SEENN: Towards Temporal Spiking Early Exit Neural Networks

Authors: Yuhang Li, Tamar Geller, Youngeun Kim, Priyadarshini Panda

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To demonstrate the efficacy and the efficiency of our SEENN, we conduct experiments on popular image recognition datasets, including CIFAR10, CIFAR100 [47], Image Net [48], and an event-stream dataset CIFAR10-DVS [49].
Researcher Affiliation Academia Yuhang Li Yale University New Haven, CT, USA yuhang.li@yale.edu Tamar Geller Yale University New Haven, CT, USA tamar.geller@yale.edu Youngeun Kim Yale University New Haven, CT, USA youngeun.kim@yale.edu Priyadarshini Panda Yale University New Haven, CT, USA priya.panda@yale.edu
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes Code is shared at https://github.com/Intelligent-Computing-Lab-Yale/SEENN.
Open Datasets Yes To demonstrate the efficacy and the efficiency of our SEENN, we conduct experiments on popular image recognition datasets, including CIFAR10, CIFAR100 [47], Image Net [48], and an event-stream dataset CIFAR10-DVS [49].
Dataset Splits Yes For example, our SEENN-II Res Net-19 can achieve 96.1% accuracy with an average of 1.08 timesteps on the CIFAR-10 test dataset. and N = |CT | + |W| is the total number of samples in the validation dataset. and In Fig. 2, we demonstrate the distribution of confidence scores of Res Net-19 on the CIFAR-10 validation dataset.
Hardware Specification Yes we directly use a GPU (NVIDIA Tesla V100) to evaluate the latency (or throughput) of SEENN.
Software Dependencies No The paper mentions software components like 'stochastic gradient descent optimizer', 'cosine annealing schedule', 'Cutout', and 'Auto Augment' but does not specify version numbers for these or other key software dependencies like PyTorch or Python.
Experiment Setup Yes All models are trained with a stochastic gradient descent optimizer with a momentum of 0.9 for 300 epochs. The learning rate is 0.1 and decayed following a cosine annealing schedule [50]. The weight decay is set to 5e 4. For ANN pre-training in QCFS, we set the step l to 4. We use Cutout [51] and Auto Augment [52] for better accuracy as adopted in [31,32].