06-08, 16:15–17:00 (Europe/London), Hardwick Hub
Transfer learning has revolutionised machine learning by enabling models trained on large datasets to generalise effectively to tasks with limited data. This talk explores strategies for adapting pretrained models to new domains, focusing on audio processing as a case study. Using YAMNet, Whisper, and wav2vec2 for laughter detection, we demonstrate how to extract meaningful representations, fine-tune models efficiently, and handle severe class imbalances. The session covers feature extraction, model fusion techniques, and best practices for optimising performance in data-scarce environments. Attendees will gain practical insights into applying transfer learning across various modalities beyond audio, maximising model effectiveness when labelled data is scarce.
This talk provides a comprehensive exploration of transfer learning, focusing on how pretrained models can be leveraged for tasks with limited labelled data. It begins with an introduction to the core principles of transfer learning, covering different strategies such as feature extraction, fine-tuning, and domain adaptation. The session then delves into the benefits and challenges of using pretrained models, helping attendees understand when and how to apply these techniques effectively.
We will discuss how to choose and adapt pretrained models, with a specific focus on YAMNet, Whisper, and wav2vec2 for audio processing. The talk will cover strategies for handling limited data and severe class imbalance, including data augmentation, synthetic data generation, and advanced loss functions. Attendees will gain insights into fine-tuning techniques, such as layer-wise training and regularisation, to optimise model performance while preventing overfitting. A case study on laughter detection will illustrate these concepts in practice, demonstrating how multiple models can be combined for improved accuracy. Finally, we will explore applications beyond audio, including transfer learning in NLP and computer vision, highlighting cross-domain adaptation techniques and emerging trends in multimodal AI.
No previous knowledge expected
Salman Khan is the Director of Data Science at Afiniti, where he drives innovative solutions to complex business challenges through data science. With a specialization in machine learning, statistical modelling, and a strong focus on generative AI, Salman leads multiple teams of data scientists and engineers in the development and deployment of cutting-edge AI-driven applications. Salman has led AI projects delivering measurable business value, including real-time prediction systems, advanced language models, semantic search platforms, and generative AI applications. Salman’s expertise spans deep learning, probabilistic modelling, and a broad range of data science techniques, with advanced proficiency in Python, R, and SQL.