, we first present a systematic study informed by second-order indicators to better understand algorithm effectiveness in FL. Interestingly, we find that standard regularization methods are surprisingly strong performers in mitigating data heterogeneity effects. Based on our findings, we further propose a simple and effective method, FedAlign, to overcome data heterogeneity and the pitfalls of previous methods. FedAlign achieves competitive accuracy with state-of-the-art FL methods across a variety of settings while minimizing computation and memory overhead. Code is available at https://github.com/mmendiet/FedAlign.

Requiring less data for accurate models, few-shot learning has shown robustness and generality in many application domains. However, deploying few-shot models in untrusted environments may inflict privacy concerns, e.g., attacks or adversaries that may breach the privacy of user-supplied data. This paper studies the privacy enhancement for the few-shot learning in an untrusted environment, e.g., the cloud, by establishing a novel privacy-preserved embedding space that preserves the privacy of data and maintains the accuracy of the model. We examine the impact of various image privacy methods such as blurring, pixelization, Gaussian noise, and differentially private pixelization (DP-Pix) on few-shot image classification and propose a method that learns privacy-preserved representation through the joint loss. The empirical results show how privacy-performance trade-off can be negotiated for privacy-enhanced few-shot learning.

A vast amount of unstructured text that contains valuable information is available over the web. This text is changing and proliferating, making it hard for people to process, read, and remember. Data mining and information extraction algorithms are used to develop new automation techniques to process the unstructured text. Among this publicly available text, there are a considerable amount of online medical articles, which provides valuable information about diseases, symptoms, operations, treatments, drugs, etc. Automatic unstructured text classification offers practical information management that does not depend on the subjective criteria of classification. It also provides useful information by obtaining and correlating relevant data present in documents. It also classifies, identifies and presents all sources of knowledge and reduces the time for retrieving information by simplifying access to content. Therefore, medical information needs to be classified into their respected categories (such as Diabetes, Cancer, Depression, Pediatrics, etc.). In this paper, we propose to use a deep learning approach for unstructured medical text classification at the document level. In our classification model we used two types of features: (i) content-based features (stylistic and complexity), and (ii) health domain-specific features. Moreover, rather than dealing with binary classification, this work handles multiclasses medical articles classification. This classification is done based on linguistic features that are extracted from the text, it also incorporates medical domain-specific terms/keywords as part of the classification feature set. These domain-specific features are extracted by applying topic modeling technique to spot the most probable terms for each medical class. Our experiments shows a reasonable classification accuracy for such a large number of classes.

Multi-hop networks (e.g., mesh, ad-hoc, and sensor networks) are important and cost-efficient communication backbones. Over the last few years wireless data traffic has drastically increased due to the changes in the way today's society creates, shares, and consumes information. This demands the efficient and intelligent utilization of limited network resources to optimize network performance. Traffic engineering (TE) optimizes network performance and enables optimal forwarding and routing rules to meet the quality of service (QoS) requirements for a large volume of traffic flows. This paper proposes a distributed model-free TE solution based on stochastic policy gradient reinforcement learning (RL), which aims to learn a stochastic routing policy for each router so that each router can send a packet to the next-hop router according to the learned optimal probability. The proposed policy-gradient solution naturally leads to multi-path TE strategies, which can effectively distribute the high traffic loads among all available routing paths to minimize the E2E delay. Moreover, a distributed software-defined networking architecture is proposed, which enables the fast prototyping of the proposed multi-agent actor-critic TE (MA-AC TE) algorithm and in-nature supports automated TE through multi-agent RL learning.

Deep reinforcement learning (DRL) has been leading to state-of-the-art performance to learn control policies for a wide range of applications. However, it does not provide an explanation of how a policy is learned and how the learned policy performs on a given task. In this paper, we answer to the inquiry: what scenes does a machine learning agent need to memorize for efficient learning and additional explanation regarding performance? Proposing a monitoring model to record the most important moments from experience-called snapshot images-we examine them for analysis. Sparse Bayesian Reinforcement Learning (SBRL) is known to remember sparse input samples during training and to construct bases for value function approximation. Also, SBRL has successfully maintained the snapshot memory for sparse input sampling. We apply our method to a visual maze problem and Atari games to observe the recorded snapshot images. Analyzing the images, we evaluate the efficacy of the proposed monitoring model and the quality of collected snapshots.

A reinforcement learning (RL) agent needs a fair amount of experience to find a near-optimal policy. Transfer learning has been investigated as a means to reduce the amount of experience required. Transfer learning, however, requires another similar reinforcement learning task as a transfer source, which can also be costly in the amount of experience required. In this research, we examine the possible “practice” approach that transfers knowledge from a non-RL task to a target RL task to avoid the expensive data sampling. We analyze how practice captures the distributions of state and action spaces in an environment. For this, we develop a novel learning approach that acquires important samples from practice and then applies them to a target RL task without changing learned bases. Results show an improved learning efficiency through practice in classical benchmark problems and limitations in OpenAI Gym problems.

To be applicable to real world problems, much reinforcement learning (RL) research has focused on continuous state spaces with function approximations. Some problems also require continuous actions, but searching for good actions in a continuous action space is problematic. This paper suggests a novel relevance vector sampling approach to action search in an RL framework with relevance vector machines (RVM-RL). We hypothesize that each relevance vector (RV) is placed on the modes of the value approximation surface as the learning converges. From the hypothesis, we select actions in RVs to maximize the estimated state-action values. We report the efficiency of the proposed approach by controlling a simulated octopus arm with RV-sampled actions.

Deep learning algorithms have recently appeared that pretrain hidden layers of neural networks in unsupervised ways, leading to state-of-the-art performance on large classification problems. These methods can also pretrain networks used for reinforcement learning. However, this ignores the additional information that exists in a reinforcement learning paradigm via the ongoing sequence of state, action, new state tuples. This paper demonstrates that learning a predictive model of state dynamics can result in a pretrained hidden layer structure that reduces the time needed to solve reinforcement learning problems.

We combine a convergent TD-learning method and direct continuous action search with neural networks for function approximation to obtain both stability and generalization over inexperienced state-action pairs. We extend linear Greedy-GQ to nonlinear neural networks for convergent learning. Direct continuous action search with back-propagation leads to efficient high-precision control. A high dimensional continuous state and action problem, octopus arm control, is examined to test the proposed algorithm. Comparing TD, linear Greedy-GQ, and nonlinear Greedy-GQ, we discuss how the correction term contributes to learning with nonlinear Greedy-GQ algorithm and how continuous action search contributes to learning speed and stability.