AI is getting bigger, but does bigger always mean better? As Large Language Models (LLMs) dominate the scene, their staggering resource consumption raises urgent questions about sustainability and efficiency. In this webinar, we challenge the notion that AI must be massive to be powerful. We introduce the Expressive Neural Network (ENN), a novel architecture that rethinks activation functions through the lens of classical signal processing – specifically, the Discrete Cosine Transform (DCT). This innovative approach not only enhances a network’s flexibility and expressiveness but also leads to faster convergence and significantly smaller models, reducing both energy consumption and computational costs.
Our discussion bridges the gap between traditional signal processing techniques and modern AI, demonstrating how established mathematical tools can inspire next-generation machine learning. We explore how ENNs can revolutionize edge computing, enabling efficient AI in resource-constrained environments, and why expressiveness – not just size – is the key to the future of neural networks. If LLMs are the brute force of AI, could ENNs be its precision tool?
Speaker(s):
Ana Isabel Pérez-Neira
Director, Centre Tecnològic de Telecomunicacions de Catalunya, Spain
Moderator(s):
Ian F. Akyildiz
ITU J-FET Editor-in-Chief and Truva Inc., USA
