Multi-Resolution Neural Networks – MuReNN

"Less is more", once the foundational motto of minimalist art, is making its way into artificial intelligence. After a maximalist decade of larger computers training larger neural networks on larger datasets (2012-2022), a countertrend arises. What if human-level performance could be achieved with less computing, less memory, and less supervision?

The research prospect of "doing more with less" in deep learning proceeds from a double motivation: to save energy and to save human effort. In this context, the MuReNN project imagines a "less is more" approach to AI systems, specifically to one of its most fundamental constituents: the deep convolutional network, or convnet for short. The overarching goal of MuReNN is to improve both the energy efficiency and annotation efficiency of convnets without compromising their ability for statistical generalization in high dimension.

MuReNN encourages a dialogue between two fields: computational harmonic analysis and end-to-end representation learning. On one hand, it will revisit the theory of multiresolution approximations (MRA) to adopt a critical lens on the current generation of convnets. On the other hand, it will provide an open-source implementation of various kinds of MRA in PyTorch—a deep learning package in Python with automatic differentiation and GPU portability.

MuReNN is the first AI system which proposes to learn an auditory filterbank by combining a dual-tree complex wavelet transform (DTCWT) with a one-dimensional convnet whose learned kernels operate at multiple resolutions. Its main originality consists in the resort to advanced methods in wavelet theory to solve practical problems in deep learning for audio content analysis.

MuReNN contributes to the ongoing diversification of machine listening systems: i.e., not just speech and music, but also sounds from cities, wildlife, factories, and the human body. Hence, while being driven by fundamental research in artificial intelligence (AI), MuReNN has the potential to benefit a broad range of scientific domains: urban planning, animal conservation, Industry 4.0, and healthcare.

The added value of the project is to integrate all stages of frugal deep learning, from data curation to sensor deployment; by way of algorithm design, neural network training, and hardware manufacturing. This will set MuReNN in a unique position to address a long-standing issue: namely, the need for low-cost and low-carbon bioacoustic sensors which are capable to perform audio event detection on device and can be reprogrammed to recognize vocalizations from any species from a limited number of labelled instances. Such sensor have the potential to transform protocols in remote sensing in ecology and conservation while reducing their environmental footprint.

The coordinator is an expert on artificial intelligence for audio content analysis, particularly bioacoustic classification. The consortium spans a wide range of experts: P. Balazs (mathematics of frame theory), M. Lagrange (environmental sound classification), A. Volkova (computer arithmetic), and F. de Dinechin (embedded computing).