Humans learn by interacting with their environments and perceiving the outcomes of their actions. A landmark in artificial intelligence has been the development of deep reinforcement learning (dRL) algorithms capable of doing the same in video games, on par with or better than humans. However, it remains unclear whether the successes of dRL models reflect advances in visual representation learning, the effectiveness of reinforcement learning algorithms at discovering better policies, or both. To address this question, we introduce the Learning Challenge Diagnosticator (LCD), a tool that separately measures the perceptual and reinforcement learning demands of a task. We use LCD to discover a novel taxonomy of challenges in the Procgen benchmark, and demonstrate that these predictions are both highly reliable and can instruct algorithmic development. More broadly, the LCD reveals multiple failure cases that can occur when optimizing dRL algorithms over entire video game benchmarks like Procgen, and provides a pathway towards more efficient progress.
Variational Autoencoders (VAEs) have become a cornerstone in generative modeling and representation learning within machine learning. This paper explores a nuanced aspect of VAEs, focusing on interpreting the Kullback-Leibler (KL) Divergence, a critical component within the Evidence Lower Bound (ELBO) that governs the trade-off between reconstruction accuracy and regularization. Meanwhile, the KL Divergence enforces alignment between latent variable distributions and a prior imposing a structure on the overall latent space but leaves individual variable distributions unconstrained. The proposed method redefines the ELBO with a mixture of Gaussians for the posterior probability, introduces a regularization term to prevent variance collapse, and employs a PatchGAN discriminator to enhance texture realism. Implementation details involve ResNetV2 architectures for both the Encoder and Decoder. The experiments demonstrate the ability to generate realistic faces, offering a promising solution for enhancing VAE-based generative models.
Single-stage multi-person human pose estimation (MPPE) methods have shown great performance improvements, but existing methods fail to disentangle features by individual instances under crowded scenes. In this paper, we propose a bounding box-level instance representation learning called BoIR, which simultaneously solves instance detection, instance disentanglement, and instance-keypoint association problems. Our new instance embedding loss provides a learning signal on the entire area of the image with bounding box annotations, achieving globally consistent and disentangled instance representation. Our method exploits multi-task learning of bottom-up keypoint estimation, bounding box regression, and contrastive instance embedding learning, without additional computational cost during inference. BoIR is effective for crowded scenes, outperforming state-of-the-art on COCO val (0.8 AP), COCO test-dev (0.5 AP), CrowdPose (4.9 AP), and OCHuman (3.5 AP). Code will be available at https://github.com/uyoung-jeong/BoIR
Understanding indoor scenes is crucial for urban studies. Considering the dynamic nature of indoor environments, effective semantic segmentation requires both real-time operation and high accuracy.To address this, we propose AsymFormer, a novel network that improves real-time semantic segmentation accuracy using RGB-D multi-modal information without substantially increasing network complexity. AsymFormer uses an asymmetrical backbone for multimodal feature extraction, reducing redundant parameters by optimizing computational resource distribution. To fuse asymmetric multimodal features, a Local Attention-Guided Feature Selection (LAFS) module is used to selectively fuse features from different modalities by leveraging their dependencies. Subsequently, a Cross-Modal Attention-Guided Feature Correlation Embedding (CMA) module is introduced to further extract cross-modal representations. The AsymFormer demonstrates competitive results with 54.1% mIoU on NYUv2 and 49.1% mIoU on SUNRGBD. Notably, AsymFormer achieves an inference speed of 65 FPS (79 FPS after implementing mixed precision quantization) on RTX3090, demonstrating that AsymFormer can strike a balance between high accuracy and efficiency.
