Shallow RNN: Accurate Time-series Classification on Resource Constrained Devices
Authors: Don Dennis, Durmus Alp Emre Acar, Vikram Mandikal, Vinu Sankar Sadasivan, Venkatesh Saligrama, Harsha Vardhan Simhadri, Prateek Jain
NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct experiments to study: a) performance of MI-Sha RNN with varying hidden state dimensions at both the layers Rp1q and Rp2q to understand how its accuracy stacks up against baseline models across different model sizes, b) inference cost improvement that MI-Sha RNN produces for standard time-series classification problems over baseline models and MI-RNN models, c) if MI-Sha RNN can enable certain time-series classification tasks on devices based on the tiny Cortex M4 with only 100MHz processor and 256KB RAM. |
| Researcher Affiliation | Collaboration | Don Kurian Dennis Carnegie Mellon University Durmus Alp Emre Acar Boston University Vikram Mandikal: University of Texas at Austin Vinu Sankar Sadasivan: IIT Gandhinagar Harsha Vardhan Simhadri Microsoft Research India Venkatesh Saligrama Boston University Prateek Jain Microsoft Research India |
| Pseudocode | No | The paper does not contain a clearly labeled 'Pseudocode' or 'Algorithm' section or block. |
| Open Source Code | Yes | The implementation of our algorithm is released as part of the Edge ML [11] library. |
| Open Datasets | Yes | Datasets: We benchmark our method on standard datasets from different domains like audio keyword detection (Google-13), wake word detection (STCI-2), activity recognition (HAR-6), sports activity recognition (DSA-19), gesture recognition (Gesture Pod-5). The number after hyphen in dataset name indicates the number of classes in the dataset. See Table 3 in appendix for more details about the datasets. All the datasets are available online (see Table 3) except STCI-2 which is a proprietary wake word detection dataset. |
| Dataset Splits | No | The paper mentions training data and test results, but does not explicitly provide details about training/validation/test splits, such as percentages or sample counts for each partition. |
| Hardware Specification | Yes | For example, we can deploy audio-keyword classification on tiny Cortex M4 devices (100MHz processor, 256KB RAM, no DSP available) which was not possible using standard RNN models. |
| Software Dependencies | No | We implemented all the algorithms on Tensor Flow and used Adam for training the models [19]. The paper mentions 'Tensor Flow' but does not specify a version number. |
| Experiment Setup | Yes | The main hyperparameters are: a) hidden state sizes for both the layers of MI-Sha RNN. b) brick-size k for MI-Sha RNN. In addition, the number of time-steps T is associated with each dataset. MI-RNN prunes down T and works with T 1 ď T time-steps. |