Selected work

Representative Papers

Abstract Compilation as Abstraction of Operator Semantics, applied to Cost Analysis

Aug 10, 2026

Traditional least fixed-point semantics often fails to support precise static cost analysis due to its neglect of recursive structural information. This work proposes operator semantics as an intermediate representation bridging syntax and denotational semantics, treating programs as operators and constructing a higher-order abstract domain grounded in category theory, with composition as the core primitive. This framework enables abstract compilation that simultaneously achieves soundness, precision, and modularity. The approach supports cost analysis for general functional unknowns and generalized fold-based metrics, leveraging a solver-agnostic technique for extracting optimal recurrence relations. Consequently, it facilitates precise static cost analysis of recursive programs over algebraic data types, encompassing generalized size metrics beyond the reach of conventional methods.

0 citationsRead paper

MAUPITI: On-Device Prototype-Based Learning on a Smart Infrared Sensor

Aug 07, 2026

This work addresses privacy-preserving online continual learning of human pose and gestures on ultra-low-power embedded infrared sensors with less than 32 kB of memory. It proposes a replay-free streaming prototype update mechanism that combines an offline-trained, quantized CNN encoder with a nearest-class-mean (NCM) classifier to dynamically maintain class prototypes on-device, thereby eliminating the need for backpropagation and replay buffers. Implemented on a RISC-V microcontroller equipped with low-precision SIMD extensions and a 16×16 TMOS infrared array, the system achieves accuracy comparable to conventional methods on two datasets, with a total latency overhead below 0.29% and power consumption of approximately 1.5 mW—demonstrating, for the first time, efficient, low-latency online continual learning on intelligent infrared sensors.

0 citationsRead paper

Cooking beyond Frames: A Stereo Event Camera Dataset in the Kitchen

Aug 05, 2026

This work addresses the scarcity of event-based datasets capturing unscripted human activities in naturalistic settings by introducing EventKitchen, a novel multimodal dataset recorded in real-world kitchen environments. Using head-mounted sensors, the dataset captures stereoscopic event streams from ten participants engaged in spontaneous cooking tasks across thirteen kitchens, synchronized with RGB, depth, and IMU data. EventKitchen provides the first large-scale, human-centric, and unscripted event-camera recordings of everyday kitchen activities, comprising 5.5 hours of stereo event data, 10,762 action annotations, and 13,482 object bounding boxes. It supports benchmarking for action recognition, object detection, and stereo depth estimation, and includes baseline model evaluations to establish a new standard for neuromorphic vision in daily-life perception.

0 citationsRead paper

QScheduler: Adaptive Gradient Sampling for Zeroth-Order On-Device Training on INT8 NPUs

Jul 21, 2026

This work addresses the challenge of setting the gradient estimation sample size \( q \) in zeroth-order optimization on neural processing unit (NPU)-based microcontrollers, where small \( q \) introduces high noise and premature convergence, while large \( q \) incurs excessive computational overhead—both traditionally requiring costly hyperparameter tuning. To overcome this, the paper proposes QScheduler, an adaptive algorithm that dynamically adjusts \( q \) during training without any pre-specified or tuned values. QScheduler enables, for the first time, INT8-quantized zeroth-order on-device training on the STM32N6 Neural-ART NPU. Experiments on EuroSAT and STL-10 demonstrate that QScheduler achieves accuracy comparable to meticulously tuned fixed-\( q \) baselines while substantially reducing hyperparameter tuning costs, facilitating efficient deployment of models such as ResNet18 and MobileNetV2.

0 citationsRead paper

Towards an Automated Reasoning Tool for Complexity Analysis of Automated Reasoners

Jun 22, 2026

This work addresses the high degree of manual effort and tediousness inherent in existing automated reasoning algorithms—such as those for hyper-exponential quantifier elimination—for complexity analysis. The paper proposes a higher-order abstract interpretation framework grounded in operator semantics, which automatically abstracts symbolic programs into numerical recurrence relations. By integrating termination analysis, fixed-point theory, and SMT solving techniques, the method enables fully automated derivation and verification of asymptotic upper bounds on computational complexity. This approach substantially reduces human intervention while significantly enhancing the automation, efficiency, and scalability of complexity analysis for intricate algorithms.