Recent years have witnessed significant advancements in self-supervised learning (SSL) methods for speech-processing tasks. Various speech-based SSL models have been developed and present promising performance on a range of downstream tasks including speech recognition. However, existing speech-based SSL models face a common dilemma in terms of computational cost, which might hinder their potential application and in-depth academic research. To address this issue, we first analyze the computational cost of different modules during HuBERT pre-training and then introduce a stack of efficiency optimizations, which is named Fast-HuBERT in this paper. The proposed Fast-HuBERT can be trained in 1.1 days with 8 V100 GPUs on the Librispeech 960h benchmark, without performance degradation, resulting in a 5.2x speedup, compared to the original implementation. Moreover, we explore two well-studied techniques in the Fast-HuBERT and demonstrate consistent improvements as reported in previous work.
Self-supervised representation learning (SSRL) has demonstrated superior performance than supervised models for tasks including phoneme recognition. Training SSRL models poses a challenge for low-resource languages where sufficient pre-training data may not be available. A common approach is cross-lingual pre-training. Instead, we propose to use audio augmentation techniques, namely: pitch variation, noise addition, accented target language and other language speech to pre-train SSRL models in a low resource condition and evaluate phoneme recognition. Our comparisons found that a combined synthetic augmentations (noise/pitch) strategy outperformed accent and language knowledge transfer. Furthermore, we examined the scaling factor of augmented data to achieve equivalent performance to model pre-trained with target domain speech. Our findings suggest that for resource-constrained languages, combined augmentations can be a viable option than other augmentations.
How to accurately learn task-relevant state representations from high-dimensional observations with visual distractions is a realistic and challenging problem in visual reinforcement learning. Recently, unsupervised representation learning methods based on bisimulation metrics, contrast, prediction, and reconstruction have shown the ability for task-relevant information extraction. However, due to the lack of appropriate mechanisms for the extraction of task information in the prediction, contrast, and reconstruction-related approaches and the limitations of bisimulation-related methods in domains with sparse rewards, it is still difficult for these methods to be effectively extended to environments with distractions. To alleviate these problems, in the paper, the action sequences, which contain task-intensive signals, are incorporated into representation learning. Specifically, we propose a Sequential Action--induced invariant Representation (SAR) method, in which the encoder is optimized by an auxiliary learner to only preserve the components that follow the control signals of sequential actions, so the agent can be induced to learn the robust representation against distractions. We conduct extensive experiments on the DeepMind Control suite tasks with distractions while achieving the best performance over strong baselines. We also demonstrate the effectiveness of our method at disregarding task-irrelevant information by deploying SAR to real-world CARLA-based autonomous driving with natural distractions. Finally, we provide the analysis results of generalization drawn from the generalization decay and t-SNE visualization. Code and demo videos are available at https://github.com/DMU-XMU/SAR.git.
Data-hungry HSI classification methods require high-quality labeled HSIs, which are often costly to obtain. This characteristic limits the performance potential of data-driven methods when dealing with limited annotated samples. Bridging the domain gap between data acquired from different sensors allows us to utilize abundant labeled data across sensors to break this bottleneck. In this paper, we propose a novel Attention-Gated Tuning (AGT) strategy and a triplet-structured transformer model, Tri-Former, to address this issue. The AGT strategy serves as a bridge, allowing us to leverage existing labeled HSI datasets, even RGB datasets to enhance the performance on new HSI datasets with limited samples. Instead of inserting additional parameters inside the basic model, we train a lightweight auxiliary branch that takes intermediate features as input from the basic model and makes predictions. The proposed AGT resolves conflicts between heterogeneous and even cross-modal data by suppressing the disturbing information and enhances the useful information through a soft gate. Additionally, we introduce Tri-Former, a triplet-structured transformer with a spectral-spatial separation design that enhances parameter utilization and computational efficiency, enabling easier and flexible fine-tuning. Comparison experiments conducted on three representative HSI datasets captured by different sensors demonstrate the proposed Tri-Former achieves better performance compared to several state-of-the-art methods. Homologous, heterologous and cross-modal tuning experiments verified the effectiveness of the proposed AGT. Code has been released at: \href{https://github.com/Cecilia-xue/AGT}{https://github.com/Cecilia-xue/AGT}.
