Self-Supervised Learning of Pretext-Invariant Representations (PIRL) aims to learn image representations that are invariant to image transformations, unlike traditional pretext tasks that encourage covariance. PIRL uses a jigsaw puzzle pretext task to learn invariant representations, which outperform covariant counterparts in various vision tasks. The method is evaluated on multiple benchmarks and shows state-of-the-art results in self-supervised learning, even surpassing supervised pre-training in object detection. PIRL uses a contrastive loss function with a memory bank of negative samples to encourage invariance. It is effective with both Jigsaw and Rotation pretext tasks and performs well on image classification and object detection. PIRL's representations are invariant to transformations and retain semantic information, leading to improved performance across tasks. The method is robust to changes in data distribution and outperforms prior approaches in terms of accuracy and efficiency. PIRL's results demonstrate the potential of self-supervised learning with invariant representations for image recognition tasks.Self-Supervised Learning of Pretext-Invariant Representations (PIRL) aims to learn image representations that are invariant to image transformations, unlike traditional pretext tasks that encourage covariance. PIRL uses a jigsaw puzzle pretext task to learn invariant representations, which outperform covariant counterparts in various vision tasks. The method is evaluated on multiple benchmarks and shows state-of-the-art results in self-supervised learning, even surpassing supervised pre-training in object detection. PIRL uses a contrastive loss function with a memory bank of negative samples to encourage invariance. It is effective with both Jigsaw and Rotation pretext tasks and performs well on image classification and object detection. PIRL's representations are invariant to transformations and retain semantic information, leading to improved performance across tasks. The method is robust to changes in data distribution and outperforms prior approaches in terms of accuracy and efficiency. PIRL's results demonstrate the potential of self-supervised learning with invariant representations for image recognition tasks.