From Black-Box AI to Trustworthy AIoT: Explainable Crop Disease Detection with PRISM

Agricultural disease detection is increasingly benefiting from artificial intelligence, yet high predictive accuracy alone does not guarantee that an AI system can be trusted in real-world agricultural environments. This session presents PRISM, an explainable AIoT framework designed to make crop disease detection more transparent, reliable, and practically meaningful.

The talk explores how lightweight deep learning can be combined with explainability and reliability-aware components to support disease classification while providing greater insight into model decisions. The proposed framework is evaluated through a crop disease detection pipeline, with experiments demonstrating strong predictive performance alongside additional reliability-oriented evaluation.

The session will discuss the design of the PRISM framework, its feature-fusion strategy, explainability considerations, evaluation methodology, and the challenges of transferring AI models across different perennial crops. Particular attention will be given to the gap between “accurate AI” and “trustworthy AI” in agricultural applications.

The presentation is intended for researchers, students, AI practitioners, and anyone interested in Explainable AI, AIoT, computer vision, and intelligent agricultural systems. It will conclude with future directions toward robust, cross-crop, and deployment-oriented AI systems for climate-resilient agriculture.

Can AI Learn Across Crops? Transfer Learning for Generalizable Plant Disease Detection

Plant disease detection models can achieve impressive results when trained and tested within the same dataset, but real-world agricultural environments are far more diverse. Different crops, disease patterns, imaging conditions, and data distributions can make it difficult for a model trained on one crop to generalise effectively to another.

This session explores cross-crop transfer learning as a practical approach toward more generalizable agricultural AI. Using deep learning-based crop disease detection as a case study, the presentation examines how knowledge learned from one crop can be transferred to another, the role of pretrained lightweight architectures, and the challenges introduced by differences between agricultural domains.

The talk presents experimental insights from a perennial-crop disease detection pipeline, including transfer learning with MobileNetV2 and the development of a broader Cross-Perennial Domain Adaptation (CPDA) direction. It discusses model initialization, feature reuse, selective fine-tuning, dataset-shift challenges, and evaluation strategies for assessing whether transferred knowledge is genuinely useful.

Rather than treating high accuracy on a single dataset as sufficient evidence of generalization, the session highlights why cross-domain validation is essential for building agricultural AI systems that can move beyond controlled datasets toward practical deployment.

The session is intended for AI/ML students, researchers, computer-vision practitioners, and professionals interested in transfer learning, domain adaptation, and sustainable AI applications. Basic familiarity with machine learning and deep learning is recommended.

Beyond Tokens: A Hybrid Character-Phonetic Transformer for Low-Resource NLP

Traditional subword tokenisation models like BERT often degrade when processing morphologically rich languages, informal text, or noisy user inputs common in chatbot environments. This technical presentation introduces an alternative: a robust token-free, character-level sequence framework engineered explicitly for low-resource NLP pipelines. Drawing directly from successful architectural implementations, we will pull back the curtain on how a CANINE-based encoder handles character variations natively without an explicit vocabulary. We will dive deep into the math and implementation behind integrating custom Indic phoneme mapping layers and building attention-based fusion mechanisms ($\alpha = \text{softmax}(QK^T / \sqrt{d})$) to merge orthographic and pronunciation-aware feature representations. Attendees will walk away with a comprehensive structural blueprint for building error-resilient language models, supported by empirical performance metrics and post-hoc model explainability frameworks using SHAP analysis.

Session Format & Duration: 45-Minute Technical Session / Deep Dive.

Target Audience: Machine Learning Engineers, NLP Researchers, and Data Engineers working on neural network architectures and low-resource language problems.

Technical Requirements: Projector, HDMI connection, and standard audio setup.

Presentation Outline & Timing (Designed to fill a 45-minute slot perfectly):

00:00–00:10 | The Tokenisation Problem: Discussing OOV (Out-of-Vocabulary) limits, subword breakdown in Indic scripts, and why character-level modelling is required.

00:10–00:25 | Core Architecture Deep Dive: Stepping through the CANINE encoder block, phoneme feature extraction matrices, and the mathematical mechanics of the cross-attention fusion layer.

00:25–00:35 | Benchmarks & Metrics: Comparing performance metrics (Character Error Rate vs. Word Error Rate) across CANINE, ByT5, and IndicBERT baselines, highlighting a +6.1% accuracy gain over base models.

00:35–00:45 | Explainable AI & Open Q&A: Utilising SHAP analysis to visually map feature attribution (42% Character, 37% Phonetic, 21% Fusion) to demystify black-box decoder predictions.

Building End-to-End Pipelines for Low-Resource NLP: From Raw Text to Phonetic Mapping

Developing production-ready NLP systems for low-resource languages involves unique hurdles that standard pre-trained architectures cannot solve out of the box. This session provides a practical, engineering-focused look at the data pipelines required to handle noisy, informal, and morphologically complex text corpora. We will break down the structural implementation of a massive 7.19-million-sentence Wikipedia dataset pipeline, detailing the critical preprocessing stages: Unicode normalization, handling complex word formations, and token-free character tokenization. Furthermore, we will explore the architectural logic of mapping text to Indic phoneme representations to capture pronunciation-aware features. This session is a pure engineering blueprint for developers looking to prepare noisy textual data for high-performance sequence models

Session Format & Duration: 30–45 Minute Technical Session. Target Audience: Data Engineers, NLP Developers, and Software Engineers interested in data preprocessing pipelines for non-English scripts. Technical Requirements: Standard projector setup with HDMI. Presentation Outline (35 Minutes):00:00–00:10 | Data Engineering Foundation: Sourcing and structuring lexical resources, cleaning raw text, and implementing strict Unicode normalization rules. 00:10–00:25 | Pipeline Mechanics: Building character-level tokenization filters and implementing the algorithmic mapping for Indic phoneme dictionaries. 00:25–00:35 | Integration & Architecture Output: How this structured data maps into embedding tensors for deep learning models, ending with an open technical Q&A.

Demystifying Deep Learning: Model Evaluation Metrics and Interpretability with SHAP

Building an advanced neural network is only half the battle; evaluating its true performance and making its decisions interpretable is vital for real-world deployment. This session explores how to look inside the "black box" of sequence-to-sequence neural networks using post-hoc explainability tools and specialised evaluation metrics. We will discuss why standard accuracy metrics fall short on character-level and text-generation tasks, introducing Character Error Rate (CER) and Word Error Rate (WER) as industry standards. The core of the session will walk through the implementation of SHAP (Shapley Additive exPlanations) analysis on complex Transformer components. Attendees will learn how to visually quantify the exact feature attribution of different layers—such as character embeddings versus phonetic fusion networks—so they can build trustworthy, explainable AI systems.

Session Format & Duration: 30-Minute Technical Talk. Target Audience: Machine Learning Engineers, Developers, and Product Managers who want to implement Explainable AI (XAI) in production. Technical Requirements: Standard projector setup with HDMI. Presentation Outline (30 Minutes):00:00–00:10 | The Evaluation Gap: Transitioning from basic accuracy to sequence metrics like Character Error Rate (CER) and Word Error Rate (WER). 00:10–00:25 | SHAP Framework Deep Dive: Step-by-step breakdown of feature attribution mechanics, explaining how to calculate layer-by-layer contributions visually. 00:25–00:30 | Explainability for Trust: Best practices for presenting model explanations to non-technical stakeholders, followed by Q&A.