研究对比了四种多模态大型语言模型与放射科医生在乳腺密度评估、BI-RADS评估、活检候选确定及恶性概率估计上的表现，发现放射科医生在分类任务上更优，但某些模型在连续恶性概率估计中接近人类水平。
This study addresses the challenge of intuitively and unobtrusively visualizing continuously sensed emotional data within the constrained screen space of wearable devices. Through a cross-cultural user study involving 105 participants from Poland and Turkey, it systematically investigates the universality and cultural variability of visual parameters—including color, shape, size, and animation speed—in conveying emotional states. Findings reveal cross-cultural consistency in the interpretation of color and object size, whereas animation speed exhibits significant cultural differences. Building on these insights, the work proposes an abstract geometric animation model tailored for global audiences, offering a theoretically grounded framework for emotion feedback design in wearables. This model further enables generative algorithms to transform physiological sensor data into intuitive, culturally adaptive dynamic visualizations.
This work addresses the high computational cost and energy consumption of deep learning—stemming from its reliance on massive floating-point multiplications—which hinders deployment on edge devices. Inspired by biological neural systems, the authors propose a novel multiplication-free neural network architecture that restricts signal amplitude through channel width and introduces trainable “neurotransmitter” parameters to modulate signal transmission under symbolic logic. Forward propagation is achieved using only addition, subtraction, and bitwise operations. By translating the physical saturation mechanism of biological synapses into a trainable, multiplication-free structure, the method achieves 100% accuracy on nonlinear tasks such as XOR and majority functions, demonstrating its capacity to model complex decision boundaries. This approach establishes a new paradigm for low-power, GPU-independent AI deployment.
Handwritten digit recognition on resource-constrained, low-power embedded FPGA platforms demands high real-time performance and predictable timing. Method: This work proposes a fully hand-coded Verilog binary neural network (BNN) inference accelerator, eschewing high-level synthesis tools. A custom hardware architecture is designed with bit-level optimizations and fine-grained timing control to achieve stable operation at 80 MHz. The end-to-end flow—including MNIST training, binarization, and hardware mapping—is deployed on a Xilinx Artix-7 FPGA. Contribution/Results: We present the first fully RTL-level, manually designed BNN accelerator that jointly optimizes reconfigurability, resource efficiency, and deployment transparency. Experimental results demonstrate 84% classification accuracy on MNIST, significantly reduced power consumption, and highly deterministic timing behavior. To foster reproducibility and further research, we publicly release both the complete training scripts and the synthesizable RTL code.
To address the lack of systematic, large-scale evaluation benchmarks for low-resource languages like Turkish, this paper introduces TR-MMLU—the first comprehensive multitask language understanding benchmark for Turkish. It comprises 6,200 expert-validated multiple-choice questions spanning 62 subjects aligned with the Turkish national education curriculum. TR-MMLU enables fine-grained assessment of language comprehension, logical reasoning, and cross-domain knowledge acquisition. Systematic evaluation across leading open- and closed-weight large language models reveals substantial performance gaps in Turkish, particularly in domain-specific reasoning and long-range dependency modeling. TR-MMLU fills a critical gap in low-resource language model evaluation, providing a standardized, reproducible benchmark to guide future model development, data curation, and capability diagnostics. It establishes a new de facto standard for Turkish NLP evaluation.
研究对比了四种多模态大型语言模型与放射科医生在乳腺密度评估、BI-RADS评估、活检候选确定及恶性概率估计上的表现，发现放射科医生在分类任务上更优，但某些模型在连续恶性概率估计中接近人类水平。
This study addresses the challenge of intuitively and unobtrusively visualizing continuously sensed emotional data within the constrained screen space of wearable devices. Through a cross-cultural user study involving 105 participants from Poland and Turkey, it systematically investigates the universality and cultural variability of visual parameters—including color, shape, size, and animation speed—in conveying emotional states. Findings reveal cross-cultural consistency in the interpretation of color and object size, whereas animation speed exhibits significant cultural differences. Building on these insights, the work proposes an abstract geometric animation model tailored for global audiences, offering a theoretically grounded framework for emotion feedback design in wearables. This model further enables generative algorithms to transform physiological sensor data into intuitive, culturally adaptive dynamic visualizations.
This work addresses the high computational cost and energy consumption of deep learning—stemming from its reliance on massive floating-point multiplications—which hinders deployment on edge devices. Inspired by biological neural systems, the authors propose a novel multiplication-free neural network architecture that restricts signal amplitude through channel width and introduces trainable “neurotransmitter” parameters to modulate signal transmission under symbolic logic. Forward propagation is achieved using only addition, subtraction, and bitwise operations. By translating the physical saturation mechanism of biological synapses into a trainable, multiplication-free structure, the method achieves 100% accuracy on nonlinear tasks such as XOR and majority functions, demonstrating its capacity to model complex decision boundaries. This approach establishes a new paradigm for low-power, GPU-independent AI deployment.
Handwritten digit recognition on resource-constrained, low-power embedded FPGA platforms demands high real-time performance and predictable timing. Method: This work proposes a fully hand-coded Verilog binary neural network (BNN) inference accelerator, eschewing high-level synthesis tools. A custom hardware architecture is designed with bit-level optimizations and fine-grained timing control to achieve stable operation at 80 MHz. The end-to-end flow—including MNIST training, binarization, and hardware mapping—is deployed on a Xilinx Artix-7 FPGA. Contribution/Results: We present the first fully RTL-level, manually designed BNN accelerator that jointly optimizes reconfigurability, resource efficiency, and deployment transparency. Experimental results demonstrate 84% classification accuracy on MNIST, significantly reduced power consumption, and highly deterministic timing behavior. To foster reproducibility and further research, we publicly release both the complete training scripts and the synthesizable RTL code.
To address the lack of systematic, large-scale evaluation benchmarks for low-resource languages like Turkish, this paper introduces TR-MMLU—the first comprehensive multitask language understanding benchmark for Turkish. It comprises 6,200 expert-validated multiple-choice questions spanning 62 subjects aligned with the Turkish national education curriculum. TR-MMLU enables fine-grained assessment of language comprehension, logical reasoning, and cross-domain knowledge acquisition. Systematic evaluation across leading open- and closed-weight large language models reveals substantial performance gaps in Turkish, particularly in domain-specific reasoning and long-range dependency modeling. TR-MMLU fills a critical gap in low-resource language model evaluation, providing a standardized, reproducible benchmark to guide future model development, data curation, and capability diagnostics. It establishes a new de facto standard for Turkish NLP evaluation.