Fine-Grained Zero-Shot Learning with DNA as Side Information
Authors: Sarkhan Badirli, Zeynep Akata, George Mohler, Christine Picard, Mehmet M Dundar
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | On the benchmark CUB dataset we show that DNA can be equally promising, yet in general a more accessible alternative than word vectors as a side information. On a newly compiled fine-grained insect dataset that uses DNA information from over a thousand species we show that the Bayesian approach outperforms state-of-the-art by a wide margin. |
| Researcher Affiliation | Academia | Sarkhan Badirli Department of Computer Science Purdue University, West Lafayette, IN, USA Indiana University-Purdue University Indianapolis, IN, USA Zeynep Akata University of Tübingen Max Planck Institute for Informatics Max Planck Institute for Intelligent Systems George Mohler Computer and Information Science Department Indiana University Purdue University Indianapolis, IN, USA Christine J. Picard Department of Biology Indiana University Purdue University Indianapolis, IN, USA Murat Dundar Computer and Information Science Department Indiana University Purdue University Indianapolis, IN, USA mdundar@iupui.edu |
| Pseudocode | No | The paper describes the generative model using mathematical equations but does not include any pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | Yes | The code and dataset are available at https://github.com/sbadirli/Fine-Grained-ZSL-with-DNA |
| Open Datasets | Yes | In this study, we present the fine-grained INSECT dataset with 21, 212 matching image/DNA pairs from 1, 213 species (see Fig. 1 for sample images). All insect images and associated DNA barcodes in our dataset come from the Barcode of Life Data System (BOLD) (35; 36). The code and dataset are available at https://github.com/sbadirli/Fine-Grained-ZSL-with-DNA |
| Dataset Splits | Yes | We randomly chose 10% of all species as unseen classes for the test set leading to 1, 092 seen and 121 unseen classes. Similarly, we randomly chose 10% of the 1, 092 training classes as unseen classes for the validation set. Samples from seen classes were split by a 80/20 ratio in a stratified fashion to create seen portion of the train and test datasets. (...) The training set is further split into two as train (80%) and validation (20%) by random sampling. |
| Hardware Specification | No | The paper mentions that training 'takes a few hours on a relatively sophisticated GPU' but does not provide specific details such as the GPU model, CPU type, or memory specifications. |
| Software Dependencies | Yes | all sequences are aligned with respect to this consensus sequence using a progressive alignment technique implemented in MATLAB R2020A (Natick, MA, USA). The model is developed in Python with Tensorflow-Keras API. |
| Experiment Setup | Yes | We used ADAM optimizer for training the model for five epochs with a batch size of 32 (with a step-decay initial learning rate = 0.0005 and drop factor= 0.5, β1 = 0.9, β2 = 0.999). The CNN architecture is completed by a softmax layer. We used the output of the fully-connected layer as the embeddings for DNA. Class level attributes are computed by the mean embedding of each class. (...) Images are resized to 256 256 and center-cropped before fed to the Res Net model. No other pre-processing is applied to the images. |