Recursive Transitional Gated Unit (RTGU): A Novel Recurrent Architecture Demonstrating Competitiveness Against LSTM and GRU

This paper presents a comprehensive study focused on the comparative performance between the traditional Long Short-Term Memory (LSTM) network and the novel Recursive Transitional Gated Unit (RTGU) model under various controlled conditions, with emphasis on parameter budgets. The experiments observe convergence speed, predictive accuracy, and parametric efficiency and expressivity. This particular novel structure demonstrates consistent equal or superior performance relative to the LSTM — even with fewer parameters. These findings indicate a potential for broader adoption of the RTGU as an alternative to the LSTM, offering improved efficiency without compromising predictive capability for at least two tasks.