Gregor Lenz

SynSense

• Company focuses on ultra-low-power inference on the edge
• Hardware is digital asynchronous for Speck and digital synchronous for
•  

What I do

• Designing the machine learning pipeline to train our models
   
• Focus on engineering the tools for gradient-based optimization of SNNs, data wrangling
   
• Establish the use of good practices to make training and deployment faster

Research

• Training SNNs for better performance through data augmentation
   
• Architectural exploration such as recurrence in the network, spike frequency adaptation
   
• How to save energy, fewer spikes is the goal

Research

• Training Spiking Neural Networks Using Lessons from Deep Learning: Eshragian, Neftci, Wang, Lenz
   
• Adversarial Attacks on Spiking Convolutional Networks: Büchel, Lenz, Sheik, Sorbaro
   
• EXODUS: Stable and Efficient Training of Spiking Neural Networks: Bauer, Lenz, Haghighatshoar, Sheik
   
• under submission: Ultra-low-power image classification on neuromorphic hardware

Patents

• Converting an ANN to an SNN using temporal coding and a single spike per neuron
   
• A method to normalise ANN activation which scales to deeper layers

Outreach

• CapoCaccia workshop: presented our chips, gave tutorials and demos
   
• AMLD: a full-day workshop with our hardware, outreach to new students
   
• Open Neuromorphic: Platform for neuromorphic open source code and hardware
   
• NeuroBench: benchmarking and metrics for neuromorphic computing

Summary

• Leading role in the algorithms vision team
   
• Strong engineering background
   
• 3 papers published in the last year, 1 under submission
   
• 2 patents submitted
   
• Involved in numerous community efforts