Towards Sustainable Models via Binary Neural Networks and Quantum Computing

The field of Artificial Intelligence (AI) is evolving at an unprecedented pace, but concerns regarding energy consumption have been raised due to the increasing computational cost of training and deploying deep learning models. Binary Neural Networks (BiNNs) - neural networks with single-bit precision - have emerged as a promising solution, reducing memory and power requirements while maintaining competitive performance. However, their optimization remains challenging due to the limitations of traditional training algorithms. In this work, we explore the potential of Quantum HyperNetworks to enhance BiNN optimization by leveraging quantum computing. Specifically, Variational Quantum Algorithms (VQAs) facilitate the generation of binary weights through quantum circuit measurements, while quantum principles such as superposition and entanglement enable exploration of a broader solution space. Moreover, we derive the Evidence Lower Bound (ELBO) for quantum state distributions and introduce a Surrogate version of the ELBO for scenarios where only implicit quantum distributions are accessible, as is the case of real quantum hardware. Additionally, we explore scalability techniques, including Tensor Network (TN) representations and Multi Basis Encoding (MBE), to efficiently scale quantum models to larger systems. Beyond theoretical insights, we provide a detailed implementation of the proposed approach within a Python environment, explaining the methods used to simulate quantum algorithms on classical hardware. Finally, we analyze the energy consumption associated with running VQAs on classical simulators and project the potential energy savings when these algorithms are executed on future Quantum Processing Units (QPUs). This comparative analysis offers valuable insights into the potential energy savings and computational benefits that quantum technologies could bring to AI systems.