Face Model Compression by Distilling Knowledge from Neurons

Authors: Ping Luo, Zhenyao Zhu, Ziwei Liu, Xiaogang Wang, Xiaoou Tang

AAAI 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments Face Data For face model compression, we train all the students using data the same as (Sun, Wang, and Tang 2014)... In test, we evaluate all the models on LFW (Huang et al. 2007)... The face verification performance on LFW is reported as the Area under ROC curve (AUC)...
Researcher Affiliation Academia Ping Luo1,3 , Zhenyao Zhu1 , Ziwei Liu1, Xiaogang Wang2,3, and Xiaoou Tang1,3 1Department of Information Engineering, The Chinese University of Hong Kong 2Department of Electronic Engineering, The Chinese University of Hong Kong 3Shenzhen Key Lab of Comp. Vis. & Pat. Rec., Shenzhen Institutes of Advanced Technology, CAS, China
Pseudocode No The paper describes the proposed method using equations and textual descriptions, but it does not include any structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any concrete statement or link regarding the public availability of its source code.
Open Datasets Yes For face model compression, we train all the students using data the same as (Sun, Wang, and Tang 2014), which combined two face databases for training, Celeb Faces+ (Sun et al. 2014) and WDRef (Chen et al. 2012), resulting in a training set of 290 thousands face images of 12,294 identities. In test, we evaluate all the models on LFW (Huang et al. 2007)... We employ the Celeb A dataset (Liu et al. 2015) as a validation set
Dataset Splits Yes For face model compression, we train all the students using data the same as (Sun, Wang, and Tang 2014), which combined two face databases for training, Celeb Faces+ (Sun et al. 2014) and WDRef (Chen et al. 2012), resulting in a training set of 290 thousands face images of 12,294 identities. In test, we evaluate all the models on LFW (Huang et al. 2007), which is the most well known benchmark for face recognition, containing 13,233 face images of 5,749 identities collected from the Internet... The face verification performance on LFW is reported as the Area under ROC curve (AUC) with respect to 3,000 positive and 3,000 negative face pairs. We employ the Celeb A dataset (Liu et al. 2015) as a validation set, which contains 20 thousand face images and each image is annotated with 40 attributes.
Hardware Specification Yes Efficiency is measured with implementation on a Intel Core 2.0GHz CPU. To simulate the environment of embedded or portable devices, the runtime is evaluated on CPU instead of GPU.
Software Dependencies No The paper mentions general techniques and components like 'standard back-propagation' and 'rectified linear unit (Nair and Hinton 2010)', but does not provide specific software dependencies with version numbers (e.g., Python, TensorFlow, PyTorch versions).
Experiment Setup Yes We formulate the objective function of model compression as a regression problem... optimized by the stochastic gradient descent with standard back-propagation... We examine different temperatures for soft targets. Note that when t = 1 and t = + , soft targets turn into label probabilities (after softmax) and logits, respectively. The best performance of soft targets is achieved when t = 10.