Over the past decade, deep learning has proven to be a highly effective tool for learning meaningful features from raw data. However, it remains an open question how deep networks perform hierarchical feature learning across layers. In this work, we attempt to unveil this mystery by investigating the structures of intermediate features. Motivated by our empirical findings that linear layers mimic the roles of deep layers in nonlinear networks for feature learning, we explore how deep linear networks transform input data into output by investigating the output (i.e., features) of each layer after training in the context of multi-class classification problems. Toward this goal, we first define metrics to measure within-class compression and between-class discrimination of intermediate features, respectively. Through theoretical analysis of these two metrics, we show that the evolution of features follows a simple and quantitative pattern from shallow to deep layers when the input data is nearly orthogonal and the network weights are minimum-norm, balanced, and approximate low-rank: Each layer of the linear network progressively compresses within-class features at a geometric rate and discriminates between-class features at a linear rate with respect to the number of layers that data have passed through. To the best of our knowledge, this is the first quantitative characterization of feature evolution in hierarchical representations of deep linear networks. Empirically, our extensive experiments not only validate our theoretical results numerically but also reveal a similar pattern in deep nonlinear networks which aligns well with recent empirical studies. Moreover, we demonstrate the practical implications of our results in transfer learning. Our code is available at \url{https://github.com/Heimine/PNC_DLN}.
We study deep neural networks for the multi-label classification (MLab) task through the lens of neural collapse (NC). Previous works have been restricted to the multi-class classification setting and discovered a prevalent NC phenomenon comprising of the following properties for the last-layer features: (i) the variability of features within every class collapses to zero, (ii) the set of feature means form an equi-angular tight frame (ETF), and (iii) the last layer classifiers collapse to the feature mean upon some scaling. We generalize the study to multi-label learning, and prove for the first time that a generalized NC phenomenon holds with the "pick-all-label" formulation. Under the natural analog of the unconstrained feature model (UFM), we establish that the only global classifier of the pick-all-label cross entropy loss display the same ETF geometry which further collapse to multiplicity-1 feature class means. Besides, we discover a combinatorial property in generalized NC which is unique for multi-label learning that we call "tag-wise average" property, where the feature class-means of samples with multiple labels are scaled average of the feature class-means of single label tags. Theoretically, we establish global optimality result for the pick-all-label cross-entropy risk for the UFM. Additionally, We also provide empirical evidence to support our investigation into training deep neural networks on multi-label datasets, resulting in improved training efficiency.
Autonomous vehicles need to accomplish their tasks while interacting with human drivers in traffic. It is thus crucial to equip autonomous vehicles with artificial reasoning to better comprehend the intentions of the surrounding traffic, thereby facilitating the accomplishments of the tasks. In this work, we propose a behavioral model that encodes drivers' interacting intentions into latent social-psychological parameters. Leveraging a Bayesian filter, we develop a receding-horizon optimization-based controller for autonomous vehicle decision-making which accounts for the uncertainties in the interacting drivers' intentions. For online deployment, we design a neural network architecture based on the attention mechanism which imitates the behavioral model with online estimated parameter priors. We also propose a decision tree search algorithm to solve the decision-making problem online. The proposed behavioral model is then evaluated in terms of its capabilities for real-world trajectory prediction. We further conduct extensive evaluations of the proposed decision-making module, in forced highway merging scenarios, using both simulated environments and real-world traffic datasets. The results demonstrate that our algorithms can complete the forced merging tasks in various traffic conditions while ensuring driving safety.
Federated learning (FL) is emerging as a sought-after distributed machine learning architecture, offering the advantage of model training without direct exposure of raw data. With advancements in network infrastructure, FL has been seamlessly integrated into edge computing. However, the limited resources on edge devices introduce security vulnerabilities to FL in the context. While blockchain technology promises to bolster security, practical deployment on resource-constrained edge devices remains a challenge. Moreover, the exploration of FL with multiple aggregators in edge computing is still new in the literature. Addressing these gaps, we introduce the Blockchain-empowered Heterogeneous Multi-Aggregator Federated Learning Architecture (BMA-FL). We design a novel light-weight Byzantine consensus mechanism, namely PBCM, to enable secure and fast model aggregation and synchronization in BMA-FL. We also dive into the heterogeneity problem in BMA-FL that the aggregators are associated with varied number of connected trainers with Non-IID data distributions and diverse training speed. We proposed a multi-agent deep reinforcement learning algorithm to help aggregators decide the best training strategies. The experiments on real-word datasets demonstrate the efficiency of BMA-FL to achieve better models faster than baselines, showing the efficacy of PBCM and proposed deep reinforcement learning algorithm.
