Low-shot Learning via Covariance-Preserving Adversarial Augmentation Networks

Authors: Hang Gao, Zheng Shou, Alireza Zareian, Hanwang Zhang, Shih-Fu Chang

NeurIPS 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical results show that our model can generate realistic yet diverse examples, leading to substantial improvements on the Image Net benchmark over the state of the art.
Researcher Affiliation Academia 1Columbia University, 2Nanyang Technological University
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code No The paper refers to a GitHub link for a ResNet-10 backbone used in experiments (Released on https://github.com/facebookresearch/low-shot-shrink-hallucinate), but this is a third-party component, not the open-source code for the methodology described in this paper.
Open Datasets Yes We evaluate our method on the real-world benchmark proposed by Hariharan et al. [11]. This is a challenging task because it requires us to learn a large variety of Image Net [18]
Dataset Splits Yes Following [11], we split the 1000 Image Net classes into four disjoint class sets Ytest b , Ytest n , Yval b , Yval n , which consist of 193, 300, 196, 311 classes respectively. All of our parameter tuning is done on validation splits, while final results are reported using held-out test splits.
Hardware Specification No The paper does not specify any hardware details (e.g., GPU/CPU models, memory) used for running the experiments.
Software Dependencies No The paper mentions "Our implementation is based on Py Torch [34]" but does not specify its version or any other software dependencies with version numbers.
Experiment Setup Yes Our GAN models are trained for 100000 episodes by ADAM [35] with initial learning rate fixed at 0.0001 which anneals by 0.5 every 20000 episodes. We fix the hyper-parameter m = 10 for computing truncated SVD. For loss term contributions, we set λcyc = 5 and λcov = 0.5 for all final objectives. We choose Z = 100 as the dimension of noise vectors for Gb s input, and C = 50 for the Gaussian mixture. We empirically set batch size B = 1000, and Nb = 20 and Kb = 10 for all training