Beyond Learning Features: Training a Fully-Functional Classifier with ZERO Instance-Level Labels
Authors: Deepak Babu Sam, Abhinav Agarwalla, Venkatesh Babu Radhakrishnan2162-2170
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments reveal that the method works on major nominal as well as ordinal classification datasets and delivers significant performance. We evaluate our approach on standard classification datasets employed by the unsupervised learning community. |
| Researcher Affiliation | Academia | Video Analytics Lab, Department of Computational and Data Sciences, Indian Institute of Science, Bangalore, India deepaksam@iisc.ac.in, agarwallaabhinav@gmail.com, venky@iisc.ac.in |
| Pseudocode | No | The paper describes the method conceptually and with diagrams (e.g., Figure 3: Training stages), but does not contain a structured pseudocode block or an explicitly labeled algorithm. |
| Open Source Code | No | The paper does not provide an unambiguous statement or a direct link to the source code for the methodology described in this paper. The only link provided (https://github.com/explosion/spaCy) refers to a third-party library used. |
| Open Datasets | Yes | Datasets: We evaluate our approach on standard classification datasets employed by the unsupervised learning community. First is the CIFAR-10 (Krizhevsky, Hinton et al. 2009) dataset... (and other dataset citations). |
| Dataset Splits | No | For Adience: "We report the performance using the cross-validation splits released with the data." For Aesthetics: "We split the dataset into training and testing set in an 80-20 ratio." For Eye PACS: "...divide the dataset into training and validation images (using a 90-10 split) and report the results on the validation split." While some splits are mentioned, they are not consistently provided (e.g., no validation split for Aesthetics, specific details of "cross-validation splits" for Adience are not provided, and for standard datasets like CIFAR-10, no split is explicitly stated). |
| Hardware Specification | No | The paper does not provide specific details about the hardware (e.g., GPU models, CPU types, or memory) used for running its experiments. It only mentions the use of ResNet-50 as the backbone network. |
| Software Dependencies | No | The paper mentions using 'Adam optimizer' and the 'spaCy' library but does not provide specific version numbers for these or any other software dependencies, which are required for full reproducibility. |
| Experiment Setup | Yes | We use the Adam (Kingma and Ba 2014) optimizer with learning rate of 10 3 and a batch size of 128 for all datasets except CIFAR100-20... We use a batch size of 512 for CIFAR100-20... Here the sinkhorn clustering employs Adam optimizer with learning rate of 0.1 for a total of 5K steps... To stabilize the training, the learning rate is set to 10 5. For distribution matching, we use a batch size of 500, entropy regularization constant β as 0.01 and training step as 10K. The three most negative classes are also randomly shuffled while sampling from the prior P. We clip the gradient norms above the value of 100. |