LTC-SE: Expanding the Potential of Liquid Time-Constant Neural Networks for Scalable AI and Embedded Systems
This work addresses the need for efficient neural network implementations in resource-constrained embedded systems, but it is incremental as it builds on existing foundational algorithms.
The paper tackles the challenge of making Liquid Time-Constant neural networks more practical for scalable AI and embedded systems by introducing LTC-SE, an improved version that enhances flexibility, compatibility, and code organization, resulting in better user experience and integration with frameworks like TensorFlow 2.x.
We present LTC-SE, an improved version of the Liquid Time-Constant (LTC) neural network algorithm originally proposed by Hasani et al. in 2021. This algorithm unifies the Leaky-Integrate-and-Fire (LIF) spiking neural network model with Continuous-Time Recurrent Neural Networks (CTRNNs), Neural Ordinary Differential Equations (NODEs), and bespoke Gated Recurrent Units (GRUs). The enhancements in LTC-SE focus on augmenting flexibility, compatibility, and code organization, targeting the unique constraints of embedded systems with limited computational resources and strict performance requirements. The updated code serves as a consolidated class library compatible with TensorFlow 2.x, offering comprehensive configuration options for LTCCell, CTRNN, NODE, and CTGRU classes. We evaluate LTC-SE against its predecessors, showcasing the advantages of our optimizations in user experience, Keras function compatibility, and code clarity. These refinements expand the applicability of liquid neural networks in diverse machine learning tasks, such as robotics, causality analysis, and time-series prediction, and build on the foundational work of Hasani et al.