Resource Center

The research behind physics-based computing, in full.

  • White Paper

    An Efficient Approach to the Supervised Training of Deep Neural Networks using MemComputing

    Gradient-based methods such as ADAM and SGD are the standard for supervised training, but they lean on a stack of extensions (dropout, momentum, weight decay, variable learning rates) and still leave training closer to art than science. This paper introduces a MemComputing-based training method and benchmarks it against ADAM and SGD on a high-energy particle detection problem, reaching better accuracy in fewer epochs with lower error rates and no additional algorithmic scaffolding.

    Download PDF
  • White Paper

    Efficient Restricted Boltzmann Machine Training for Deep Learning

    Unsupervised pre-training of deep networks still relies on Contrastive Divergence, whose slow-mixing Gibbs sampling is noisy, time-consuming and prone to local minima. Quantum annealing has been proposed as a way out but faces hard implementation limits. This paper applies MemComputing's nonlocal collective-state computation to Restricted Boltzmann Machine training and reports comparative results on MNIST.

    Download PDF
  • White Paper

    Leveraging Physics & Memory: A Paradigm Shift for Efficient Computing

    Every computer from a phone to a supercomputer separates memory from compute. Digital Memcomputing Machines take the opposite view, using memory as a physical property of computation to build self-organizing logic gates that do not distinguish input from output. The paper lays out the framework, its finite-precision realization, and demonstrated performance on Max-SAT, machine learning, subset-sum and integer linear programming, and how it compares with quantum computing.

    Download PDF
  • White Paper

    Confirmation of Exponential Speed-up Using MemComputing to Solve Hard Optimization Problems

    For NP-hard optimization problems, compute time grows exponentially as inputs and constraints grow linearly; a single industrial problem can run for days or weeks. This paper documents a controlled test in which MemComputing's self-organizing logic gates and linear-time scalability deliver a confirmed exponential speed-up over conventional solvers on hard optimization instances.

    Download PDF
  • White Paper

    Integer Linear Programming with MemComputing

    Integer linear programming underpins a vast class of industrial and academic optimization problems, yet the hardest MIPLIB instances still defeat conventional algorithms. Using a new architecture of self-organizing algebraic circuits, MemComputing found feasible solutions within minutes to f2000, a MIPLIB problem whose feasibility had been an open question since 2010.

    Download PDF
  • White Paper

    Solve Optimization Problems with Unparalleled Speed

    As data has exploded, the time to solve constrained optimization problems with classical methods has grown until many are effectively unsolvable, forcing industry onto heuristics that return approximations. This introductory paper explains MemComputing's physics-inspired, non-combinatorial approach and its software coprocessor, and demonstrates it on a hard optimization challenge.

    Download PDF