# Closed-form Continuous-Depth Models

**Authors:** Ramin Hasani, Mathias Lechner, Alexander Amini, Lucas Liebenwein, Max Tschaikowski, Gerald Teschl, Daniela Rus  
**arXiv:** [2106.13898](https://arxiv.org/abs/2106.13898)  
**Date:** 2021-06-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2106.13898v1/

## Key points

- Introduction of Closed-form Continuous-depth (CfC) networks
- Overcoming limitations of existing continuous-depth neural models
- Derived from analytical closed-form solution, eliminating the need for complex DE solvers
- CfC networks are simpler to describe and significantly faster than ODE-based counterparts
- Equally strong modeling abilities as advanced recurrent models
- Outperforming recurrent models in various time-series prediction tasks
- Potential deployment in resource-constrained settings where performance and efficiency are crucial
- Application of CfC networks in PhysioNet Challenge 2012 dataset for predicting ICU patient mortality based on time series medical measurements
- Application of CfC networks in sentiment analysis using the IMD dataset
- Innovative approach to training and deploying continuous neural models by leveraging closed-form solutions instead of complex DE solvers

## Summary

In this paper, the authors introduce a new family of models called Closed-form Continuous-depth (CfC) networks. These networks are designed to process sequential data and overcome the limitations of existing continuous-depth neural models that rely on advanced numerical differential equation solvers. The CfC networks are derived from the analytical closed-form solution of a subset of time-continuous models, eliminating the need for complex DE solvers. The authors demonstrate that CfC networks are not only simpler to describe but also significantly faster compared to their ODE-based counterparts, with at least one order of magnitude reduction in computational cost. Despite their simplicity and efficiency, CfC networks exhibit equally strong modeling abilities as advanced recurrent models. Experimental evaluations show that CfC networks outperform recurrent models in various time-series prediction tasks, including those with long-term dependencies and irregularly sampled data. This suggests that CfC networks have the potential to be deployed in resource-constrained settings where both performance and efficiency are crucial. Furthermore, the authors discuss the application of CfC networks in two specific domains. Firstly, they apply these models to the PhysioNet Challenge 2012 dataset, which involves predicting the mortality of ICU patients based on time series medical measurements. Secondly, they explore sentiment analysis using the IMD dataset. Overall, this research presents an innovative approach to training and deploying continuous neural models by leveraging closed-form solutions instead of relying on complex DE solvers. The findings highlight the potential for CfC networks to improve performance and efficiency in various sequential data processing tasks.