Large Language Models (LLMs) often encode latent biases that are difficult to eliminate through training or design alone. Like humans requiring education to regulate instinctive behaviour, LLMs need external safeguards to guide their outputs responsibly. Guardrail techniques have thus become essential for promoting fairness, safety and cultural sensitivity while preserving the model’s original capabilities. However, many existing guardrails reflect Western-centric norms, which can overlook linguistic diversity and national contexts, especially in smaller or multicultural societies. A case study on Māori language shows how rigid safeguards can distort meaning and suppress Indigenous perspectives. Responsible AI requires not only technical control but also culturally grounded approaches that are informed by local languages, values and policy environments.
This paper presents an Explainable AI (XAI) method for explaining Spiking Neural Networks (SNNs) through surrogate modeling. The proposed method involves translating trained SNNs into an equivalent Multi-Layer Perceptron (MLP) model, and enabling the use of standard post-hoc explanation techniques such as Shapley Additive Explanation (SHAP). The translation framework, implemented in Nengo, includes a custom weight-mapping and Leaky Integrate-and-Fire Rate (LIFRate)-based activation to approximate spiking behaviour. On a binary Distributed Denial-of-Service (DDoS) detection task, the translated model achieved lower accuracy (0.87) than benchmark MLPs (0.99–1.00), but identified the same key features. Further experiments showed that Recursive Least Squares (RLS)-trained SNNs consistently outperformed Prescribed Error-Sensitivity (PES)-trained variants.
Grad-CAM is a widely used feature attribution method for visualizing class-relevant regions in image classification models, but its attribution maps often suffer from low spatial resolution. To address this problem, we propose an enhanced version of Grad-CAM that is integrated with FeatUp, a model-agnostic upsampling technique for constructing high-resolution feature maps. In FeatUp, we adopt the Implicit variant, which uses multi-layer perceptrons (MLPs) to generate continuous representations over arbitrary spatial coordinates. Whereas the original FeatUp paper briefly mentioned its potential for explainability in deep learning, our study conducted an in-depth investigation of its effectiveness. As a result, thanks to FeatUp, the proposed method successfully enhances the resolution of Grad-CAM without modifying the model architecture under study. We evaluated the proposed method with ViT, ResNet50, and VGG19 both qualitatively and quantitatively. Experimental results indicated a certain improvement in quantitative evaluation, and a notable improvement in qualitative evaluation, especially in terms of localization and visual clarity. These findings suggest that high-resolution feature representations can benefit feature attribution in image classification.
Traditional office resource management systems often suffer from inefficiencies, including 35–40
The rapid growth of data generated by various domains presents new opportunities for research and innovation. However, this data is often underutilized due to the technical expertise required to interpret and apply it effectively. To address this challenge, we propose an automated framework that leverages large language models (LLMs) to generate natural language dataset profiles. These profiles aim to help non-expert users understand what a dataset contains, what it represents, and how it can be used. The framework operates in four stages: data input, metadata extraction from raw datasets, profile generation using few-shot prompting with LLMs, and evaluation of the generated profiles. Each profile summarizes key aspects such as data types, subject domain, semantic content, and potential use cases. We evaluate our method using datasets from Kaggle and assess the output through both human annotation and automated metrics. Our results show that the generated profiles are coherent, relevant, and align well with human judgments, suggesting the potential of LLMs to support accessible and scalable dataset understanding.
