Abstract:Causal disentangled representation learning is a key technique for extracting causally independent latent factors from high-dimensional data, with important applications in tasks such as controllable generation and robust classification. Although existing causal disentangled models can capture relationships among latent factors using structural causal models, they often suffer from insufficient suppression of feature redundancy and a lack of closed-loop consistency, thus limiting their application in critical fields such as medical diagnosis and autonomous driving. To address these issues, this study proposes a causal disentangled representation learning framework integrating closed-loop transcription and sparse coding. Its core design includes two parts. First, a closed-loop transcription mechanism is constructed based on the principle of maximal coding rate reduction. The bidirectional interaction and dynamic feedback between the encoder and decoder effectively improve disentanglement accuracy and stability. Second, a convolutional sparse coding layer is introduced to replace standard convolutional layers in traditional deep neural networks, effectively reducing feature redundancy. Extensive experiments are conducted on the Pendulum synthetic dataset and the CelebA real-world dataset. The sample efficiency reaches 99.11% on the Pendulum dataset and 99.32% on the CelebA dataset. Experimental results show that the proposed framework outperforms existing methods in terms of sample efficiency and distributional robustness in downstream tasks, providing a superior solution for the practical application of causal disentangled representations.