RNNPool: Efficient Non-linear Pooling for RAM Constrained Inference
Authors: Oindrila Saha, Aditya Kusupati, Harsha Vardhan Simhadri, Manik Varma, Prateek Jain
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical evaluation indicates that an RNNPool layer can effectively replace multiple blocks in a variety of architectures such as Mobile Nets, Dense Net when applied to standard vision tasks like image classification and face detection. Our experiments demonstrate that RNNPool can be used as an effective replacement for multi-layered, expensive CNN blocks in a variety of architectures such as Mobile Nets, Dense Nets, S3FD, and for varied tasks such as image classification and face detection. |
| Researcher Affiliation | Collaboration | Microsoft Research India, University of Washington {t-oisaha,harshasi,manik,prajain}@microsoft.com, kusupati@cs.washington.edu |
| Pseudocode | Yes | Algorithm 1 RNNPool Operation |
| Open Source Code | Yes | Code is released at https://github.com/Microsoft/Edge ML. |
| Open Datasets | Yes | We first focus on Image Net-10, a 10 class subset of Image Net-1K [7] where the classes correspond to the categories in CIFAR-10 [25]. We train and validate these architectures with the WIDER FACE dataset [47]. |
| Dataset Splits | Yes | The dataset is divided into 1300 images for training and 50 for validation per class. Table 4 compares validation Mean Average Precision (MAP) for easy, medium, and hard subsets. |
| Hardware Specification | Yes | RNNPool-Face-M4 model processes a single image in 10.45 seconds on an ARM Cortex-M4 microcontroller based STM32F439-M4 device clocked at 168 MHz. |
| Software Dependencies | No | While the paper mentions using frameworks like PyTorch [34] and a compiler like See Dot [12], it does not provide specific version numbers for these or any other software dependencies crucial for replication. |
| Experiment Setup | No | See Appendix G for more details about model training and hyperparameters used for the experiments. This statement indicates that experimental setup details are located in the appendix rather than the main body of the paper. |