Abstract
Deep learning has achieved state-of-the-art accuracies on several computer vision tasks. However, the computational and energy requirements associated with training such deep neural networks can be quite high. In this paper, we propose a cumulative training strategy with Net2Net transformation that achieves training computational efficiency without incurring large accuracy loss, in comparison to a model trained from scratch. We achieve this by first training a small network (with lesser parameters) on a small subset of the original dataset, and then gradually expanding the network using Net2Net transformation to train incrementally on larger subsets of the dataset. This incremental training strategy with Net2Net utilizes function-preserving transformations that transfers knowledge from each previous small network to the next larger network, thereby, reducing the overall training complexity. Our experiments demonstrate that compared with training from scratch, cumulative training yields ~2x reduction in computational complexity for training TinyImageNet using VGG19 at iso-accuracy. Besides training efficiency, a key advantage of our cumulative training strategy is that we can perform pruning during Net2Net expansion to obtain a final network with optimal configuration (~0.4x lower inference compute complexity) compared to conventional training from scratch. We also demonstrate that the final network obtained from cumulative training yields better generalization performance and noise robustness. Further, we show that mutual inference from all the networks created with cumulative Net2Net expansion enables improved adversarial input detection.
Introduction
Training state-of-the-art deep neural networks (DNNs) is an expensive endeavor. Standard practice involves training a large, fixed-size model on a massive dataset from scratch. This "brute force" approach is computationally intensive, consuming vast amounts of energy and memory because every iteration requires forward and backward passes through the entire model using the full dataset. This research identifies a missed opportunity: the full capacity of a massive model is often unnecessary to learn initial features. However, simply training a small model and then restarting with a larger one wastes prior computational effort. The challenge lies in growing a model's size and its data exposure incrementally without losing the progress made in earlier stages.
Methodology: Cumulative Training with Net2Net
The proposed strategy, Cumulative Training with Net2Net Transformation, co-evolves the network architecture and the dataset size. The process begins by training a small "base" network (e.g., VGG11) on a small subset of the data (e.g., the first 50 classes of CIFAR100). Once this model converges, it is expanded into a larger architecture (e.g., VGG16 or VGG19) using Net2Net transformations.
Net2Net utilizes function-preserving transformations—such as initializing new layers as identity mappings—to ensure the new, larger model inherits the exact performance of its predecessor. This provides a "warm start," eliminating the initial period of high gradient volatility seen in random initialization. Following expansion, more data is introduced (e.g., adding the next 20 classes), and training continues. The authors also incorporate L1-norm kernel pruning during expansion steps to "trim the fat" from learned layers, resulting in a final model that is efficient for both training and inference.
Experimental Results
The approach was evaluated on CIFAR-10, CIFAR-100, and TinyImageNet. A key finding is the significant reduction in training complexity. The cumulative approach achieved a 0.5x to 0.8x reduction in total training complexity (measured in MACs/OPS) compared to training from scratch, while maintaining nearly identical final accuracy.
Beyond efficiency, the strategy yielded unexpected benefits in robustness. Cumulatively trained models showed improved resilience to Gaussian noise and random neuron ablation, suggesting the incremental process leads to more stable, generalizable minima in the loss landscape. Furthermore, by employing "Mutual Inference" (MI)—using the intermediate models in an ensemble—the system achieved a higher True Negative Rate (TNR) for detecting adversarial attacks like FGSM, as the different model versions possess distinct failure modes.
Conclusion and Future Work
Cumulative training with Net2Net offers a compelling path toward more sustainable AI by reducing the "brute force" requirements of standard training. While it introduces some pipeline complexity and memory overhead for Mutual Inference, the gains in training efficiency (~2x for TinyImageNet) and robustness are substantial. Future research should explore how these strategies translate to modern Transformer-based architectures, which currently dominate the field but remain exceptionally resource-hungry.
Summary updated