Gallery¶

Interactive, data-backed visualisations of the project. Every chart below is a live Plotly figure (pan, zoom, hover) — the result charts are generated directly from the results database by experiments/build_plotly_gallery.py, and the reservoir visualisations by esnfed.viz.

Results from the thesis¶

Echo State Network hyper-parameters¶

Larger reservoirs and a spectral radius near 1 minimise error on NARMA-10.

Reservoir topology comparison¶

Topology matters in a task-dependent way (note the log scale).

Federated vs local as the federation grows¶

Exact federated ridge stays flat; local-only degrades with more, smaller clients.

FedAvg convergence vs exact ridge¶

Iterative FedAvg converges slowly toward the one-shot closed-form solution.

Prediction ensemble (heterogeneous reservoirs)¶

The ensemble beats local-only and approaches the centralized model.

Structural alignment¶

Parameter aggregation only becomes competitive as reservoirs are homogenised.

Federated counterparty-risk forecasting¶

On real, non-stationary financial data local-only training collapses (log scale).

FedResPrompt vs federated LoRA¶

Orders-of-magnitude communication and edge-compute savings across model scales.

Acceleration backends¶

Harvest throughput for the optional Numba / float32 / sparse accelerators.

Benchmark classification accuracy¶

Federated = centralized; ensemble in between; local-only far behind.

Reservoir heterogeneity extensions¶

Depth and heterogeneous leaking rates help long-memory Mackey-Glass; multi-type nonlinearities help the faster Lorenz task (see Advanced reservoirs).

Reservoir visualisations¶

Animations¶

The spectral radius and the echo state property:

A reservoir's fading response to an input impulse: