The paper introduces a highly efficient method for computing approximate sparse codes, which are essential for feature extraction in various applications such as visual neuroscience and image restoration. The main contribution is a learning-based approach that trains a non-linear, feed-forward predictor to approximate the optimal sparse code, significantly reducing the computational cost compared to traditional methods. The method, called Learned Iterative Shrinkage-Thresholding Algorithm (LISTA), is based on truncated versions of the Iterative Shrinkage and Thresholding Algorithm (ISTA) and Coordinate Descent (CoD). LISTA and its variant, Learned Coordinate Descent (LCoD), are designed to handle overcomplete dictionaries and allow for mutual inhibition between code components, enabling better feature extraction. Experimental results show that LISTA and LCoD achieve much lower prediction errors with significantly fewer computational steps, making them suitable for real-time applications. The method also demonstrates improved performance in object recognition tasks using the MNIST dataset.The paper introduces a highly efficient method for computing approximate sparse codes, which are essential for feature extraction in various applications such as visual neuroscience and image restoration. The main contribution is a learning-based approach that trains a non-linear, feed-forward predictor to approximate the optimal sparse code, significantly reducing the computational cost compared to traditional methods. The method, called Learned Iterative Shrinkage-Thresholding Algorithm (LISTA), is based on truncated versions of the Iterative Shrinkage and Thresholding Algorithm (ISTA) and Coordinate Descent (CoD). LISTA and its variant, Learned Coordinate Descent (LCoD), are designed to handle overcomplete dictionaries and allow for mutual inhibition between code components, enabling better feature extraction. Experimental results show that LISTA and LCoD achieve much lower prediction errors with significantly fewer computational steps, making them suitable for real-time applications. The method also demonstrates improved performance in object recognition tasks using the MNIST dataset.