Selected work

Representative Papers

High-Level Synthesis of Digital Circuits from Template Haskell and SDF-AP

Apr 10, 2025International Conference / Workshop on Embedded Computer Systems: Architectures, Modeling and Simulation

To address the lack of explicit temporal semantics and execution-order modeling in functional languages for high-level synthesis (HLS), this paper proposes a novel hardware description methodology integrating the Synchronous Dataflow with Actor Parameters (SDF-AP) model and Template Haskell. It is the first to embed SDF-AP’s production/consumption timing constraints directly into functional specifications, leveraging higher-order function reuse and dataflow patterns to jointly characterize resource allocation and critical-path latency. Built upon the Clash compiler framework, the approach automatically generates VHDL/Verilog code featuring deterministic timing behavior and complete control- and data-path implementations. Experimental evaluation across multiple benchmarks demonstrates stable resource utilization and strict cycle-accurate timing predictability. Compared to Vitis HLS, the method achieves 23–41% average latency reduction and up to 18% lower resource consumption in selected designs.

3 citationsRead paper

Carefree multiple testing with e-processes

Jan 31, 2025

Existing e-BH procedures lack order-invariance over e-processes, causing test conclusions to reverse spuriously upon addition of irrelevant data and failing to control the false discovery rate (FDR) — or even the family-wise error rate (FWER) — under arbitrary dependence. This paper provides the first rigorous proof that e-BH violates FDR control in this setting. Method: We propose a novel, order-invariant multiple testing framework built on e-process upper bounds, featuring a dependence-structure-adaptive calibrator. Contribution/Results: Our method guarantees strict FDR control at level α (i.e., FDR-sup ≤ α) for arbitrary dependence structures among hypotheses. It eliminates temporal instability in rejection sets induced by sequential data arrival, ensuring robustness and reproducibility in dynamic data environments. Theoretical guarantees are established without restrictive assumptions on dependence, and the procedure is computationally tractable.

2 citations1 influentialRead paper

Physical Analog Kolmogorov-Arnold Networks based on Reconfigurable Nonlinear-Processing Units

Feb 07, 2026

This work addresses the challenge of efficiently implementing the learnable nonlinear edge functions in Kolmogorov–Arnold Networks (KANs) in hardware by proposing an analog KAN architecture based on a Reconfigurable Nonlinear Processing Unit (RNPU). Leveraging multi-terminal nanosilicon devices that natively support programmable nonlinear transformations, the design employs the RNPU as its fundamental computational element, integrated with analog mixed-signal interfaces to achieve high parameter efficiency, low power consumption, and minimal silicon area for edge neural network deployment. Experimental results demonstrate that, at comparable approximation error levels, the proposed architecture reduces energy consumption by two to three orders of magnitude and chip area by approximately one order of magnitude relative to digital fixed-point MLP implementations, achieving a single-inference energy cost of merely 250 pJ with a latency of about 600 ns.

1 citationsRead paper

Memory Wall is not gone: A Critical Outlook on Memory Architecture in Digital Neuromorphic Computing

Jul 06, 2025IEEE Computer Society Annual Symposium on VLSI

Although digital neuromorphic computing adopts near-memory architectures, on-chip memory systems still impose a significant memory wall that limits energy and area efficiency, hindering deployment in edge scenarios. This work challenges the prevailing notion that the memory wall has been resolved by demonstrating, for the first time, that on-chip memories such as SRAM and STT-MRAM themselves constitute primary bottlenecks for performance and energy efficiency. By developing energy and area models for various memory technologies and integrating them into representative digital neuromorphic architectures, the study quantifies the overheads of different memory solutions, revealing that current designs fall short of edge application requirements. The analysis further identifies critical directions for future memory architecture re-design to overcome these limitations.

1 citationsRead paper

Revisiting the attacker's knowledge in inference attacks against Searchable Symmetric Encryption

Apr 14, 2025

This work investigates the dependence of inference attacks in Searchable Symmetric Encryption (SSE) on the quality of “similar data” available to the adversary. We propose the first general statistical analysis framework that formally defines “similar data” and reveals how its non-uniqueness critically impacts attack robustness. We prove that index size constraints significantly degrade inference attack efficacy and derive a provably secure lower bound on the required index size. Within the leakage-abuse model, we integrate probabilistic modeling with statistical estimation theory and empirically validate our findings on the Enron dataset: imposing an index size cap of 200 reduces the optimal inference attack’s accuracy to below 5% with high probability. Our results yield the first quantifiable, data-similarity-aware defense configuration guideline for SSE systems—bridging theoretical security guarantees with practical deployment constraints.