0 citationsRead paper

Recent publications

Latest Papers

Abstract Compilation as Abstraction of Operator Semantics, applied to Cost Analysis

Aug 10, 2026

Traditional least fixed-point semantics often fails to support precise static cost analysis due to its neglect of recursive structural information. This work proposes operator semantics as an intermediate representation bridging syntax and denotational semantics, treating programs as operators and constructing a higher-order abstract domain grounded in category theory, with composition as the core primitive. This framework enables abstract compilation that simultaneously achieves soundness, precision, and modularity. The approach supports cost analysis for general functional unknowns and generalized fold-based metrics, leveraging a solver-agnostic technique for extracting optimal recurrence relations. Consequently, it facilitates precise static cost analysis of recursive programs over algebraic data types, encompassing generalized size metrics beyond the reach of conventional methods.

0 citationsRead paper

MAUPITI: On-Device Prototype-Based Learning on a Smart Infrared Sensor

Aug 07, 2026

This work addresses privacy-preserving online continual learning of human pose and gestures on ultra-low-power embedded infrared sensors with less than 32 kB of memory. It proposes a replay-free streaming prototype update mechanism that combines an offline-trained, quantized CNN encoder with a nearest-class-mean (NCM) classifier to dynamically maintain class prototypes on-device, thereby eliminating the need for backpropagation and replay buffers. Implemented on a RISC-V microcontroller equipped with low-precision SIMD extensions and a 16×16 TMOS infrared array, the system achieves accuracy comparable to conventional methods on two datasets, with a total latency overhead below 0.29% and power consumption of approximately 1.5 mW—demonstrating, for the first time, efficient, low-latency online continual learning on intelligent infrared sensors.

0 citationsRead paper

Cooking beyond Frames: A Stereo Event Camera Dataset in the Kitchen

Aug 05, 2026

This work addresses the scarcity of event-based datasets capturing unscripted human activities in naturalistic settings by introducing EventKitchen, a novel multimodal dataset recorded in real-world kitchen environments. Using head-mounted sensors, the dataset captures stereoscopic event streams from ten participants engaged in spontaneous cooking tasks across thirteen kitchens, synchronized with RGB, depth, and IMU data. EventKitchen provides the first large-scale, human-centric, and unscripted event-camera recordings of everyday kitchen activities, comprising 5.5 hours of stereo event data, 10,762 action annotations, and 13,482 object bounding boxes. It supports benchmarking for action recognition, object detection, and stereo depth estimation, and includes baseline model evaluations to establish a new standard for neuromorphic vision in daily-life perception.

0 citationsRead paper

QScheduler: Adaptive Gradient Sampling for Zeroth-Order On-Device Training on INT8 NPUs

Jul 21, 2026

This work addresses the challenge of setting the gradient estimation sample size \( q \) in zeroth-order optimization on neural processing unit (NPU)-based microcontrollers, where small \( q \) introduces high noise and premature convergence, while large \( q \) incurs excessive computational overhead—both traditionally requiring costly hyperparameter tuning. To overcome this, the paper proposes QScheduler, an adaptive algorithm that dynamically adjusts \( q \) during training without any pre-specified or tuned values. QScheduler enables, for the first time, INT8-quantized zeroth-order on-device training on the STM32N6 Neural-ART NPU. Experiments on EuroSAT and STL-10 demonstrate that QScheduler achieves accuracy comparable to meticulously tuned fixed-\( q \) baselines while substantially reducing hyperparameter tuning costs, facilitating efficient deployment of models such as ResNet18 and MobileNetV2.

0 citationsRead paper

Towards an Automated Reasoning Tool for Complexity Analysis of Automated Reasoners

Jun 22, 2026

This work addresses the high degree of manual effort and tediousness inherent in existing automated reasoning algorithms—such as those for hyper-exponential quantifier elimination—for complexity analysis. The paper proposes a higher-order abstract interpretation framework grounded in operator semantics, which automatically abstracts symbolic programs into numerical recurrence relations. By integrating termination analysis, fixed-point theory, and SMT solving techniques, the method enables fully automated derivation and verification of asymptotic upper bounds on computational complexity. This approach substantially reduces human intervention while significantly enhancing the automation, efficiency, and scalability of complexity analysis for intricate algorithms.

0 citationsRead paper