In the past few years, there has been an explosive surge in the use of machine learning (ML) techniques to address combinatorial optimization (CO) problems, especially mixed-integer linear programs (MILPs). Despite the achievements, the limited availability of real-world instances often leads to sub-optimal decisions and biased solver assessments, which motivates a suite of synthetic MILP instance generation techniques. However, existing methods either rely heavily on expert-designed formulations or struggle to capture the rich features of real-world instances. To tackle this problem, we propose G2MILP, which to the best of our knowledge is the first deep generative framework for MILP instances. Specifically, G2MILP represents MILP instances as bipartite graphs, and applies a masked variational autoencoder to iteratively corrupt and replace parts of the original graphs to generate new ones. The appealing feature of G2MILP is that it can learn to generate novel and realistic MILP instances without prior expert-designed formulations, while preserving the structures and computational hardness of real-world datasets, simultaneously. Thus the generated instances can facilitate downstream tasks for enhancing MILP solvers under limited data availability. We design a suite of benchmarks to evaluate the quality of the generated MILP instances. Experiments demonstrate that our method can produce instances that closely resemble real-world datasets in terms of both structures and computational hardness.
Following the success of GPT4, there has been a surge in interest in multimodal large language model (MLLM) research. This line of research focuses on developing general-purpose LLMs through fine-tuning pre-trained LLMs and vision models. However, catastrophic forgetting, a notorious phenomenon where the fine-tuned model fails to retain similar performance compared to the pre-trained model, still remains an inherent problem in multimodal LLMs (MLLM). In this paper, we introduce EMT: Evaluating MulTimodality for evaluating the catastrophic forgetting in MLLMs, by treating each MLLM as an image classifier. We first apply EMT to evaluate several open-source fine-tuned MLLMs and we discover that almost all evaluated MLLMs fail to retain the same performance levels as their vision encoders on standard image classification tasks. Moreover, we continue fine-tuning LLaVA, an MLLM and utilize EMT to assess performance throughout the fine-tuning. Interestingly, our results suggest that early-stage fine-tuning on an image dataset improves performance across other image datasets, by enhancing the alignment of text and visual features. However, as fine-tuning proceeds, the MLLMs begin to hallucinate, resulting in a significant loss of generalizability, even when the image encoder remains frozen. Our results suggest that MLLMs have yet to demonstrate performance on par with their vision models on standard image classification tasks and the current MLLM fine-tuning procedure still has room for improvement.
Understanding the intention of vehicles in the surrounding traffic is crucial for an autonomous vehicle to successfully accomplish its driving tasks in complex traffic scenarios such as highway forced merging. In this paper, we consider a behavioral model that incorporates both social behaviors and personal objectives of the interacting drivers. Leveraging this model, we develop a receding-horizon control-based decision-making strategy, that estimates online the other drivers' intentions using Bayesian filtering and incorporates predictions of nearby vehicles' behaviors under uncertain intentions. The effectiveness of the proposed decision-making strategy is demonstrated and evaluated based on simulation studies in comparison with a game theoretic controller and a real-world traffic dataset.
Foundation models (FMs), including large language models, have become increasingly popular due to their wide-ranging applicability and ability to understand human-like semantics. While previous research has explored the use of FMs in semantic communications to improve semantic extraction and reconstruction, the impact of these models on different system levels, considering computation and memory complexity, requires further analysis. This study focuses on integrating FMs at the effectiveness, semantic, and physical levels, using universal knowledge to profoundly transform system design. Additionally, it examines the use of compact models to balance performance and complexity, comparing three separate approaches that employ FMs. Ultimately, the study highlights unresolved issues in the field that need addressing.
Reconfigurable intelligent surfaces (RISs) have received extensive concern to improve the performance of wireless communication systems. In this paper, a subarray-based scheme is investigated in terms of its effects on ergodic spectral efficiency (SE) and energy efficiency (EE) in RIS-assisted systems. In this scheme, the adjacent elements divided into a subarray are controlled by one signal and share the same reflection coefficient. An upper bound of ergodic SE is derived and an optimal phase shift design is proposed for the subarray-based RIS. Based on the upper bound and optimal design, we obtain the maximum of the upper bound. In particular, we analytically evaluate the effect of the subarray-based RIS on EE since it reduces SE and power consumption simultaneously. Numerical results verify the tightness of the upper bound, demonstrate the effectiveness of the optimal phase shift design for the subarray-based RIS, and reveal the effects of the subarray-based scheme on SE and EE.