LCM : A Latent-Connected MLP Architecture for Universal Deep Learning with Fast Convergence and low Computational Cost

This repository presents the LCM (Latent-Connected Model), a universal deep learning architecture designed to bridge the gap between traditional MLPs and modern attention-based models.

The LCM introduces a latent connection mechanism that merges dual perception of logits and connects residual outputs between past and present representations, enabling efficient learning across multiple domains — including NLP, Computer Vision, and Tabular forecasting tasks.

Compared to Transformer and RNN-based architectures, LCM offers faster convergence and lower computational cost, achieved through purely linear operations combined with gated and residual mechanisms.

This work includes both the full research paper and implementation code (PyTorch) for reproducibility and benchmarking.

Keywords: Latent Connected Model, Deep Learning, MLP, Universal Architecture, Transformers, AI Research, Neural Network, Fast Convergence, Low Computational Cost