Learning Network Architecture for Open-Set Recognition
Authors: Xuelin Zhang, Xuelian Cheng, Donghao Zhang, Paul Bonnington, Zongyuan Ge3362-3370
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the benefits of this learning pipeline on 5 OSR datasets, including MNIST, SVHN, CIFAR10, CIFARAdd10, and CIFARAdd50, where our approach outperforms prior state-of-the-art networks designed by humans. We evaluated our model on 5 common OSR datasets and the model is able to outperform state-of-the-art OSR methods on all datasets while having fewer parameters than the baseline model. The ablation studies show the benefits of different components designed for the open-set problem in our method. |
| Researcher Affiliation | Academia | Xuelin Zhang1,3, Xuelian Cheng1,3, Donghao Zhang2,3, Paul Bonnington2, Zongyuan Ge2,3 1Monash University, 2Monash e Research Centre, 3Monash Medical AI {xuelin.zhang,xuelian.cheng,donghao.zhang,paul.bonnington,zongyuan.ge}@monash.edu |
| Pseudocode | Yes | Algorithm 1: SEARCHING PIPELINE |
| Open Source Code | Yes | To spark research in this field, our code is available at https://github.com/zxl101/NAS OSR. |
| Open Datasets | Yes | Dataset We conduct the architecture search on CIFAR10 dataset (Krizhevsky and Hinton 2009). Once the search is completed, we adopt our searched best architecture to conduct evaluations on MNIST (LECUN 2012), SVHN (Netzer et al. 2011), CIFAR10, CIFARAdd10 (C+10), and CIFARAdd50 (C+50). |
| Dataset Splits | Yes | We first use two disjoint training sets Dtrain p and Dtrain w for network weights w and network architecture parameters {α, β} optimization respectively. Then for each training set, we pick several classes as pseudo KUCs, namely Dtrain p pkuc and Dtrain w pkuc. The rest of original classes are denoted as Dtrain p kkc and Dtrain w kkc. We do alternating optimization for w and {α, β}. ... Algorithm 1: SEARCHING PIPELINE ... validation dataset Dval |
| Hardware Specification | No | The paper mentions running experiments but does not specify any particular hardware components such as GPU models, CPU types, or memory. |
| Software Dependencies | No | The paper does not list specific software dependencies with version numbers (e.g., Python version, library versions like PyTorch or TensorFlow). |
| Experiment Setup | Yes | The loss functions used in our model are summarized as follows: L = γce Lce + γkl Lkl + γre Lre + γcr Lcr (4) where γce, γkl, γre, γcr are coefficients of different losses. We fix the values of γkl and γre, and have only tried γce and γcr from a small range of numbers. ... We perform the architecture search for a total of 60 epochs. The first 30 epochs are to pretrain the network weights w and the rest 30 epochs targets for alternatively updating α, β and w. |