Yang Xiang

Yexing Du, Youcheng Pan, Ziyang Ma et al.

Multimodal Large Language Models (MLLMs) have achieved significant success in Speech-to-Text Translation (S2TT) tasks. While most existing research has focused on English-centric translation directions, the exploration of many-to-many translation is still limited by the scarcity of parallel data. To address this, we propose a three-stage curriculum learning strategy that leverages the machine translation capabilities of large language models and adapts them to S2TT tasks, enabling effective learning in low-resource settings. We trained MLLMs with varying parameter sizes (3B, 7B, and 32B) and evaluated the proposed strategy using the FLEURS and CoVoST-2 datasets. Experimental results show that the proposed strategy achieves state-of-the-art average performance in $15\times14$ language pairs, requiring fewer than 10 hours of speech data per language to achieve competitive results. The source code and models are released at https://github.com/yxduir/LLM-SRT.

Junyi Fan, Yuxuan Han, Jialin Zeng et al.

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 $ε$-CCE with a provable optimal accuracy bound $O(ε^{-2})$ 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.

Yang Xiang, Michael P. Beddoes, David L Poole

In this paper, the feasibility of using finite totally ordered probability models under Alelinnas's Theory of Probabilistic Logic [Aleliunas, 1988] is investigated. The general form of the probability algebra of these models is derived and the number of possible algebras with given size is deduced. Based on this analysis, we discuss problems of denominator-indifference and ambiguity-generation that arise in reasoning by cases and abductive reasoning. An example is given that illustrates how these problems arise. The investigation shows that a finite probability model may be of very limited usage.

Yang Xiang, David L. Poole, Michael P. Beddoes

Current Bayesian net representations do not consider structure in the domain and include all variables in a homogeneous network. At any time, a human reasoner in a large domain may direct his attention to only one of a number of natural subdomains, i.e., there is ?localization' of queries and evidence. In such a case, propagating evidence through a homogeneous network is inefficient since the entire network has to be updated each time. This paper presents multiply sectioned Bayesian networks that enable a (localization preserving) representation of natural subdomains by separate Bayesian subnets. The subnets are transformed into a set of permanent junction trees such that evidential reasoning takes place at only one of them at a time. Probabilities obtained are identical to those that would be obtained from the homogeneous network. We discuss attention shift to a different junction tree and propagation of previously acquired evidence. Although the overall system can be large, computational requirements are governed by the size of only one junction tree.

Yang Xiang

Recent developments show that Multiply Sectioned Bayesian Networks (MSBNs) can be used for diagnosis of natural systems as well as for model-based diagnosis of artificial systems. They can be applied to single-agent oriented reasoning systems as well as multi-agent distributed probabilistic reasoning systems. Belief propagation between a pair of subnets plays a central role in maintenance of global consistency in a MSBN. This paper studies the operation UpdateBelief, presented originally with MSBNs, for inter-subnet propagation. We analyze how the operation achieves its intended functionality, which provides hints as for how its efficiency can be improved. We then define two new versions of UpdateBelief that reduce the computation time for inter-subnet propagation. One of them is optimal in the sense that the minimum amount of computation for coordinating multi-linkage belief propagation is required. The optimization problem is solved through the solution of a graph-theoretic problem: the minimum weight open tour in a tree.

Michael S. K. M. Wong, C. J. Butz, Yang Xiang

This paper discusses a method for implementing a probabilistic inference system based on an extended relational data model. This model provides a unified approach for a variety of applications such as dynamic programming, solving sparse linear equations, and constraint propagation. In this framework, the probability model is represented as a generalized relational database. Subsequent probabilistic requests can be processed as standard relational queries. Conventional database management systems can be easily adopted for implementing such an approximate reasoning system.

Yang Xiang, Michael S. K. M. Wong, N. Cercone

In learning belief networks, the single link lookahead search is widely adopted to reduce the search space. We show that there exists a class of probabilistic domain models which displays a special pattern of dependency. We analyze the behavior of several learning algorithms using different scoring metrics such as the entropy, conditional independence, minimal description length and Bayesian metrics. We demonstrate that single link lookahead search procedures (employed in these algorithms) cannot learn these models correctly. Thus, when the underlying domain model actually belongs to this class, the use of a single link search procedure will result in learning of an incorrect model. This may lead to inference errors when the model is used. Our analysis suggests that if the prior knowledge about a domain does not rule out the possible existence of these models, a multi-link lookahead search or other heuristics should be used for the learning process.

Jun Hu, Yang Xiang

A pseudo independent (PI) model is a probabilistic domain model (PDM) where proper subsets of a set of collectively dependent variables display marginal independence. PI models cannot be learned correctly by many algorithms that rely on a single link search. Earlier work on learning PI models has suggested a straightforward multi-link search algorithm. However, when a domain contains recursively embedded PI submodels, it may escape the detection of such an algorithm. In this paper, we propose an improved algorithm that ensures the learning of all embedded PI submodels whose sizes are upper bounded by a predetermined parameter. We show that this improved learning capability only increases the complexity slightly beyond that of the previous algorithm. The performance of the new algorithm is demonstrated through experiment.

TongSheng Chu, Yang Xiang

It has been shown that a class of probabilistic domain models cannot be learned correctly by several existing algorithms which employ a single-link look ahead search. When a multi-link look ahead search is used, the computational complexity of the learning algorithm increases. We study how to use parallelism to tackle the increased complexity in learning such models and to speed up learning in large domains. An algorithm is proposed to decompose the learning task for parallel processing. A further task decomposition is used to balance load among processors and to increase the speed-up and efficiency. For learning from very large datasets, we present a regrouping of the available processors such that slow data access through file can be replaced by fast memory access. Our implementation in a parallel computer demonstrates the effectiveness of the algorithm.