Prakash, Surya and Mehta, Utkal V. and Sharma, Bibhya N. (2025) Blending additive and multiplicative gates: new flexGate - long short - term memory architecture. Engineered Science, NA . NA. ISSN 2576-988X
Full text not available from this repository.Abstract
In many real-world applications, sequential data exhibit additive and multiplicative dependencies between features and their temporal context. Traditional Long Short-Term Memory (LSTM) networks update their state through additive interactions modulated by multiplicative gates, which can limit flexibility when stronger feature interactions are present. Conversely, architecture such as the Multiplicative-integration Recurrent Neural Network (Mi-RNN) relies purely on multiplicative fusion, often sacrificing stability and interpretability. We introduce FlexGate-LSTM, a recurrent architecture that adaptively blends additive and multiplicative operations inside each gate. A learnable parameter vector in each gate continuously tunes the trade-off between the two interaction modes, allowing the network to select the most suitable integration strategy for the current task or time step. The proposed FlexGate-LSTM is evaluated in five controlled synthetic scenarios: additive, multiplicative, conditional, noisy, and non-stationary, where it matches or exceeds the performance of both vanilla LSTM and Mi-RNN baselines. To demonstrate real-world usefulness, we further test the model on the ETTh1 electricity-transformer temperature dataset. The investigation shows that the FlexGate-LSTM performs competitively in specialized cases and significantly outperforms other architectures when the data exhibits mixed or uncertain dynamics. In addition, the analysis of the learnt parameters provides insight into the internal adaptation strategies of the model, improving the interpretability.
| Item Type: | Journal Article |
|---|---|
| Subjects: | Q Science > QA Mathematics > QA75 Electronic computers. Computer science T Technology > T Technology (General) |
| Divisions: | School of Information Technology, Engineering, Mathematics and Physics (STEMP) |
| Depositing User: | Ms Shalni Sanjana |
| Date Deposited: | 24 Oct 2025 03:11 |
| Last Modified: | 24 Oct 2025 03:11 |
| URI: | https://repository.usp.ac.fj/id/eprint/15187 |
Actions (login required)
 |
View Item |