To improve word representation learning, we propose a probabilistic prior which can be seamlessly integrated with word embedding models. Different from previous methods, word embedding is taken as a probabilistic generative model, and it enables us to impose a prior regularizing word representation learning. The proposed prior not only enhances the representation of embedding vectors but also improves the model's robustness and stability. The structure of the proposed prior is simple and effective, and it can be easily implemented and flexibly plugged in most existing word embedding models. Extensive experiments show the proposed method improves word representation on various tasks.
Self-supervised representation learning for human action recognition has developed rapidly in recent years. Most of the existing works are based on skeleton data while using a multi-modality setup. These works overlooked the differences in performance among modalities, which led to the propagation of erroneous knowledge between modalities while only three fundamental modalities, i.e., joints, bones, and motions are used, hence no additional modalities are explored. In this work, we first propose an Implicit Knowledge Exchange Module (IKEM) which alleviates the propagation of erroneous knowledge between low-performance modalities. Then, we further propose three new modalities to enrich the complementary information between modalities. Finally, to maintain efficiency when introducing new modalities, we propose a novel teacher-student framework to distill the knowledge from the secondary modalities into the mandatory modalities considering the relationship constrained by anchors, positives, and negatives, named relational cross-modality knowledge distillation. The experimental results demonstrate the effectiveness of our approach, unlocking the efficient use of skeleton-based multi-modality data. Source code will be made publicly available at https://github.com/desehuileng0o0/IKEM.
Self-supervised representation learning has seen remarkable progress in the last few years, with some of the recent methods being able to learn useful image representations without labels. These methods are trained using backpropagation, the de facto standard. Recently, Geoffrey Hinton proposed the forward-forward algorithm as an alternative training method. It utilizes two forward passes and a separate loss function for each layer to train the network without backpropagation. In this study, for the first time, we study the performance of forward-forward vs. backpropagation for self-supervised representation learning and provide insights into the learned representation spaces. Our benchmark employs four standard datasets, namely MNIST, F-MNIST, SVHN and CIFAR-10, and three commonly used self-supervised representation learning techniques, namely rotation, flip and jigsaw. Our main finding is that while the forward-forward algorithm performs comparably to backpropagation during (self-)supervised training, the transfer performance is significantly lagging behind in all the studied settings. This may be caused by a combination of factors, including having a loss function for each layer and the way the supervised training is realized in the forward-forward paradigm. In comparison to backpropagation, the forward-forward algorithm focuses more on the boundaries and drops part of the information unnecessary for making decisions which harms the representation learning goal. Further investigation and research are necessary to stabilize the forward-forward strategy for self-supervised learning, to work beyond the datasets and configurations demonstrated by Geoffrey Hinton.
As 3D human pose estimation can now be achieved with very high accuracy in the supervised learning scenario, tackling the case where 3D pose annotations are not available has received increasing attention. In particular, several methods have proposed to learn image representations in a self-supervised fashion so as to disentangle the appearance information from the pose one. The methods then only need a small amount of supervised data to train a pose regressor using the pose-related latent vector as input, as it should be free of appearance information. In this paper, we carry out in-depth analysis to understand to what degree the state-of-the-art disentangled representation learning methods truly separate the appearance information from the pose one. First, we study disentanglement from the perspective of the self-supervised network, via diverse image synthesis experiments. Second, we investigate disentanglement with respect to the 3D pose regressor following an adversarial attack perspective. Specifically, we design an adversarial strategy focusing on generating natural appearance changes of the subject, and against which we could expect a disentangled network to be robust. Altogether, our analyses show that disentanglement in the three state-of-the-art disentangled representation learning frameworks if far from complete, and that their pose codes contain significant appearance information. We believe that our approach provides a valuable testbed to evaluate the degree of disentanglement of pose from appearance in self-supervised 3D human pose estimation.
