Disentangled Representation Learning Based On Causal Inference

Most of the current researches on disentangled representation learning is based on the assumption that the underlying generative factors are independent of each other.They all employ separable latent representations to control generative factors.However,traditional disentangled representation learning cannot explore the causal logic information fully between data because of causally related generative factors in real world.Therefore,causal inference and disentangled representation learning are combined to explore the causal relationships between the underlying generative factors.By using the causal graph to represent the causal structure information between variables and synthesizing causal representations through structural causal models,we complete the causal representation learning and image reconstruction tasks in two stages of disentanglement and generation.Thus,a causal disentangled representation learning model is established that can fully explore causal logic information between data.Currently,there exists an imbalance between disentanglement and generative performance in causal disentangled representation learning,which results in the inability of model to effectively represent causal information among data.To address this issue,we propose a sampling-based differentiable DAG learning algorithm,which efficiently learns the DAG structure among variables.Secondly,a causal representation learning approach based on differentiable DAG sampling is introduced,which improves the disentanglement performance by enhancing the efficiency of causal representation learning.Lastly,a high-fidelity causal disentangled representation learning method is proposed,which utilizes adversarial training to improve image reconstruction quality and enhance the generative performance.This thesis improves the causal disentanglement representation learning model in terms of both disentanglement and generative performance,thereby enhancing the ability of model to learn causal mechanisms among data.The research content and main achievements of the thesis include:(1)To achieve efficient learning of directed acyclic graph(DAG)structures between causal variables,we propose a DAG learning algorithm based on differentiable sampling.First,the nodes are linearly sorted and sampled,and the edges with consistent node ordering are sampled.Then,the equal acyclic constraint is added in the model training stage to avoid repeated ineffective training in the representation learning process.Finally,combined with variational inference,the posterior probability of the DAG directed edge in observable data is approximated to learn the DAG structure from observable data.Compared with existing algorithms,this research can achieve fast and differentiable DAG sampling without complex augmented Lagrangian iterative learning,which is suitable for continuous optimization.Extensive experiments on ten synthetic datasets and the Sach real dataset show that this algorithm significantly improves the ability of model to learn the DAG structure.(2)To improve the disentanglement performance of the causal disentangled representation learning model,we propose a causal representation learning algorithm based on differentiable sampling.In the inference stage,the differentiable sampling DAG learning algorithm is used to learn the DAG with the help of the acyclic equal constraint,and the causal representation is generated through a structural causal model.In the generation stage,the causal information is transmitted through the reconstruction layer,and the causal representation is fed into the generator to synthesize the reconstruction image.The experimental results of the model on the Pendulum synthetic dataset and the Celeb A real dataset show that this network can quickly output the DAG and significantly improve the training speed.The learned causal representation is interpretable semantically,and counterfactual data can be generated by intervention operation.The disentanglement performance of the model is significantly improved.(3)To solve the imbalance problem between the disentanglement performance and the generation performance of causal disentangled representation learning models,we propose a high-fidelity causal disentangled representation learning framework.In the inference stage,the assumption of independent latent variables adopted by most previous methods is abandoned,and a causal layer is added to the encoder to learn causal representation.In the generation stage,the idea of generative adversarial network is adopted,and high-fidelity images are generated through adversarial training between the generator and discriminator.Experiments conducted on the CANDLE synthetic dataset and Celeb A real dataset demonstrate that causal disentangled generation factors are successfully extracted through the inference network,and the learned causal representation is interpretable semantically,and high-fidelity images can be generated through adversarial training while ensuring the disentanglement performance.