This paper proposes a joint deep learning and clustering approach for learning K-means-friendly latent representations. The key idea is to jointly optimize dimensionality reduction (DR) and K-means clustering by learning a deep neural network (DNN) to approximate the nonlinear transformation from the latent space to the data space. Unlike previous methods that assume a linear transformation, this approach treats the transformation as an unknown nonlinear function. The proposed method includes a reconstruction term to avoid trivial solutions and a clustering regularization term to promote K-means-friendly latent representations. The optimization is performed using an alternating stochastic gradient algorithm with effective initialization. The method is validated on both synthetic and real-world datasets, showing significant improvements over existing methods. The results demonstrate that the proposed approach can effectively recover K-means-friendly latent representations even under complex nonlinear generative models. The method is scalable and can be applied to various types of data, including text and images. The paper also discusses the importance of joint optimization in DR and clustering, and highlights the advantages of using deep neural networks for this task.This paper proposes a joint deep learning and clustering approach for learning K-means-friendly latent representations. The key idea is to jointly optimize dimensionality reduction (DR) and K-means clustering by learning a deep neural network (DNN) to approximate the nonlinear transformation from the latent space to the data space. Unlike previous methods that assume a linear transformation, this approach treats the transformation as an unknown nonlinear function. The proposed method includes a reconstruction term to avoid trivial solutions and a clustering regularization term to promote K-means-friendly latent representations. The optimization is performed using an alternating stochastic gradient algorithm with effective initialization. The method is validated on both synthetic and real-world datasets, showing significant improvements over existing methods. The results demonstrate that the proposed approach can effectively recover K-means-friendly latent representations even under complex nonlinear generative models. The method is scalable and can be applied to various types of data, including text and images. The paper also discusses the importance of joint optimization in DR and clustering, and highlights the advantages of using deep neural networks for this task.