Riemannian meta-optimization provides a promising approach to solving non-linear constrained optimization problems, which trains neural networks as optimizers to perform optimization on Riemannian manifolds. However, existing Riemannian meta-optimization methods take up huge memory footprints in large-scale optimization settings, as the learned optimizer can only adapt gradients of a fixed size and thus cannot be shared across different Riemannian parameters. In this paper, we propose an efficient Riemannian meta-optimization method that significantly reduces the memory burden for large-scale optimization via a subspace adaptation scheme. Our method trains neural networks to individually adapt the row and column subspaces of Riemannian gradients, instead of directly adapting the full gradient matrices in existing Riemannian meta-optimization methods. In this case, our learned optimizer can be shared across Riemannian parameters with different sizes. Our method reduces the model memory consumption by six orders of magnitude when optimizing an orthogonal mainstream deep neural network (e.g., ResNet50). Experiments on multiple Riemannian tasks show that our method can not only reduce the memory consumption but also improve the performance of Riemannian meta-optimization.

Query: I want to buy a PS5 for each child in the photo. Thought: Use the `facedetection` tool to detect Thought: First analyze the image 1 to find the number human faces in the two images. Faces in Image 1: 4 bounding boxes Thought: There are 4 children in total. The price of Price of PS5: $479.99 a PS5 is approximately $500, so the cost is 4* 500. Thought: Using the price of $479.99 for each console. Query: The men in the picture want to buy one NVIDIA GeForce RTX 4070 SUPER each. According to the price in January, how many dollars will they need to spend in total? Observation: This image does not provide any price. On January 8, 2024, Nvidia released the RTX Thought: I cannot obtain useful information. I 4070 SUPER at $599, think the price is about $1800 for three men. Thought: The price is $599. Our agent chooses more precise tools based on the given files and intermediate observations. The advancement of large language models (LLMs) prompts the development of multi-modal agents, which are used as a controller to call external tools, providing a feasible way to solve practical tasks. In this paper, we propose a multi-modal agent tuning method that automatically generates multi-modal tool-usage data and tunes a vision-language model (VLM) as the controller for powerful tool-usage reasoning. To preserve the data quality, we prompt the GPT-4o mini model to generate queries, files, and trajectories, followed by query-file and trajectory verifiers. Based on the data synthesis pipeline, we collect the MM-Traj dataset that contains 20K tasks with trajectories of tool usage. Then, we develop the T3-Agent via Trajectory Tuning on VLMs for Tool usage using MM-Traj. Evaluations on the GTA and GAIA benchmarks show that the T3-Agent consistently achieves improvements on two popular VLMs: MiniCPM-V-8.5B Integrating external tools to solve diverse multi-modal tasks is a promising research direction towards multi-modal agents (Surís et al., 2023; Gupta & Kembhavi, 2023; Gao et al., 2024; Yuan et al., 2024; Zhong et al., 2023). Existing agents usually use a large language model (LLM) as the controller that generates plans via prompt engineering to call tools, achieving impressive performance in multiple domains, such as image editing (Wu et al., 2023), robotic manipulation (ichter et al., 2023), question answering (Shen et al., 2024), video understanding (Fan et al., 2024), and desktop APPs (Trivedi et al., 2024). Despite their success, prompt engineering faces limited reasoning abilities for tool usage in tackling practical tasks, as shown in Figure 1.

Compositional generalization is the capability of a model to understand novel compositions composed of seen concepts. There are multiple levels of novel compositions including phrase-phrase level, phrase-word level, and word-word level. Existing methods achieve promising compositional generalization, but the consistency of compositional generalization across multiple levels of novel compositions remains unexplored. The consistency refers to that a model should generalize to a phrase-phrase level novel composition, and phrase-word/word-word level novel compositions that can be derived from it simultaneously. In this paper, we propose a meta-learning based framework, for achieving consistent compositional generalization across multiple levels. The basic idea is to progressively learn compositions from simple to complex for consistency. Specifically, we divide the original training set into multiple validation sets based on compositional complexity, and introduce multiple meta-weight-nets to generate sample weights for samples in different validation sets. To fit the validation sets in order of increasing compositional complexity, we optimize the parameters of each meta-weight-net independently and sequentially in a multilevel optimization manner. We build a GQA-CCG dataset to quantitatively evaluate the consistency. Experimental results on visual question answering and temporal video grounding, demonstrate the effectiveness of the proposed framework. We release GQA-CCG at https://github.com/NeverMoreLCH/CCG.

