Goal:  Utilizing computer vision and object detection algorithms, we are developing AI solutions to assess the quality of agricultural products at the end of the production chain.

Case Study 1: Fruit Inspection System: YOLO-based Model for Identifying Defects on Plums Surface 

    Thesis
         -  Intelligent Fruit Inspection System: Developing a YOLO-based Model for Identifying Defects on Plums Surface
         -  Computer Vision for Segmentation and Quantification of Damage Surfaces on African Plum Fruits
   Publications
     Indaba 2023: https://openreview.net/group?id=DeepLearningIndaba.com/2023/Conference#tab-active-submissions
        - https://openreview.net/forum?id=uBPApDJPfn
       - https://drive.google.com/file/d/19rQQzdQmNKfFjxGm1IxUbEy13oA0sqq5/view

Case Study 2: Tomato Quality Assessment Using Object detection-Based Models
     Task: An AI-driven vision system, specifically utilizing Deep Learning to efficiently sort tomatoes.
      Data collection: IN progress.

The research explores the application of Artificial Intelligence (AI) and Machine Learning (ML) in medical diagnosis, particularly in the context of the Covid-19 pandemic. The authors propose a neural-symbolic system called Neural HPLP, which combines Deep 3D Convolutional Neural Networks (3D-CNNs) and Decision Trees (DTs) to predict whether a Covid-19 patient will end up in critical condition. The system analyzes lung CT scans using 3D-CNNs and clinical data using DTs, and integrates them using Hierarchical Probabilistic Logic Programs (HPLPs).
The ability to predict if a Covid-19 patient will deteriorate into a critical condition is valuable for managing limited resources, such as intensive care units in hospitals. It also allows doctors to gain early knowledge about patients and provide specialized treatment to those predicted to be at risk. The Neural HPLP system achieves good performance in terms of accuracy and precision, with values of about 0.96 for both metrics. Additionally, the system is designed to provide explanations for its predictions, making them explainable, interpretable, and reliable.
Publications
- Neural-Symbolic System for Predicting COVID-19 Positivity
- Neural-Symbolic Ensemble Learning for early-stage prediction of Critical State of Covid-19 patients

This research improves the previously defined Lifted Probabilistic Logic Programming (LPLP) language to favor the integration of the logical approach of AI and the connectionist approach based on artificial neural networks. An extension of the LPLP language has been proposed, called Hierarchical Probabilistic Logic Programs (HPLP), in which clauses and predicates are hierarchically organized. A program in this language can be converted into a series of deep neural networks and the reasoning is done by evaluating the networks. Several systems have been proposed to perform reasoning and learning in HPLPs. To learn HPLP parameters from data, Arnaud Nguembang Fadja proposed the algorithm, “Parameter learning for HIerarchical probabilistic Logic programs (PHIL)”. Two variants of PHIL, Deep PHIL (DPHIL) and EMPHIL (EMPHIL), and their regularized versions were presented. Furthermore, he proposed the SLEAHP algorithm, for Structure LEArning of Hierarchical Probabilistic logic programming, to learn both the structure and the parameters of HPLP from the data.
Publications
- Learning hierarchical probabilistic logic programs
- Learning the parameters of deep probabilistic logic programs
- Deep probabilistic logic programming
-  Deep learning for probabilistic logic programming

In the real world, information could have levels of uncertainty, so it becomes essential to provide methods for representing such type of information. To cover this need, the research group on Artificial Intelligence of the University of Ferrara, of which Arnaud Nguembang Fadja is a member, proposed two reasoning systems based on the language of distributional semantics a few years ago. The two systems, called EMBLEM and SLIPCOVER, are very effective in terms of the result provided but are very expensive in terms of learning and reasoning time. Arnaud Nguembang Fadja then proposed a new knowledge representation language called "Liftable Probabilistic Logic Programs (LPLP)" in which reasoning and learning are much faster. He has implemented several systems of reasoning and learning both parameters and the structure of programs based on the LPLP language. He developed the algorithm, LIFTCOVER for LIFTed slipCOVER, which learns the structure of LPLP from the data. Two versions of LIFTCOVER have been proposed: the first, LIFTCOVER-EM, uses the Expectation Maximization (EM) algorithm as a subroutine for learning the parameters, and the second, LIFTCOVER-LBFGS, uses an optimization method called BFGS with limited memory.
Publications
- Lifted discriminative learning of probabilistic logic programs, published in a Machine learning journal.
- Lifted discriminative learning of probabilistic logic programs, published at the 27th International Conference on Inductive Logic Programming, ILP 2017, 2017