Performance evaluation of efficient interpretable CNN-transformer model for redshift prediction

The popularity of deep learning models precisely transformer-based models has immensely increased in image classification tasks recently. In this article, we investigate the performance of a transformer based model on redshift estimation, a regression task from cosmology. Primarily estimation of redshift through galaxy images is a crucial application to comprehend the structure of our universe. Here, we propose a new variant of deep learning framework consisting of a Convolution Neural network (CNN) coupled with a inception and attention mechanism to estimate redshift from images and photometric features acquired through Sloan Digital Sky Survey (SDSS). The proposed architecture combines the inception block layer together with transformer layers to improve both generalization and capacity of the model and designed specifically to handle mixed input data (image and magnitude features) together in parallel. Extensive experiment shows that model achieves slightly better results in terms of Mean Squared Error (0.026), bias (-0.0002), and competetive Normalized Mean Absolute Deviation (0.013), than contemporary models when tested on our dataset. We empirically show that the proposed model leverages the data better than non-transformer-based models. Furthermore, to gain better insight about inner working procedure of the proposed framework, we applied Saliency map and SHapley Additive exPlanations (SHAP) to explore the essential information in activation maps.