Agency for Science, Technology and Research

Selected work

Representative Papers

Generative Multiform Bayesian Optimization

May 13, 2022IEEE Transactions on Cybernetics

Bayesian optimization (BO) of expensive black-box functions over complex input spaces—such as discrete or non-Euclidean domains—remains challenging. Existing generative BO (GBO) methods suffer from suboptimal convergence and solution quality due to reliance on a single latent space and inability to handle variable-dimensional inputs robustly. Method: We propose a multimodal generative BO framework featuring: (i) parallel optimization across multiple cooperative latent spaces; (ii) a generative model (VAE/GAN) with positive-correlation constraints to preserve fidelity between latent representations and objective values; and (iii) two cross-space information exchange strategies to reconcile the trade-off between dimension selection and the accuracy–convergence rate balance. Results: Evaluated on airfoil design, cantilever beam optimization, and area maximization tasks, our method achieves significantly faster convergence and higher-quality solutions than both single-latent-space GBO and conventional BO under limited evaluation budgets.

8 citations1 influentialRead paper

ViLa-MIL: Dual-scale Vision-Language Multiple Instance Learning for Whole Slide Image Classification

Jun 16, 2024Computer Vision and Pattern Recognition

In whole-slide image (WSI) classification for digital pathology, existing multiple instance learning (MIL) methods suffer from heavy reliance on abundant bag-level annotations and poor generalizability, while vision-language models (VLMs) are hindered by pathology-agnostic text prompts and prohibitively high pretraining costs, yielding limited performance gains. To address these limitations, we propose ViLa-MIL—a dual-scale vision-language MIL framework. It introduces the first pathology-informed, dual-scale descriptive text prompting mechanism; designs a prototype-guided patch decoder and a context-guided text decoder to enable cross-modal, multi-granularity feature co-modeling; and integrates a frozen large language model, prototype clustering, and vision-language alignment. Evaluated on three multi-cancer, multi-center datasets, ViLa-MIL significantly outperforms state-of-the-art methods, demonstrating low annotation dependency, strong cross-center generalizability, and high robustness.

4 citations1 influentialRead paper

MToP: A MATLAB Optimization Platform for Evolutionary Multitasking

Dec 13, 2023arXiv.org

The multi-task optimization (MTO) community lacks a unified, open-source evaluation platform, hindering reproducibility, fair benchmarking, and practical validation of evolutionary multitasking (EMT) algorithms. Method: This paper introduces EMT-Platform—the first open-source MATLAB platform dedicated to EMT research—featuring a modular architecture, plugin-based algorithm interfaces, a graphical user interface, and built-in knowledge transfer mechanisms. It integrates over 50 multitasking algorithms (including systematically adapted single-task baselines), 200+ standardized benchmark problem instances, and 20+ performance metrics. Contribution/Results: EMT-Platform enables standardized algorithm development, rigorous cross-algorithm evaluation, and intuitive result visualization. It significantly enhances reproducibility, accelerates empirical research, and supports real-world application validation across diverse domains. Widely adopted by the EMT research community, it serves as a foundational infrastructure for advancing both theoretical and applied multitasking optimization.

2 citations1 influentialRead paper

Differentiable Rule Induction from Raw Sequence Inputs

Feb 14, 2026International Conference on Learning Representations

Existing differentiable inductive logic programming (ILP) approaches struggle to learn symbolic rules directly from raw continuous data—such as time-series or images—primarily due to the explicit label leakage problem: without supervision from feature-level labels, they cannot reliably map continuous inputs to symbolic variables. This work proposes an end-to-end neuro-symbolic framework that integrates self-supervised differentiable clustering with a novel differentiable ILP formulation, enabling direct learning of interpretable symbolic rules from raw data without requiring explicit labels. By circumventing the label leakage bottleneck for the first time, the method preserves rule interpretability while substantially improving generalization and applicability. Experiments on both temporal and visual tasks demonstrate its ability to discover accurate and intuitively meaningful symbolic rules.

2 citationsRead paper

Wyckoff Transformer: Generation of Symmetric Crystals

Mar 04, 2025

Existing crystal generation models neglect space-group symmetry constraints, leading to physically implausible and energetically unstable structures. To address this, we propose the first physics-driven generative framework grounded in discrete Wyckoff position representations. Our method introduces a novel position-encoding-free, permutation-invariant autoregressive Transformer that explicitly enforces space-group symmetry as a hard constraint during generation—thereby unifying symmetry fidelity, structural stability, and property predictability. Evaluated across multiple benchmarks, our approach significantly outperforms state-of-the-art models in symmetry fidelity (measured via space-group consistency), energy stability (via DFT-calculated formation energies), property prediction accuracy (e.g., bandgap, bulk modulus), and inference speed. Notably, it achieves, for the first time, controllable generation of high-quality, symmetry-compliant crystals under strict space-group constraints.

2 citationsRead paper

Recent publications