Abstract: In this study, we investigate the expressive power of deep neural networks with ReLU activations in binary classification tasks, where the final readout is given by a linear hyperplane. Our focus is on the geometric and topological properties of the negative region induced in the input space. Using structural properties of the tope graph associated with hyperplane arrangements, we analyze the topology of the decision region at the output layer and reveal certain constraints on its structure. We then examine how this region is propagated backward through the layers of the network via the ReLU transformations. This viewpoint allows us to study how topological features evolve under layer-wise pullbacks. Based on this framework, we propose a conjectural principle describing the stability of topological properties of decision regions under such maps. This work establishes a connection between neural network expressivity and combinatorial topology.

Abstract: The growing use of machine learning in data-driven organizations has increased the need for models that are not only accurate, but also interpretable, reliable, and operationally meaningful. Particular attention is given to decision environments shaped by complex organizational data, where transparency and robustness play a critical role in the quality of managerial actions. Within this context, machine learning is addressed as a core component of decision intelligence supported by analytical reasoning, domain knowledge, and implementation constraints. This perspective establishes a connection between interpretable machine learning and trustworthy decision intelligence in management information systems.

Abstract: Rotating machinery serves as the core component of China's "national heavy equipment," underpinning energy, transportation, manufacturing, and defense. Sudden failures can cause catastrophic economic losses and safety accidents, making early fault diagnosis of urgent theoretical and practical value. With the rapid advancement of artificial intelligence, deep learning-based fault diagnosis has attracted increasing attention. This report focuses on fault diagnosis under domain adaptation, partial domain adaptation, and domain generalization scenarios. It reviews the challenges posed by domain shift between training and testing data, including distribution discrepancy, label space mismatch, and unseen target domains. Finally, future research directions toward robust and generalizable intelligent diagnosis for rotating machinery are discussed.

Abstract: As IoT devices become increasingly integrated into critical infrastructure, scalable and effective firmware security testing has never been more important. Yet, despite a decade of research and rapid advances in large language models (LLMs), how far are we from reliably emulating and testing real-world IoT firmware at scale? In this talk, we systematically revisit over 20 representative firmware emulation and fuzzing techniques developed over the past decade, examining their fundamental challenges, technical advances, and limitations. Moving beyond small, curated benchmarks, we further conduct a large-scale evaluation using over 4,000 real-world black-box firmware samples across an end-to-end security analysis pipeline.Our findings reveal a striking reality: only 1,580 firmware samples (34%) can be successfully fuzzed by at least one existing tool, with an average code coverage of approximately 10% and frequent false crashes and hangs. By uncovering the root causes of these limitations, this talk highlights the critical gaps in today's techniques and explores future opportunities for scalable, intelligent, and LLM-assisted firmware security analysis.