Centrum Wiskunde & Informatica
EDL P16-25 P7: Low Power Platforms for Medical Applications
Efficient Deep Learning for System Health Management in Radio Astronomy
With the development of Deep learning, the applications of AI are increasingly seeping into our daily lives. But, because of the energy cost of operation, most applications are hardly ever done locally. During my PhD research, I’m looking for solutions to achieve > 100 times better energy efficiency based on brain-inspired neural networks – in particular spiking neural networks (SNN).
- Designing efficient learning algorithms for Spiking recurrent Neural Networks. This is challenging, as the nonlinearity of the spiking neuron makes it non-differentiable, and the continuous-time dynamics of the neuron raises a temporal credit assignment problem. We aim to design a learning framework for SNNs compatible with auto-differentiation platforms.
- Innovating online learning algorithms for network personalization. Networks should be easily personalized and locally updated based on user’s data. In this way, the network can be more user-friendly and more private. We are looking for efficient online learning algorithms fitting SNNs.
- Exploring the possibility of applying SNNs in hardware. We are working with IMEC to discover how to make the spiking neural network algorithms run more precisely and efficiently on neuromorphic chips.
Effective and Efficient Computation with Multiple-timescale Spiking Recurrent Neural Networks. Bojian Yin, Federico Corradi, Sander Bohte. ICONS. Conference proceedings. Open Access-Green.
Accurate and efficient time-domain classification with adaptive spiking recurrent neural networks. Bojian Yin, Federico Corradi, Sander Bohte. In: Nature Machine Intelligence. Journal publication. Open Access-Green
LocalNorm: Robust Image Classification Through Dynamically Regularized Normalization. Bojian Yin, H. Steven Scholte, Sander Bohte.
International Conference on Artificial Neural Networks, 2021. Conference proceedings. Open Access-Green.