ARC (Academic Research Collection) (College Dubin)
Not a member yet
359 research outputs found
Sort by
A Comparative Evaluation of the Effectiveness of Mel Frequency Cepstral Coefficients and Difference Files for Audio Effect Identification Using Convolutional Recurrent Neural Networks.
No full textThe field of Music information retrieval (MIR) is concerned with computational systems which help humans better make sense of the processing, searching, organizing, and accessing of music-related data and takes in disciplines such as music theory, computer science, psychology, neuroscience, library science, electrical engineering and machine learning. MIR processes relating to audio classification have applications in fields such as speech recognition, automatic bandwidth allocation and Audio Database Indexing which is especially of relevance to large audio collections in broadcasting facilities, the movie industry or music content providers.
These indexing aspects of MIR have been applied to classifying audio into musical genres , identifying individual instruments within a piece of music and automatically transcribing music but as of yet none have focused on the identification and cataloguing of the various audio effects that can be applied to music. If it can be established what class of audio effect has been applied to an audio file it has a variety of potential applications such as aiding with the identification of the underlying instruments in the audio or facilitating any library or cataloguing based activities.
The novelty of this paper lies in the following two aspects: 1) no previous studies have been uncovered on the classification of audio effects; 2) The use of a difference file as an input into a neural network is heretofore unexplored.
The dataset used in the study consisted of 7,648 files which were processed through eight different audio effects to give 61,184 unique inputs. A total of 18 different neural network runs were undertaken with three different sets of inputs: MFCCs generated from sub samples of the audio files and images generated from the full length audio files in both MFCC and difference file formats. Variations in the length of the audio file inputs, the number of inputs provided to the model, and the number of epochs run were assessed for their impact on model performance.
The best performing Convolutional Recurrent Neural Network (CRNN) using sub samples of the audio files converted to MFCCs achieved an accuracy of 97.3% via the use of 10 second subsamples generating 100,000 inputs being run for 30 epochs. The overall best performing model used 256*256 px full length difference files as inputs and achieved an accuracy score of 98.4% on just 10 epochs. These results demonstrate the effectiveness of CRNNs at classifying audio effects and the potential of difference files as an image input format for neural networks
Brain Tumor Classification using Deep Learning: Custom CNN vs. ResNet50
No full textThe project presents a deep learning solution to classify brain tumors through MRI images. Following the CRISP-DM framework, two Convolutional Neural Network (CNN) models were developed and evaluated, a custom CNN designed from scratch and a pretrained ResNet50 that was transfer learned and fine-tuned. Both models were assessed using standard performance metrics such as accuracy, precision, recall and F1-score. Despite the higher test accuracy achieved by the custom CNN, further interpretability indicated inconsistent attention to the actual tumor regions also known as shortcut learning. On the other hand, ResNet50 showed more reliable and clinically relevant focus which supported its selection as the final model. The selected model was deployed using Gradio, to demonstrate a real-world application with real-time predictions and visual explanation to reduce the black box nature of AI. The results showcase that despite the good performance of custom and tailored CNN models on specific datasets, generalization and real-world relevance are equally important for reliable deploymen
The use of deep learning solutions to develop a practice tool to support Lámh language for communication partners
No full textThis study has proposed an alternative to promote the learning and enhancement of Lámh language for communication partners that support current users by creating a real time detection tool to recognise 20 chosen Lámh signs based on existing studies in the field. This implementation was carried out by generating primary data composed by MediaPipe landmark numpy arrays of 40 frames and 45 repetitions per sign. The Neural Networks were built using the Python library Keras and the applied SVM models were built with the library sklearn. The real time detection was carried out by integrating the mentioned elements with the library OpenCV. Neural Networks with different architectures with Long Short-Term Memory (LSTM) and 1D Convolutional Neural Network (CNN) were compared with SVM classifications applied with cross-validations to achieve the optimal hyperparameters in order to determine the most appropriate model.
The final chosen model after the assessment of the training and testing accuracy and loss was the two 1-D CNN layers with 32 and 64 nodes respectively, a dropout of 0.2 followed by two LSTM layers with 32 and 64 nodes respectively and a dense layer of 32 nodes. The training accuracy was 99.86%, the testing accuracy was 93.33%, the training loss was 0.0035 and the testing loss was 0.1791. This was the model which performed better in a real-time detection environment, easily detecting 8 different Lámh signs and detecting other 6 with reservations.
For future work, some skeletal motion signs should be captured again and other data augmentation strategies should be adopted, like capturing hips and legs landmarks alongside the signs and explore the augmentation of the data by promoting offset measures of the landmark coordinates of the skeletons captured by MediaPipe. Once the corrections of the methodology achieve better real time results, works toward tool accessibility and user experience should be investigated in order to generate a Lámh language real-time detection tool that could potentially promote Lámh and become a learning alternative for communication partners
Brain Tumor Classification using Deep Learning- Poster
No full textThe project presents deep learning solutions to classify brain tumors through MRI images. Two Convolutional Neural Network (CNN) models were developed, a custom CNN designed from scratch and a pretrained ResNet50 that was transfer learned and fine-tuned.
Both models were implemented following CRISP-DM methodology from data understanding to deployment, and they were evaluated using different metrics such as accuracy, precision, recall and F1-score.
Key Highlights: •The custom CNN model achieved higher accuracy but failed to locate tumors. •ResNet50 provided a good performance while balancing explainability through Grad-CAM. •Model was deployed through Gradio to demonstrate a real-world use of the solution
Customer Service Support. Utilizing machine learning to classify, prioritize and summarize issues.
