Enabling Ensemble Learning for Heterogeneous Large Language Models with Deep Parallel Collaboration

Large language models (LLMs) have shown complementary strengths in various tasks and instances, motivating the research of ensembling LLMs to push the frontier leveraging the wisdom of the crowd. Existing work achieves this objective via training the extra reward model or fusion model to select or fuse all candidate answers. However, these methods pose a great challenge to the generalizability of the trained models. Besides, existing methods use the textual responses as communication media, ignoring the rich information in the inner representations of neural networks. Therefore, we propose a training-free ensemble framework DEEPEN, averaging the probability distributions outputted by different LLMs. A key challenge in this paradigm is the vocabulary discrepancy between heterogeneous LLMs, which hinders the operation of probability distribution averaging. To address this challenge, DEEPEN maps the probability distribution of each model from the probability space to a universe relative space based on the relative representation theory, and performs aggregation. Then, the result of aggregation is mapped back to the probability space of one LLM via a search-based inverse transformation to determine the generated token. We conduct experiments on the ensemble of various LLMs of 6B to 70B. Experimental results show that DEEPEN achieves consistent improvements across six popular benchmarks involving subject examination, reasoning and knowledge-QA, proving the effectiveness of our approach.

RL in Markov Games with Independent Function Approximation: Improved Sample Complexity Bound under the Local Access Model

Efficiently learning equilibria with large state and action spaces in general-sum Markov games while overcoming the curse of multi-agency is a challenging problem. Recent works have attempted to solve this problem by employing independent linear function classes to approximate the marginal $Q$-value for each agent. However, existing sample complexity bounds under such a framework have a suboptimal dependency on the desired accuracy $\varepsilon$ or the action space. In this work, we introduce a new algorithm, Lin-Confident-FTRL, for learning coarse correlated equilibria (CCE) with local access to the simulator, i.e., one can interact with the underlying environment on the visited states. Up to a logarithmic dependence on the size of the state space, Lin-Confident-FTRL learns $\epsilon$-CCE with a provable optimal accuracy bound $O(\epsilon^{-2})$ and gets rids of the linear dependency on the action space, while scaling polynomially with relevant problem parameters (such as the number of agents and time horizon). Moreover, our analysis of Linear-Confident-FTRL generalizes the virtual policy iteration technique in the single-agent local planning literature, which yields a new computationally efficient algorithm with a tighter sample complexity bound when assuming random access to the simulator.

Domain Adversarial Active Learning for Domain Generalization Classification

Domain generalization models aim to learn cross-domain knowledge from source domain data, to improve performance on unknown target domains. Recent research has demonstrated that diverse and rich source domain samples can enhance domain generalization capability. This paper argues that the impact of each sample on the model's generalization ability varies. Despite its small scale, a high-quality dataset can still attain a certain level of generalization ability. Motivated by this, we propose a domain-adversarial active learning (DAAL) algorithm for classification tasks in domain generalization. First, we analyze that the objective of tasks is to maximize the inter-class distance within the same domain and minimize the intra-class distance across different domains. To achieve this objective, we design a domain adversarial selection method that prioritizes challenging samples. Second, we posit that even in a converged model, there are subsets of features that lack discriminatory power within each domain. We attempt to identify these feature subsets and optimize them by a constraint loss. We validate and analyze our DAAL algorithm on multiple domain generalization datasets, comparing it with various domain generalization algorithms and active learning algorithms. Our results demonstrate that the DAAL algorithm can achieve strong generalization ability with fewer data resources, thereby reducing data annotation costs in domain generalization tasks.

