Auxiliary Image Regularization for Deep CNNs with Noisy Labels
Authors: Samaneh Azadi, Jiashi Feng, Stefanie Jegelka, Trevor Darrell
ICLR 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Comprehensive experiments on benchmark data sets clearly demonstrate our proposed regularized CNN model is resistant to label noise in training data. |
| Researcher Affiliation | Academia | 1 Department of EECS, University of California, Berkeley 2 Department of ECE, National University of Singapore 3 Department of EECS, Massachusetts Institute of Technology |
| Pseudocode | No | The paper provides mathematical equations for optimization updates but does not include a clearly labeled 'Pseudocode' or 'Algorithm' block. |
| Open Source Code | No | The paper does not provide any explicit statement or link for the availability of its source code. |
| Open Datasets | Yes | First, we conduct image classification on a subset of the Image Net7k data set (Deng et al., 2010). ... We perform similar experiment on the MNIST data... We also trained the deep CNN with AIR on the CIFAR-10 data... on publicly available multi-label NUS-WIDE-LITE data set (Chua et al., July 8-10, 2009). |
| Dataset Splits | Yes | Furthermore, we set λ1 in Eq. (3) to a very small number, λg equal to 10 over the length of each feature vector, and finetune the other set of hyper-parameters (batch size, β {1.1, 1.3, 1.5} , and ρmax) in the SADMM updates of Eq. (5) on the cross-validation set for each experiment. We cross-validate the regularization parameter C for our SVM baseline from the set of {1, 10, 100}. |
| Hardware Specification | No | The paper does not specify the hardware used for running the experiments (e.g., specific CPU, GPU, or memory models). |
| Software Dependencies | No | The paper mentions using 'Alex Net CNN model' and 'Cuda Conv network', and refers to 'Caffe', but does not provide specific version numbers for any software dependencies. |
| Experiment Setup | Yes | Furthermore, we set λ1 in Eq. (3) to a very small number, λg equal to 10 over the length of each feature vector, and finetune the other set of hyper-parameters (batch size, β {1.1, 1.3, 1.5} , and ρmax) in the SADMM updates of Eq. (5) on the cross-validation set for each experiment. The initial value for ρ is 10 in all experiments. We cross-validate the regularization parameter C for our SVM baseline from the set of {1, 10, 100}. We define our loss function to be a softmax and apply the AIR regularizer on the top layer in the CNN model. |