Hyperbolic graph convolutional networks (HGCNs) have demonstrated representational capabilities of modeling hierarchical-structured graphs. However, as in general GCNs, over-smoothing may occur as the number of model layers increases, limiting the representation capabilities of most current HGCN models. In this paper, we propose residual hyperbolic graph convolutional networks (R-HGCNs) to address the over-smoothing problem. We introduce a hyperbolic residual connection function to overcome the over-smoothing problem, and also theoretically prove the effectiveness of the hyperbolic residual function. Moreover, we use product manifolds and HyperDrop to facilitate the R-HGCNs. The distinctive features of the R-HGCNs are as follows: (1) The hyperbolic residual connection preserves the initial node information in each layer and adds a hyperbolic identity mapping to prevent node features from being indistinguishable. (2) Product manifolds in R-HGCNs have been set up with different origin points in different components to facilitate the extraction of feature information from a wider range of perspectives, which enhances the representing capability of R-HGCNs. (3) HyperDrop adds multiplicative Gaussian noise into hyperbolic representations, such that perturbations can be added to alleviate the over-fitting problem without deconstructing the hyperbolic geometry. Experiment results demonstrate the effectiveness of R-HGCNs under various graph convolution layers and different structures of product manifolds.

Transformers achieve promising performance in document understanding because of their high effectiveness and still suffer from quadratic computational complexity dependency on the sequence length. General efficient transformers are challenging to be directly adapted to model document. They are unable to handle the layout representation in documents, e.g. word, line and paragraph, on different granularity levels and seem hard to achieve a good trade-off between efficiency and performance. To tackle the concerns, we propose Fast-StrucTexT, an efficient multi-modal framework based on the StrucTexT algorithm with an hourglass transformer architecture, for visual document understanding. Specifically, we design a modality-guided dynamic token merging block to make the model learn multi-granularity representation and prunes redundant tokens. Additionally, we present a multi-modal interaction module called Symmetry Cross Attention (SCA) to consider multi-modal fusion and efficiently guide the token mergence. The SCA allows one modality input as query to calculate cross attention with another modality in a dual phase. Extensive experiments on FUNSD, SROIE, and CORD datasets demonstrate that our model achieves the state-of-the-art performance and almost 1.9X faster inference time than the state-of-the-art methods.

Video captioning has shown impressive progress in recent years. One key reason of the performance improvements made by existing methods lie in massive paired video-sentence data, but collecting such strong annotation, i.e., high-quality sentences, is time-consuming and laborious. It is the fact that there now exist an amazing number of videos with weak annotation that only contains semantic concepts such as actions and objects. In this paper, we investigate using weak annotation instead of strong annotation to train a video captioning model. To this end, we propose a progressive visual reasoning method that progressively generates fine sentences from weak annotations by inferring more semantic concepts and their dependency relationships for video captioning. To model concept relationships, we use dependency trees that are spanned by exploiting external knowledge from large sentence corpora. Through traversing the dependency trees, the sentences are generated to train the captioning model. Accordingly, we develop an iterative refinement algorithm that refines sentences via spanning dependency trees and fine-tunes the captioning model using the refined sentences in an alternative training manner. Experimental results demonstrate that our method using weak annotation is very competitive to the state-of-the-art methods using strong annotation.

Current deep learning methods for action recognition rely heavily on large scale labeled video datasets. Manually annotating video datasets is laborious and may introduce unexpected bias to train complex deep models for learning video representation. In this paper, we propose an unsupervised deep learning method which employs unlabeled local spatial-temporal volumes extracted from action videos to learn midlevel video representation for action recognition. Specifically, our method simultaneously discovers mid-level semantic concepts by discriminative clustering and optimizes local spatial-temporal features by two relatively small and simple deep neural networks. The clustering generates semantic visual concepts that guide the training of the deep networks, and the networks in turn guarantee the robustness of the semantic concepts. Experiments on the HMDB51 and the UCF101 datasets demonstrate the superiority of the proposed method, even over several supervised learning methods.

Deep neural networks have shown excellent performance for stereo matching. Many efforts focus on the feature extraction and similarity measurement of the matching cost computation step while less attention is paid on cost aggregation which is crucial for stereo matching. In this paper, we present a learning-based cost aggregation method for stereo matching by a novel sub-architecture in the end-to-end trainable pipeline. We reformulate the cost aggregation as a learning process of the generation and selection of cost aggregation proposals which indicate the possible cost aggregation results. The cost aggregation sub-architecture is realized by a two-stream network: one for the generation of cost aggregation proposals, the other for the selection of the proposals. The criterion for the selection is determined by the low-level structure information obtained from a light convolutional network. The two-stream network offers a global view guidance for the cost aggregation to rectify the mismatching value stemming from the limited view of the matching cost computation. The comprehensive experiments on challenge datasets such as KITTI and Scene Flow show that our method outperforms the state-of-the-art methods.