ZO-AdaMU Optimizer: Adapting Perturbation by the Momentum and Uncertainty in Zeroth-order Optimization

Lowering the memory requirement in full-parameter training on large models has become a hot research area. MeZO fine-tunes the large language models (LLMs) by just forward passes in a zeroth-order SGD optimizer (ZO-SGD), demonstrating excellent performance with the same GPU memory usage as inference. However, the simulated perturbation stochastic approximation for gradient estimate in MeZO leads to severe oscillations and incurs a substantial time overhead. Moreover, without momentum regularization, MeZO shows severe over-fitting problems. Lastly, the perturbation-irrelevant momentum on ZO-SGD does not improve the convergence rate. This study proposes ZO-AdaMU to resolve the above problems by adapting the simulated perturbation with momentum in its stochastic approximation. Unlike existing adaptive momentum methods, we relocate momentum on simulated perturbation in stochastic gradient approximation. Our convergence analysis and experiments prove this is a better way to improve convergence stability and rate in ZO-SGD. Extensive experiments demonstrate that ZO-AdaMU yields better generalization for LLMs fine-tuning across various NLP tasks than MeZO and its momentum variants.

ElasticLaneNet: A Geometry-Flexible Approach for Lane Detection

The task of lane detection involves identifying the boundaries of driving areas. Recognizing lanes with complex and variable geometric structures remains a challenge. In this paper, we introduce a new lane detection framework named ElasticLaneNet (Elastic-interaction-energy guided Lane detection Network). A novel and flexible way of representing lanes, namely, implicit representation is proposed. The training strategy considers predicted lanes as moving curves that being attracted to the ground truth guided by an elastic interaction energy based loss function (EIE loss). An auxiliary feature refinement (AFR) module is designed to gather information from different layers. The method performs well in complex lane scenarios, including those with large curvature, weak geometric features at intersections, complicated cross lanes, Y-shapes lanes, dense lanes, etc. We apply our approach on three datasets: SDLane, CULane, and TuSimple. The results demonstrate the exceptional performance of our method, with the state-of-the-art results on the structure-diversity dataset SDLane, achieving F1-score of 89.51, Recall rate of 87.50, and Precision of 91.61.

A Deep Representation Learning-based Speech Enhancement Method Using Complex Convolution Recurrent Variational Autoencoder

Generally, the performance of deep neural networks (DNNs) heavily depends on the quality of data representation learning. Our preliminary work has emphasized the significance of deep representation learning (DRL) in the context of speech enhancement (SE) applications. Specifically, our initial SE algorithm employed a gated recurrent unit variational autoencoder (VAE) with a Gaussian distribution to enhance the performance of certain existing SE systems. Building upon our preliminary framework, this paper introduces a novel approach for SE using deep complex convolutional recurrent networks with a VAE (DCCRN-VAE). DCCRN-VAE assumes that the latent variables of signals follow complex Gaussian distributions that are modeled by DCCRN, as these distributions can better capture the behaviors of complex signals. Additionally, we propose the application of a residual loss in DCCRN-VAE to further improve the quality of the enhanced speech. {Compared to our preliminary work, DCCRN-VAE introduces a more sophisticated DCCRN structure and probability distribution for DRL. Furthermore, in comparison to DCCRN, DCCRN-VAE employs a more advanced DRL strategy. The experimental results demonstrate that the proposed SE algorithm outperforms both our preliminary SE framework and the state-of-the-art DCCRN SE method in terms of scale-invariant signal-to-distortion ratio, speech quality, and speech intelligibility.

Confounder Balancing in Adversarial Domain Adaptation for Pre-Trained Large Models Fine-Tuning

The excellent generalization, contextual learning, and emergence abilities in the pre-trained large models (PLMs) handle specific tasks without direct training data, making them the better foundation models in the adversarial domain adaptation (ADA) methods to transfer knowledge learned from the source domain to target domains. However, existing ADA methods fail to account for the confounder properly, which is the root cause of the source data distribution that differs from the target domains. This study proposes an adversarial domain adaptation with confounder balancing for PLMs fine-tuning (ADA-CBF). The ADA-CBF includes a PLM as the foundation model for a feature extractor, a domain classifier and a confounder classifier, and they are jointly trained with an adversarial loss. This loss is designed to improve the domain-invariant representation learning by diluting the discrimination in the domain classifier. At the same time, the adversarial loss also balances the confounder distribution among source and unmeasured domains in training. Compared to existing ADA methods, ADA-CBF can correctly identify confounders in domain-invariant features, thereby eliminating the confounder biases in the extracted features from PLMs. The confounder classifier in ADA-CBF is designed as a plug-and-play and can be applied in the confounder measurable, unmeasurable, or partially measurable environments. Empirical results on natural language processing and computer vision downstream tasks show that ADA-CBF outperforms the newest GPT-4, LLaMA2, ViT and ADA methods.

