# Towards Modular Machine Learning Solution Development: Benefits and Trade-offs

**Authors:** Samiyuru Menik, Lakshmish Ramaswamy  
**arXiv:** [2301.09753](https://arxiv.org/abs/2301.09753)  
**Date:** 2023-01-23  
**Summary page:** https://www.summarizepaper.com/en/paper/2301.09753v1/

## Key points

- Machine learning technologies have demonstrated immense capabilities in various domains and play a key role in the success of modern businesses.
- Adoption of machine learning technologies has untapped potential due to high cost of developing custom solutions that solve unique business problems.
- Monolithic nature prevalent in today's machine learning applications stands in the way of efficient and cost-effective customized solution development.
- Modular machine learning solutions can overcome some major solution engineering limitations of monolithic solutions.
- Trade-offs between modular and monolithic machine learning solutions were analyzed through three deep learning problems; one text-based and two image-based.
- Experimental results show that modular machine learning solutions have promising potential to reap the solution engineering advantages of modularity while gaining performance and data advantages in a way that monolithic machine learning solutions do not permit.
- Modular solutions can reuse existing pre-trained models and exploit more data to achieve higher accuracy and overcome data limitations.
- Black-box knowledge distillation is used to overcome performance challenges that modular solutions may face, showing its impact on accuracy and latency compared to original monolithic and modular solutions.
- Further investigation into multi-stage modular machine learning solution development is suggested as an interesting avenue for addressing systems engineering concerns related to wider adoption of these technologies in diverse organizations.
- Semantic modularity in machine learning is suggested as an interesting avenue to address issues by providing a more efficient and cost-effective approach for developing customized machine-learning-based solutions.

## Summary

In today's business landscape, machine learning technologies have demonstrated immense capabilities in various domains and play a key role in the success of modern businesses. However, the adoption of machine learning technologies has a lot of untapped potential due to the high cost of developing custom machine learning solutions that solve unique business problems. The monolithic nature prevalent in today's machine learning applications stands in the way of efficient and cost-effective customized solution development. To address this issue, this work explores the benefits of modular machine learning solutions and discusses how they can overcome some of the major solution engineering limitations of monolithic machine learning solutions. The authors analyze the trade-offs between modular and monolithic machine learning solutions through three deep learning problems; one text-based and two image-based. The experimental results show that modular machine learning solutions have promising potential to reap the solution engineering advantages of modularity while gaining performance and data advantages in a way that monolithic machine learning solutions do not permit. Modular solutions can reuse existing pre-trained models and exploit more data to achieve higher accuracy and overcome data limitations. Furthermore, black-box knowledge distillation is used to overcome performance challenges that modular solutions may face, showing its impact on accuracy and latency compared to original monolithic and modular solutions. The authors suggest further investigation into multi-stage modular machine learning solution development as an interesting avenue for addressing systems engineering concerns related to wider adoption of these technologies in diverse organizations. In conclusion, while acknowledging the immense potential and positive impact of current deep learning technologies, this work highlights the challenges and limitations of widespread monolithic deep learning technologies with respect to systems engineering concerns. It suggests semantic modularity in machine learning as an interesting avenue to address these issues by providing a more efficient and cost-effective approach for developing customized machine-learning-based solutions.