LLC: Accurate, Multi-purpose Learnt Low-dimensional Binary Codes

Authors: Aditya Kusupati, Matthew Wallingford, Vivek Ramanujan, Raghav Somani, Jae Sung Park, Krishna Pillutla, Prateek Jain, Sham Kakade, Ali Farhadi

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate that the learnt codes capture intrinsically important features in the data, by discovering an intuitive taxonomy over classes. We further quantitatively measure the quality of our codes by applying it to the efficient image retrieval as well as out-of-distribution (OOD) detection problems. For Image Net-100 retrieval problem, our learnt binary codes outperform 16 bit Hash Net using only 10 bits and also are as accurate as 10 dimensional real representations.
Researcher Affiliation Collaboration University of Washington, Google Research India {kusupati,mcw244,ramanv,raghavs,jspark96,pillutla,sham,ali}@cs.washington.edu, prajain@google.com
Pseudocode Yes Algorithm 1 The LLC Method
Open Source Code Yes Code is open-sourced at https://github.com/RAIVNLab/LLC.
Open Datasets Yes Image Net-1K [49] is a widely used image classification dataset with 1000 hierarchical classes. Our classification experiments use Res Net50 [25] and are trained using the 1.3M training images.
Dataset Splits Yes We evaluate all the methods using top-1 accuracy on the Image Net-1K validation set. [...] Image Net-100 has 100 classes randomly sampled from Image Net-1K. All the validation images of these classes are used as query images, all the training images (~1300 per class) of these 100 classes are used as database images. Finally, 130 samples per class from the database are used as the training set for learning binary codes or hashing functions.
Hardware Specification Yes All the implementations were in Py Torch [45] and experimented on a machine with 4 NVIDIA Titan X (Pascal) GPUs.
Software Dependencies No The paper mentions 'Py Torch [45]' but does not provide a specific version number for it or any other software dependencies.
Experiment Setup Yes See Appendix C for the hyperparameter values and other training details.