Elastic Interaction Energy Loss for Traffic Image Segmentation

Segmentation is a pixel-level classification of images. The accuracy and fast inference speed of image segmentation are crucial for autonomous driving safety. Fine and complex geometric objects are the most difficult but important recognition targets in traffic scene, such as pedestrians, traffic signs and lanes. In this paper, a simple and efficient geometry-sensitive energy-based loss function is proposed to Convolutional Neural Network (CNN) for multi-class segmentation on real-time traffic scene understanding. To be specific, the elastic interaction energy (EIE) between two boundaries will drive the prediction moving toward the ground truth until completely overlap. The EIE loss function is incorporated into CNN to enhance accuracy on fine-scale structure segmentation. In particular, small or irregularly shaped objects can be identified more accurately, and discontinuity issues on slender objects can be improved. Our approach can be applied to different segmentation-based problems, such as urban scene segmentation and lane detection. We quantitatively and qualitatively analyze our method on three traffic datasets, including urban scene data Cityscapes, lane data TuSimple and CULane. The results show that our approach consistently improves performance, especially when using real-time, lightweight networks as the backbones, which is more suitable for autonomous driving.

Energy stable neural network for gradient flow equations

In this paper, we propose an energy stable network (EStable-Net) for solving gradient flow equations. The solution update scheme in our neural network EStable-Net is inspired by a proposed auxiliary variable based equivalent form of the gradient flow equation. EStable-Net enables decreasing of a discrete energy along the neural network, which is consistent with the property in the evolution process of the gradient flow equation. The architecture of the neural network EStable-Net consists of a few energy decay blocks, and the output of each block can be interpreted as an intermediate state of the evolution process of the gradient flow equation. This design provides a stable, efficient and interpretable network structure. Numerical experimental results demonstrate that our network is able to generate high accuracy and stable predictions.

Client-side Gradient Inversion Against Federated Learning from Poisoning

Federated Learning (FL) enables distributed participants (e.g., mobile devices) to train a global model without sharing data directly to a central server. Recent studies have revealed that FL is vulnerable to gradient inversion attack (GIA), which aims to reconstruct the original training samples and poses high risk against the privacy of clients in FL. However, most existing GIAs necessitate control over the server and rely on strong prior knowledge including batch normalization and data distribution information. In this work, we propose Client-side poisoning Gradient Inversion (CGI), which is a novel attack method that can be launched from clients. For the first time, we show the feasibility of a client-side adversary with limited knowledge being able to recover the training samples from the aggregated global model. We take a distinct approach in which the adversary utilizes a malicious model that amplifies the loss of a specific targeted class of interest. When honest clients employ the poisoned global model, the gradients of samples belonging to the targeted class are magnified, making them the dominant factor in the aggregated update. This enables the adversary to effectively reconstruct the private input belonging to other clients using the aggregated update. In addition, our CGI also features its ability to remain stealthy against Byzantine-robust aggregation rules (AGRs). By optimizing malicious updates and blending benign updates with a malicious replacement vector, our method remains undetected by these defense mechanisms. To evaluate the performance of CGI, we conduct experiments on various benchmark datasets, considering representative Byzantine-robust AGRs, and exploring diverse FL settings with different levels of adversary knowledge about the data. Our results demonstrate that CGI consistently and successfully extracts training input in all tested scenarios.