Unconventional Computing

How can physical, chemical and biological dynamics implement specified computations? Covers computing with living, chemical and bodily dynamics, including how inputs are encoded and results read out.

What constitutes a computation

  • Computation and dynamical regime - Langton studies information storage, transmission and modification near transitions in cellular-automaton behavior. The proposed link between computation and criticality is a hypothesis drawn from the studied rule families. Langton 1990.
  • Physical computation criteria - Relates a physical system to an abstract computation through encoding, evolution and decoding. The proposed framework distinguishes performing a computation from observing dynamics that can be described mathematically. Horsman et al. 2014.

Molecular computation

  • DNA computation - DNA strands encode candidate paths through a seven-node graph. Amplification and successive laboratory filters isolate a Hamiltonian path, visiting each node once. Adleman 1994.
  • DNA associative memory - DNA strand displacement implements a four-neuron Hopfield network that recalls stored patterns from incomplete inputs. Weights are trained in software, then encoded as molecular concentrations. Qian et al. 2011.

Physical and biological computations

  • Pask’s electrochemical ear - Cariani documents electrochemical devices whose growing conductive structures acquired sensitivity to sound or magnetic fields. A historical account of changing the sensor itself, not only adjusting a controller. Cariani 1993.
  • Slime-mould maze solving - Physarum initially fills a small maze, then retracts into a short connection between two food sources. The assay demonstrates distributed path selection through growth and retraction. Nakagaki et al. 2000; video.
  • Physarum transport networks - Slime mould connects food sites representing Tokyo-area cities. Its networks balance cost, transport efficiency and fault tolerance comparably to the railway; a flow-dependent tube model captures adaptation. Tero et al. 2010.
  • Soft-body computation - A driven silicone arm provides nonlinear dynamics and memory for trained readouts. It emulates temporal systems better than the tested direct linear readout, but worse than an optimized software reservoir. Nakajima et al. 2015.
  • Iontronic memory - Salt redistribution gives aqueous channels fading conductance memory. Tests at three channel lengths support retention time proportional to length squared. Software classifiers use measured pulse responses to recognize digits. Kamsma et al. 2024.

Trainable physical systems

  • Physics-aware training - Physical forward passes and gradients from a differentiable digital model train optical, electronic and mechanical systems for classification. Measuring hardware during training reduces errors from simulation mismatch. Wright et al. 2022.
  • Contrastive local learning - Twin transistor networks learn XOR and nonlinear regression by locally comparing free and target-clamped voltages. Circuit dynamics update parameters without a digital model; training inputs and targets are externally supplied. Dillavou et al. 2024.
  • Physical-network training - Training can optimize a digital surrogate, combine measured hardware outputs with digital gradients, or update parameters through physical dynamics. Model accuracy, measurement access and drift constrain the approaches a substrate supports. Momeni et al. 2025.

Open questions

  • How much of a demonstrated computation is performed by the substrate versus preprocessing, readout or external feedback?
  • Does performance persist after accounting for preparation time, calibration, energy, error and repeated use?