Copyright © 2026 Authors retain the copyright of this article. This article is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

@article{208572,
        author = {Pratiksha Rajendra Dashpute and Vikrant Satish Salunkhe and Sheetal Shrikant Shevkari},
        title = {Comparative Analysis of Classical Machine Learning Vs. Deep Learning for Resource-Constrained Edge Devices},
        journal = {International Journal of Innovative Research in Technology},
        year = {2026},
        volume = {13},
        number = {no},
        pages = {512-519},
        issn = {2349-6002},
        url = {https://ijirt.org/article?manuscript=208572},
        abstract = {Deploying machine learning models directly onto ultra-low-power microcontrollers and edge hardware (TinyML) is increasingly vital for real-time, privacy-preserving Internet of Things (IoT) applications. However, selecting the optimal algorithm under severe hardware constraints typically under 512 KB of RAM and tight power budgets presents a fundamental trade-off between computational overhead and predictive power. 
This paper presents a systematic comparative analysis evaluating traditional machine learning algorithms (Decision Trees, Random Forests, and Support Vector Machines) against lightweight, quantized Deep Learning architectures (compact CNNs and Multilayer Perceptrons). 
Benchmark experiments are conducted across standardized datasets for tabular, acoustic, and vision tasks deployed on representative microcontroller platforms (e.g., ARM Cortex-M series, ESP32). 
Models are evaluated across three primary dimensions: inference latency, memory footprint (Flash and peak SRAM usage), and predictive accuracy. Findings indicate that classical algorithms achieve significantly lower SRAM overhead and near-instantaneous inference times on structured sensor data, making them highly effective for ultra-low-memory nodes. Conversely, quantized neural networks achieve higher accuracy and generalization on complex spatio-temporal features, though they require higher Flash allocation and non-trivial peak RAM. 
Based on these empirical trade-offs, we propose a lightweight decision matrix to guide model selection based on strict device hardware limits and application demands.},
        keywords = {TinyML, Edge Computing, Machine Learning Benchmarking, Decision Trees, Neural Network Quantization, Inference Latency, Memory Footprint, Microcontrollers},
        month = {September},
        }