This paper presents BELT, a novel model and learning framework for the pivotal topic of brain-to-language translation research. The translation from noninvasive brain signals into readable natural language has the potential to promote the application scenario as well as the development of brain-computer interfaces (BCI) as a whole. The critical problem in brain signal decoding or brain-to-language translation is the acquisition of semantically appropriate and discriminative EEG representation from a dataset of limited scale and quality. The proposed BELT method is a generic and efficient framework that bootstraps EEG representation learning using off-the-shelf large-scale pretrained language models (LMs). With a large LM's capacity for understanding semantic information and zero-shot generalization, BELT utilizes large LMs trained on Internet-scale datasets to bring significant improvements to the understanding of EEG signals. In particular, the BELT model is composed of a deep conformer encoder and a vector quantization encoder. Semantical EEG representation is achieved by a contrastive learning step that provides natural language supervision. We achieve state-of-the-art results on two featuring brain decoding tasks including the brain-to-language translation and zero-shot sentiment classification. Specifically, our model surpasses the baseline model on both tasks by 5.45% and over 10% and archives a 42.31% BLEU-1 score and 67.32% precision on the main evaluation metrics for translation and zero-shot sentiment classification respectively.
Choosing an appropriate representation of the environment for the underlying decision-making process of the reinforcement learning agent is not always straightforward. The state representation should be inclusive enough to allow the agent to informatively decide on its actions and disentangled enough to simplify policy training and the corresponding sim2real transfer. Given this outlook, this work examines the effect of various representations in incentivizing the agent to solve a specific robotic task: antipodal and planar object grasping. A continuum of state representations is defined, starting from hand-crafted numerical states to encoded image-based representations, with decreasing levels of induced task-specific knowledge. The effects of each representation on the ability of the agent to solve the task in simulation and the transferability of the learned policy to the real robot are examined and compared against a model-based approach with complete system knowledge. The results show that reinforcement learning agents using numerical states can perform on par with non-learning baselines. Furthermore, we find that agents using image-based representations from pre-trained environment embedding vectors perform better than end-to-end trained agents, and hypothesize that separation of representation learning from reinforcement learning can benefit sim2real transfer. Finally, we conclude that incentivizing the state representation with task-specific knowledge facilitates faster convergence for agent training and increases success rates in sim2real robot control.
Discrete audio representation, aka audio tokenization, has seen renewed interest driven by its potential to facilitate the application of text language modeling approaches in audio domain. To this end, various compression and representation-learning based tokenization schemes have been proposed. However, there is limited investigation into the performance of compression-based audio tokens compared to well-established mel-spectrogram features across various speaker and speech related tasks. In this paper, we evaluate compression based audio tokens on three tasks: Speaker Verification, Diarization and (Multi-lingual) Speech Recognition. Our findings indicate that (i) the models trained on audio tokens perform competitively, on average within $1\%$ of mel-spectrogram features for all the tasks considered, and do not surpass them yet. (ii) these models exhibit robustness for out-of-domain narrowband data, particularly in speaker tasks. (iii) audio tokens allow for compression to 20x compared to mel-spectrogram features with minimal loss of performance in speech and speaker related tasks, which is crucial for low bit-rate applications, and (iv) the examined Residual Vector Quantization (RVQ) based audio tokenizer exhibits a low-pass frequency response characteristic, offering a plausible explanation for the observed results, and providing insight for future tokenizer designs.
Recently, the AI community has made significant strides in developing powerful foundation models, driven by large-scale multimodal datasets. However, for audio representation learning, existing datasets suffer from limitations in the following aspects: insufficient volume, simplistic content, and arduous collection procedures. To establish an audio dataset with high-quality captions, we propose an innovative, automatic approach leveraging multimodal inputs, such as video frames, audio streams. Specifically, we construct a large-scale, high-quality, audio-language dataset, named as Auto-ACD, comprising over 1.5M audio-text pairs. We exploit a series of pre-trained models or APIs, to determine audio-visual synchronisation, generate image captions, object detection, or audio tags for specific videos. Subsequently, we employ LLM to paraphrase a congruent caption for each audio, guided by the extracted multi-modality clues. To demonstrate the effectiveness of the proposed dataset, we train widely used models on our dataset and show performance improvement on various downstream tasks, for example, audio-language retrieval, audio captioning, zero-shot classification. In addition, we establish a novel benchmark with environmental information and provide a benchmark for audio-text tasks.
