Energy-based learning optimises parameters to reduce the free energy of data; dense interactions allow richer energy surfaces and better modelling of complex distributions, aligning with classic energy-based model (EBM) principles. 

Our architecture can emulate fully connected Boltzmann machines at scale, which enhances expressive power and sampling efficiency compared to sparse or physically limited topologies in conventional implementations. 

Our approach replaces large physical synapse networks or hidden chains with a digital energy calculator, reducing interconnect overheads and routing complexity, a major bottleneck in dense EBMs, thereby improving hardware utilisation and enabling larger instances under fixed physical constraints.