No full textThis capstone project investigates the application of machine learning and natural language processing (NLP) to enhance customer support operations through automated ticket classification, prioritization, and summarization. Using the multilingual Customer Support Emails dataset from Kaggle, the project follows the CRISP-DM methodology, performing extensive data cleaning, preprocessing, feature engineering, and class balancing. Five machine learning models—Decision Tree, KNN, LinearSVC, Naive Bayes, and Random Forest—were evaluated using hyperparameter tuning, cross-validation, confusion matrix analysis, and learning curves. LinearSVC demonstrated the strongest performance for both queue and priority classification, achieving accuracies of 89.8% and 81.2% respectively, with consistent generalization across folds. For summarization, extractive and abstractive methods were implemented using BERT and BART, with extractive summarization selected as the most reliable for preserving technical accuracy. The results show that machine learning can significantly improve ticket routing efficiency, reduce resolution time, and support customer service agents by providing concise issue overviews. This work demonstrates a practical framework for integrating AI into customer support workflows while addressing ethical considerations such as data privacy, fairness, and robustness
Development of a Deep Learning Model for Synthetic vs. Real Image Classification Synthetic vs. Real image classification
No full textThe advancement of generative AI technologies has made it increasingly difficult to distinguish synthetic images from authentic ones. This capstone project addresses the challenge by developing a binary image classification model using deep learning techniques to differentiate AI-generated images from real photographs. Guided by the CRISP-DM methodology, we employed the DeepGuardDB dataset, consisting of 13,000 balanced image samples, evenly split between real and synthetic sources. We implemented and compared three Convolutional Neural Network (CNN) architectures through transfer learning, standardising input pipelines and integrating custom classification heads. Following a performance evaluation across multiple metrics, the best-performing model was selected for further optimisation using established techniques such as hyperparameter tuning and fine-tuning. This project demonstrates how deep learning can contribute to maintaining information integrity in media, journalism, and digital content platforms.
either AI-generated or real, addressing the growing challenge of synthetic media detection. Using the DeepGuardDB dataset and guided by the CRISP-DM methodology, we implemented and compared three Convolutional Neural Networks (CNNs) architectures via transfer learning. The best-performing model was further optimised using hyperparameter tuning and fine-tuning techniques The resulting model achieved strong accuracy and generalisation, making it a promising candidate for real-time deployment and practical use across diverse industries
Neural Networks Activation Functions and Hybrid Activations Functions accuracy and loss comparison on small dataset against large datasets for classification problems
No full textEven on the era of Big Data, small datasets are the reality of many companies and sectors. Many datasets in rare disease diagnosis, custom manufacturing, military sciences, bioengineering, and disaster events are commonly limited in size, making machine learning predictive modelling difficult. Being the Activation Function choice crucial for Neural Networks learning, it raises the question of their effectiveness in such scenarios. This study compares five standard (single) activation functions (Sigmoid, Tanh, ReLU, Leaky ReLU, ELU) and two hybrid variants (one a mix of ReLU plus Tanh and a Learnable Activation Function with a trainable weight (alpha) that balances ReLU and Tanh in each layer) in feed-forward neural networks on one small dataset (Iris) and two large datasets (Adult, Adult Expanded). The architecture is constant and different activations, optimizers (SGD, Adam, RMSprop), scalers (Standard, Min-Max, Robust), batch sizes (16, 32) and epochs (50, 100) were used as parameters creating 756 models. Performance is evaluated via accuracy and loss convergence. Hybrids (fixed ReLU and Tanh blend, Learnable Hybrid) consistently match or outperform single activations for the small dataset Iris where they achieve 100% test accuracy, and on Adult/Adult Expanded they rank among the top models (~0.853 test accuracy) and the early-loss curves show faster reductions in the first half of training for hybrid activation functions. The findings suggest hybrid activations are a top choice when data are scarce (small datasets) and a strong rule-of-thumb for large datasets as well. The findings are aligned to Kavun (2025) where the hybrid activations outperformed single activation functions. This open discussions for practical guidance and avenues for broader benchmarking when using small datasets. Neural Networks are complex with a multitude of parameters that can be explored like architectures, data types, parameters tunning and various other combinations of activation function ensembles
Big data vs Big law: The impact of big data and machine learning in anonymising or synthesizing data for use across borders.
No full textThis research investigates the viability of anonymization and synthetic data generation in the area of big data so that the data could be shared across borders and exist outside the constraints of privacy laws. These privacy laws are growing around the world to help protect individual identity and prevent open sharing of private data. These privacy laws all provide guidance on how data may be shared and the strict conditions upon how that may occur. Two methods which are growing in popularity are anonymization of data, specifically k-Anonymity, l-Diversity and t-Closeness, and generating synthetic data from a real dataset leveraging machine learning techniques. This paper explores some of these techniques and aims to effectively measure them as a solution to allow organizations to share big data outside of the constraints of privacy laws. The areas of measurement addressed are risk, utility, and usability. A number of measurements are discussed within the paper and implemented within the artifact to allow for comparative testing of different datasets. The focus for this paper is on healthcare and financial data. For anonymization, it was important to understand the quasi-identifiers within the datasets and the sensitive attributes that needed to be considered. These details were used to conduct the measurements around risk and utility. Synthetic data needed to be measured to understand how similar it was to the real data and if any potential leaks of the real data occurred. Both were measured separately, but for usability were tested together across several machine learning models. Across both experiments in healthcare and finance, the results showed that anonymized data contained minimal utility while introducing risk, while real synthetic data performed well, retained utility and demonstrated very low risk. That said, the usability measure showed that synthetic data, while close, doesn’t perform exactly the same as the real data, which could be an issue depending on use case. In conclusion, the synthetic version of the anonymized data appears to be a viable option that could be shared with low risk, good utility and potentially good usability. Keywords