Let dynamics do part of the computation.
We investigate learning systems shaped by physical principles: memory, dissipation, geometry, and dynamical evolution. Reservoir computing is a central direction, alongside physics-informed modelling and the interpretation of time-varying signals.
Temporal data carries history. Reservoir computing uses a recurrent dynamical system to transform that history into a state from which a readout can learn. We explore how deliberately structured dynamics can make memory, mixing, and stability easier to understand and control.
Questions guiding this vertical and its contribution to intelligence for inner wellbeing.
Investigate architectures with explicit controls for memory retention and mixing. Lindblad-inspired work motivates separating rotation from dissipation in classical models rather than hiding both inside one recurrent operator.
Explore how evolving geometry can encode the order and history of inputs. Chern–Simons context-reservoir work motivates testing which tasks benefit from these structures against matched, simpler controls.
Study how physical priors and interpretable state dynamics can support signal modelling. Compare candidate approaches on memory, stability, generalisation, and computational cost before considering deployment.
Map a time-varying input into a dynamical state.
Let structured recurrence mix input with retained history.
Train an output mapping for the task of interest.
Benchmark against simpler reservoirs and other temporal models.
Selected work by our scientific lead and collaborators. Each paper is credited to its authors; publication status is shown alongside the source.
A classical reservoir architecture inspired by open-system dynamics, with separate controls for rotational mixing and dissipative forgetting. Its design makes memory timescales and stability explicit.
Research implication: explore interpretable temporal models that balance recent changes with longer histories.
Read preprintA study of memory in a dynamical reservoir organised by Chern–Simons gauge dynamics. Matched comparisons find task-specific benefits for geometry- and order-sensitive processing, rather than a general advantage across reservoir tasks.
Research implication: investigate which dynamical features retain useful history, and compare them against simpler mechanisms.
Read preprintThese works inform the research agenda. They do not constitute validation of SakshiSense hardware or wellness outcomes.
Wearable signals unfold across multiple timescales. We are interested in whether structured temporal models can help represent these histories for personalised wellness research. This is a development direction to evaluate, not a claim that either preprint’s architecture is already deployed in SakshiSense.
Explore the wellness platform ↗Join us on reservoir architectures, temporal benchmarks, signal modelling, or physics-informed computation.
