2IQRnn13

2IQRnn13 is a neural network architecture designed for robust sequential data modeling. The designation combines two interquartile-range (IQR) based processing blocks with a deep recurrent backbone, commonly interpreted as a 13-layer variant of an RNN.

Architecture: The model comprises two parallel IQR modules that operate on input features and hidden states to extract robust statistics, which are then fused and passed to a 13-layer recurrent stack (LSTM or GRU cells are typical). The network supports unidirectional and bidirectional modes and can be configured for sequence-to-sequence tasks.

Training and loss: It uses a robust residual loss based on the interquartile range of prediction errors, reducing the influence of outliers. Training is typically performed with Adam or AdamW, with gradient clipping, dropout, and early stopping.

Applications and performance: Demonstrated on synthetic datasets with heavy-tailed noise and real-world time-series such as financial data and sensor streams; reported to improve resilience to outliers and maintain accuracy on clean data relative to standard LSTM/GRU baselines.

Variants and status: There exist simplified variants with fewer layers (2IQRnn9) and expanded versions (2IQRnn20); the concept has seen exploration within academic and hobbyist communities.