Few-shot point cloud semantic segmentation aims to train a model to quickly adapt to new unseen classes with only a handful of support set samples. However, the noise-free assumption in the support set can be easily violated in many practical real-world settings. In this paper, we focus on improving the robustness of few-shot point cloud segmentation under the detrimental influence of noisy support sets during testing time. To this end, we first propose a Component-level Clean Noise Separation (CCNS) representation learning to learn discriminative feature representations that separates the clean samples of the target classes from the noisy samples. Leveraging the well separated clean and noisy support samples from our CCNS, we further propose a Multi-scale Degree-based Noise Suppression (MDNS) scheme to remove the noisy shots from the support set. We conduct extensive experiments on various noise settings on two benchmark datasets. Our results show that the combination of CCNS and MDNS significantly improves the performance. Our code is available at https://github.com/Pixie8888/R3DFSSeg.
Unsupervised sentence representation learning aims to transform input sentences into fixed-length vectors enriched with intricate semantic information while obviating the reliance on labeled data. Recent strides within this domain have been significantly propelled by breakthroughs in contrastive learning and prompt engineering. Despite these advancements, the field has reached a plateau, leading some researchers to incorporate external components to enhance the quality of sentence embeddings. Such integration, though beneficial, complicates solutions and inflates demands for computational resources. In response to these challenges, this paper presents CoT-BERT, an innovative method that harnesses the progressive thinking of Chain-of-Thought reasoning to tap into the latent potential of pre-trained models like BERT. Additionally, we develop an advanced contrastive learning loss function and propose a novel template denoising strategy. Rigorous experimentation demonstrates that CoT-BERT surpasses a range of well-established baselines by relying exclusively on the intrinsic strengths of pre-trained models.
Federated Learning (FL) is currently one of the most popular technologies in the field of Artificial Intelligence (AI) due to its collaborative learning and ability to preserve client privacy. However, it faces challenges such as non-identically and non-independently distributed (non-IID) and data with imbalanced labels among local clients. To address these limitations, the research community has explored various approaches such as using local model parameters, federated generative adversarial learning, and federated representation learning. In our study, we propose a novel Clustered FedStack framework based on the previously published Stacked Federated Learning (FedStack) framework. The local clients send their model predictions and output layer weights to a server, which then builds a robust global model. This global model clusters the local clients based on their output layer weights using a clustering mechanism. We adopt three clustering mechanisms, namely K-Means, Agglomerative, and Gaussian Mixture Models, into the framework and evaluate their performance. We use Bayesian Information Criterion (BIC) with the maximum likelihood function to determine the number of clusters. The Clustered FedStack models outperform baseline models with clustering mechanisms. To estimate the convergence of our proposed framework, we use Cyclical learning rates.
Unsupervised multi-view representation learning has been extensively studied for mining multi-view data. However, some critical challenges remain. On the one hand, the existing methods cannot explore multi-view data comprehensively since they usually learn a common representation between views, given that multi-view data contains both the common information between views and the specific information within each view. On the other hand, to mine the nonlinear relationship between data, kernel or neural network methods are commonly used for multi-view representation learning. However, these methods are lacking in interpretability. To this end, this paper proposes a new multi-view fuzzy representation learning method based on the interpretable Takagi-Sugeno-Kang (TSK) fuzzy system (MVRL_FS). The method realizes multi-view representation learning from two aspects. First, multi-view data are transformed into a high-dimensional fuzzy feature space, while the common information between views and specific information of each view are explored simultaneously. Second, a new regularization method based on L_(2,1)-norm regression is proposed to mine the consistency information between views, while the geometric structure of the data is preserved through the Laplacian graph. Finally, extensive experiments on many benchmark multi-view datasets are conducted to validate the superiority of the proposed method.