- Designing an Efficient SNN Learning Algorithm for Vertical Synapse Array Structure
- Suggested SNN learning method using knowledge distillation for resource-constrained environment of 3D-CIM
- Implementation of an algorithm to reduce the occurrence of spikes for energy efficiency in model inference
- Compute capability sensitivity analysis
- Implementation of an artificial neural network inference model that is robust to device noise
- SNN model design robust to time-variant noise generated in 3D CIM
- To address the noise from high-dimensional data and parameters in artificial neural network training, robust optimization techniques are used
- By obtaining observed values that consider the worst-case scenario along with the original values, this approach aims to overcome the limitations of ECOC in neural network training
- Analyze and apply model ensemble technology, which is a type of Bayesian model
Fig 1. System diagram of task performance.
Fig 2. BNN implementation method with -1/1 as activation variable.
Fig 3. BWN-based CIM circuit configuration
Fig 4. Comparison of the accuracy and number of spikes of various neural coding methods when spike deletion (left) and spike jitter (right) noise were applied respectively (VGG16 model, CIFAR-10 dataset).
Fig 5. Accuracy comparison of various neural coding and proposed weight scaling methods when spike deletion (left) and spike jitter (right) noise are applied respectively (VGG16 model, CIFAR-10 dataset).
#Spiking neural network #Object detection #Optimization #Noise robust model
