Discovering and Overcoming Limitations of Noise-engineered Data-free Knowledge Distillation

Authors: Piyush Raikwar, Deepak Mishra

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate our approach on CIFAR10, CIFAR100, SVHN, and Food101 datasets.
Researcher Affiliation Academia Piyush Raikwar ABV-IIITM, Gwalior, India imt_2017062@iiitm.ac.in Deepak Mishra IIT Jodhpur, India dmishra@iitj.ac.in
Pseudocode Yes Algorithm 1 Training KD and Algorithm 2 Evaluation
Open Source Code Yes Code is available at: https://github.com/Piyush-555/Gaussian Distillation
Open Datasets Yes We validate our approach on CIFAR10, CIFAR100, SVHN, and Food101 datasets.
Dataset Splits No The paper mentions "validation set" once in passing, and refers to "test data" and "training data" subsets for finetuning, but does not provide specific train/validation/test split percentages, sample counts, or explicit predefined split citations for reproducibility of data partitioning.
Hardware Specification No No specific hardware details (GPU/CPU models, memory amounts, or detailed computer specifications) are provided.
Software Dependencies No The paper mentions using "Adam optimizer" but does not provide specific software dependencies like framework versions (e.g., PyTorch 1.9) or other library versions needed for replication.
Experiment Setup Yes In both cases, the batch size is 256, and an Adam optimizer with a learning rate of 10^-3 for tuning the parameters of the student network is used. For finetuning, a subset of training data is sampled randomly and a reduced learning rate of 10^-4 is used.