Memory-Efficient and Trustworthy Neural Networks via Random Seed-Based Design

Deep learning models deployed in resource-constrained and safety-critical settings must be not only compact and accurate but also trustworthy and capable of expressing predictive uncertainty. However, conventional compression techniques such as pruning and quantization often make models overconfident, while uncertainty-aware methods like Bayesian inference and ensembles ignore memory and compute budgets. In this work, we propose Random Seed-Based Design, a framework that elevates the random seed to a first-class design primitive for jointly shaping sparsity, freezing, and stochastic inference. Random seeds deterministically generate all stochastic components: Strong Lottery Ticket weights, partially frozen masks, and Monte Carlo Dropout masks, so that a deployed model can be exactly regenerated from seeds, binary masks, and a small set of interpretable hyperparameters instead of dense weight tensors. We formulate this seed-based construction as a multi-objective design problem over accuracy, calibration, uncertainty, FLOP, and memory footprint, and solve it with Bayesian optimization. This seed-based design enables us to characterize Pareto-optimal trade-offs between efficiency and trustworthiness across diverse architectures, including convolutional neural networks and graph neural networks, thereby illuminating domain-dependent relationships between computational efficiency and predictive reliability. Our method discovers models that are tens to hundreds of times smaller than their dense counterparts, reduces expected calibration error by up to an order of magnitude, and matches or improves accuracy. These results show that structured seed-controlled randomness improves both compression and trustworthiness, rather than treating them as competing objectives. The source code is available at https://github.com/itocha/SLT-TRUST.