Modality-Inconsistent Continual Learning of Multimodal Large Language Models

In this paper, we introduce Modality-Inconsistent Continual Learning (MICL), a new continual learning scenario for Multimodal Large Language Models (MLLMs) that involves tasks with inconsistent modalities (image, audio, or video) and varying task types (captioning or question-answering). Unlike existing vision-only or modality-incremental settings, MICL combines modality and task type shifts, both of which drive catastrophic forgetting. To address these challenges, we propose MoInCL, which employs a Pseudo Targets Generation Module to mitigate forgetting caused by task type shifts in previously seen modalities. It also incorporates Instruction-based Knowledge Distillation to preserve the model's ability to handle previously learned modalities when new ones are introduced. We benchmark MICL using a total of six tasks and conduct experiments to validate the effectiveness of our proposed MoInCL. The experimental results highlight the superiority of MoInCL, showing significant improvements over representative and state-of-the-art continual learning baselines.

Enhancing Robustness of CLIP to Common Corruptions through Bimodal Test-Time Adaptation

Although open-vocabulary classification models like Contrastive Language Image Pretraining (CLIP) have demonstrated strong zero-shot learning capabilities, their robustness to common image corruptions remains poorly understood. Through extensive experiments, we show that zero-shot CLIP lacks robustness to common image corruptions at increasing severity levels during test-time, necessitating the adaptation of CLIP to unlabeled corrupted images using test-time adaptation (TTA). However, we found that existing TTA methods have severe limitations in adapting CLIP due to their unimodal nature. To address these limitations, we propose \framework, a bimodal TTA method specially designed to improve CLIP's robustness to common image corruptions. The key insight of our approach is not only to adapt the visual encoders for better image feature extraction but also to strengthen the alignment between image and text features by promoting a stronger association between the image class prototype, computed using pseudo-labels, and the corresponding text feature. We evaluate our approach on benchmark image corruption datasets and achieve state-of-the-art results in TTA for CLIP, specifically for domains involving image corruption. Particularly, with a ViT-B/16 vision backbone, we obtain mean accuracy improvements of 9.7%, 5.94%, and 5.12% for CIFAR-10C, CIFAR-100C, and ImageNet-C, respectively.

GradAlign for Training-free Model Performance Inference

Architecture plays an important role in deciding the performance of deep neural networks. However, the search for the optimal architecture is often hindered by the vast search space, making it a time-intensive process. Recently, a novel approach known as training-free neural architecture search (NAS) has emerged, aiming to discover the ideal architecture without necessitating extensive training. Training-free NAS leverages various indicators for architecture selection, including metrics such as the count of linear regions, the density of per-sample losses, and the stability of the finite-width Neural Tangent Kernel (NTK) matrix. Despite the competitive empirical performance of current training-free NAS techniques, they suffer from certain limitations, including inconsistent performance and a lack of deep understanding. In this paper, we introduce GradAlign, a simple yet effective method designed for inferring model performance without the need for training. At its core, GradAlign quantifies the extent of conflicts within per-sample gradients during initialization, as substantial conflicts hinder model convergence and ultimately result in worse performance. We evaluate GradAlign against established training-free NAS methods using standard NAS benchmarks, showing a better overall performance. Moreover, we show that the widely adopted metric of linear region count may not suffice as a dependable criterion for selecting network architectures during at initialization.

Continual Audio-Visual Sound Separation

In this paper, we introduce a novel continual audio-visual sound separation task, aiming to continuously separate sound sources for new classes while preserving performance on previously learned classes, with the aid of visual guidance. This problem is crucial for practical visually guided auditory perception as it can significantly enhance the adaptability and robustness of audio-visual sound separation models, making them more applicable for real-world scenarios where encountering new sound sources is commonplace. The task is inherently challenging as our models must not only effectively utilize information from both modalities in current tasks but also preserve their cross-modal association in old tasks to mitigate catastrophic forgetting during audio-visual continual learning. To address these challenges, we propose a novel approach named ContAV-Sep (\textbf{Cont}inual \textbf{A}udio-\textbf{V}isual Sound \textbf{Sep}aration). ContAV-Sep presents a novel Cross-modal Similarity Distillation Constraint (CrossSDC) to uphold the cross-modal semantic similarity through incremental tasks and retain previously acquired knowledge of semantic similarity in old models, mitigating the risk of catastrophic forgetting. The CrossSDC can seamlessly integrate into the training process of different audio-visual sound separation frameworks. Experiments demonstrate that ContAV-Sep can effectively mitigate catastrophic forgetting and achieve significantly better performance compared to other continual learning baselines for audio-visual sound separation. Code is available at: \url{https://github.com/weiguoPian/ContAV-Sep_NeurIPS2024}.

STONE: A Submodular Optimization Framework for Active 3D Object Detection

3D object detection is fundamentally important for various emerging applications, including autonomous driving and robotics. A key requirement for training an accurate 3D object detector is the availability of a large amount of LiDAR-based point cloud data. Unfortunately, labeling point cloud data is extremely challenging, as accurate 3D bounding boxes and semantic labels are required for each potential object. This paper proposes a unified active 3D object detection framework, for greatly reducing the labeling cost of training 3D object detector. Our framework is based on a novel formulation of submodular optimization, specifically tailored to the problem of active 3D object detection. In particular, we address two fundamental challenges associated with active 3D object detection: data imbalance and the need to cover the distribution of the data, including LiDAR-based point cloud data of varying difficulty levels. Extensive experiments demonstrate that our method achieves state-of-the-art performance with high computational efficiency compared to existing active learning methods.