1 citationsRead paper

Recent publications

Latest Papers

MergeOver: Post-Training Token Merging for Recursive Vision Transformers

Aug 13, 2026

This work addresses the persistent computational and memory overheads of recursive Vision Transformers, which remain challenging despite parameter reduction through weight sharing, and their incompatibility with token merging techniques without retraining. To overcome these limitations, we propose MergeOver—a post-training, plug-and-play method that, for the first time, effectively integrates token merging into the recursively weight-shared SReT architecture. Our approach resolves spatial misalignment and merging constraints via an unmerge tracking stack, safety-constrained merge ratio control, a cross-space permutation-aware token quality synchronization mechanism, and a staged single-pass scheduling strategy. Evaluated on ImageNet-1K, MergeOver incurs only a 1.47% accuracy drop while reducing GPU peak activation memory by 37.3% (batch=1) and 38.4% (batch=16), and decreases inference latency on Raspberry Pi 5 by 2.4% (batch=1) and 17.6% (batch=16).

0 citationsRead paper

Reasoning for Social Audio-Visual Question Answering: Where Do We Stand?

Aug 13, 2026

This work addresses critical limitations in existing audio-visual social understanding benchmarks, which suffer from high noise levels and poorly designed questions, while complex reasoning approaches often incur substantial costs with marginal gains. The authors systematically evaluate the reasoning capabilities of multimodal large language models on social audio-visual question answering, introducing IntentBench-Prime—a high-quality benchmark constructed through rigorous data cleaning—and comparing diverse training strategies. Their findings reveal that a simple vanilla supervised fine-tuning (SFT) baseline matches or surpasses state-of-the-art complex methods across three benchmarks. Notably, using only textual captions achieves performance comparable to full video inputs, suggesting that linguistic modalities encode strong social priors. The study further proposes a cost-effective evaluation paradigm and publicly releases the denoised IntentBench-Prime benchmark to support future research.

0 citationsRead paper

Structural Morphisms for Nested Conditions - Full Version

Aug 12, 2026

This study addresses the formal representation and preservation of logical relationships among nested conditions in graph transformation rules. To overcome the limitations of existing approaches, which lack precise characterization of structural dependencies between nested conditions, the work introduces condition operators that emulate logical connectives and, for the first time, defines structural morphisms between nested conditions. Within a categorical framework, it establishes criteria under which these morphisms align with logical implication and proves that they preserve implication under a specific semantic interpretation. Furthermore, the paper uncovers the functorial nature and universal properties of the proposed constructions, thereby providing novel formal tools to strengthen the logical foundations of graph transformation systems.

0 citationsRead paper

Deep Learning based Detection of Fishing Vessels and Fishing Monitoring using Nightlight Images

Aug 10, 2026

This study addresses the challenge of monitoring illegal fishing by “dark vessels”—those without Automatic Identification System (AIS) signals—along India’s west coast by proposing a dual-branch YOLOv11 model specifically optimized for nighttime light imagery. For the first time, it integrates multi-resolution night-time data from the SDGSAT-1 satellite, combining 10-meter panchromatic and 40-meter RGB bands to enable high-precision detection of small fishing boats. Applied to 2022–2023 data, the method identified 31,525 vessels, with 77.3% classified as dark vessels. It achieved a precision of 0.99, recall of 0.93, and mAP@50 of 0.96. The analysis further reveals January–April as the primary fishing season and the 50–100 km offshore zone as the core operational area, significantly enhancing remote sensing capabilities for nocturnal small-target vessel monitoring.

0 citationsRead paper

Dynamic $e$-closure for online hypotheses with any-time-valid evidence: closure principles and projective mergers

Aug 10, 2026

This work addresses the challenge of controlling multiple testing errors in online hypothesis testing, where hypotheses arrive dynamically and evidence becomes observable at arbitrary times. The authors propose a dynamic e-closure method that integrates e-values, dynamic closure principles, and cross-time consistency constraints to establish the first theoretical framework capable of effectively controlling the supremum false discovery rate (SupFDR) under simultaneous stopping and ensuring the persistence of rejection sets. Key contributions include establishing universal guarantees under a canonical normalized loss process, revealing structural limitations inherent to pointwise merging procedures, and constructing a globally shared control procedure via projection merging, alongside a counterexample demonstrating the failure of consistency under certain conditions.

0 citationsRead paper