FractalNet introduces a novel design strategy for neural network architecture based on self-similarity, generating an extremely deep network with a truncated fractal structure. Unlike residual networks, which rely on residual learning, FractalNet achieves high performance without residuals, matching the state-of-the-art error rate on CIFAR-100. Fractal networks exhibit computational efficiency and implicit union of subnetworks of varying depths. They can extract high-performance fixed-depth subnetworks and demonstrate an "anytime" property, where shallow subnetworks provide quick answers while deeper ones offer higher accuracy. FractalNet's design is simple and effective, with a single loss function driving internal behavior. It is robust to depth choices and can be trained with drop-path, a regularization technique that prevents co-adaptation of subpaths. Drop-path enables the network to maintain anytime behavior, where shallow subnetworks provide quick, moderately accurate results, while deeper subnetworks offer more accurate answers. FractalNet outperforms ResNet in performance on CIFAR-100 and CIFAR-10 without data augmentation, demonstrating that residual learning is not essential for deep networks. It also shows that deep supervision and student-teacher learning can emerge naturally from its structure. FractalNet's architecture is computationally efficient and can be used to extract high-performance subnetworks. Experiments show that FractalNet achieves competitive results with and without data augmentation, and that drop-path regularization is effective in preventing overfitting. FractalNet's design is simple and efficient, with a self-similar structure that may be fundamental to neural architectures. It demonstrates that deep networks can be trained effectively without residual learning, and that self-similarity can be a key component of neural network design.FractalNet introduces a novel design strategy for neural network architecture based on self-similarity, generating an extremely deep network with a truncated fractal structure. Unlike residual networks, which rely on residual learning, FractalNet achieves high performance without residuals, matching the state-of-the-art error rate on CIFAR-100. Fractal networks exhibit computational efficiency and implicit union of subnetworks of varying depths. They can extract high-performance fixed-depth subnetworks and demonstrate an "anytime" property, where shallow subnetworks provide quick answers while deeper ones offer higher accuracy. FractalNet's design is simple and effective, with a single loss function driving internal behavior. It is robust to depth choices and can be trained with drop-path, a regularization technique that prevents co-adaptation of subpaths. Drop-path enables the network to maintain anytime behavior, where shallow subnetworks provide quick, moderately accurate results, while deeper subnetworks offer more accurate answers. FractalNet outperforms ResNet in performance on CIFAR-100 and CIFAR-10 without data augmentation, demonstrating that residual learning is not essential for deep networks. It also shows that deep supervision and student-teacher learning can emerge naturally from its structure. FractalNet's architecture is computationally efficient and can be used to extract high-performance subnetworks. Experiments show that FractalNet achieves competitive results with and without data augmentation, and that drop-path regularization is effective in preventing overfitting. FractalNet's design is simple and efficient, with a self-similar structure that may be fundamental to neural architectures. It demonstrates that deep networks can be trained effectively without residual learning, and that self-similarity can be a key component of neural network design.