Continual Distillation Learning

We study the problem of Continual Distillation Learning (CDL) that considers Knowledge Distillation (KD) in the Continual Learning (CL) setup. A teacher model and a student model need to learn a sequence of tasks, and the knowledge of the teacher model will be distilled to the student to improve the student model. We introduce a novel method named CDL-Prompt that utilizes prompt-based continual learning models to build the teacher-student model. We investigate how to utilize the prompts of the teacher model in the student model for knowledge distillation, and propose an attention-based prompt mapping scheme to use the teacher prompts for the student. We demonstrate that our method can be applied to different prompt-based continual learning models such as L2P, DualPrompt and CODA-Prompt to improve their performance using powerful teacher models. Although recent CL methods focus on prompt learning, we show that our method can be utilized to build efficient CL models using prompt-based knowledge distillation.

Adapting Pre-Trained Vision Models for Novel Instance Detection and Segmentation

Novel Instance Detection and Segmentation (NIDS) aims at detecting and segmenting novel object instances given a few examples of each instance. We propose a unified framework (NIDS-Net) comprising object proposal generation, embedding creation for both instance templates and proposal regions, and embedding matching for instance label assignment. Leveraging recent advancements in large vision methods, we utilize the Grounding DINO and Segment Anything Model (SAM) to obtain object proposals with accurate bounding boxes and masks. Central to our approach is the generation of high-quality instance embeddings. We utilize foreground feature averages of patch embeddings from the DINOv2 ViT backbone, followed by refinement through a weight adapter mechanism that we introduce. We show experimentally that our weight adapter can adjust the embeddings locally within their feature space and effectively limit overfitting. This methodology enables a straightforward matching strategy, resulting in significant performance gains. Our framework surpasses current state-of-the-art methods, demonstrating notable improvements of 22.3, 46.2, 10.3, and 24.0 in average precision (AP) across four detection datasets. In instance segmentation tasks on seven core datasets of the BOP challenge, our method outperforms the top RGB methods by 3.6 AP and remains competitive with the best RGB-D method. Code is available at: https://github.com/YoungSean/NIDS-Net

KITE: A Kernel-based Improved Transferability Estimation Method

Transferability estimation has emerged as an important problem in transfer learning. A transferability estimation method takes as inputs a set of pre-trained models and decides which pre-trained model can deliver the best transfer learning performance. Existing methods tackle this problem by analyzing the output of the pre-trained model or by comparing the pre-trained model with a probe model trained on the target dataset. However, neither is sufficient to provide reliable and efficient transferability estimations. In this paper, we present a novel perspective and introduce Kite, as a Kernel-based Improved Transferability Estimation method. Kite is based on the key observations that the separability of the pre-trained features and the similarity of the pre-trained features to random features are two important factors for estimating transferability. Inspired by kernel methods, Kite adopts centered kernel alignment as an effective way to assess feature separability and feature similarity. Kite is easy to interpret, fast to compute, and robust to the target dataset size. We evaluate the performance of Kite on a recently introduced large-scale model selection benchmark. The benchmark contains 8 source dataset, 6 target datasets and 4 architectures with a total of 32 pre-trained models. Extensive results show that Kite outperforms existing methods by a large margin for transferability estimation.

Wideband Beamforming for Near-Field Communications with Circular Arrays

The beamforming performance of the uniform circular array (UCA) in near-field wideband communication systems is investigated. Compared to uniform linear array (ULA), UCA exhibits uniform effective array aperture in all directions, thus enabling more users to benefit from near-field communications. In this paper, the unique beam squint effect in near-field wideband UCA systems is comprehensively analyzed in both the distance and angular domains. It is rigorously demonstrated that the beam focal point only exists at a specific frequency in wideband UCA systems, resulting in significant beamforming loss. To alleviate this unique beam squint effect, the true-time delay (TTD)-based beamforming architecture is exploited. In particular, two wideband beamforming optimization approaches leveraging TTD units are proposed. 1) Analytical approach: In this approach, the phase shifters (PSs) and the time delay of TTD units are designed based on the analytical formula for beamforming gain. Following this design, the minimum number of TTD units required to achieve a predetermined beamforming gain is quantified. 2) Joint-optimization approach: In this method, the PSs and the TTD units are jointly optimized under practical maximum delay constraints to approximate the optimal unconstrained analog beamformer. Specifically, an efficient alternating optimization algorithm is proposed, where the PSs and the TTD units are alternately updated using either the closed-form solution or the low-complexity linear search approach. Extensive numerical results demonstrate that 1) the proposed beamforming schemes effectively mitigate the beam squint effect, and 2) the joint-optimization approach outperforms the analytical approach in terms of array gain and achievable spectral efficiency.

PALM: Pushing Adaptive Learning Rate Mechanisms for Continual Test-Time Adaptation

Real-world vision models in dynamic environments face rapid shifts in domain distributions, leading to decreased recognition performance. Continual test-time adaptation (CTTA) directly adjusts a pre-trained source discriminative model to these changing domains using test data. A highly effective CTTA method involves applying layer-wise adaptive learning rates, and selectively adapting pre-trained layers. However, it suffers from the poor estimation of domain shift and the inaccuracies arising from the pseudo-labels. In this work, we aim to overcome these limitations by identifying layers through the quantification of model prediction uncertainty without relying on pseudo-labels. We utilize the magnitude of gradients as a metric, calculated by backpropagating the KL divergence between the softmax output and a uniform distribution, to select layers for further adaptation. Subsequently, for the parameters exclusively belonging to these selected layers, with the remaining ones frozen, we evaluate their sensitivity in order to approximate the domain shift, followed by adjusting their learning rates accordingly. Overall, this approach leads to a more robust and stable optimization than prior approaches. We conduct extensive image classification experiments on CIFAR-10C, CIFAR-100C, and ImageNet-C and demonstrate the efficacy of our method against standard benchmarks and prior methods.