English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Networks in machine learning offer examples of complex high-dimensionaldynamical systems reminiscent of biological systems. Here, we study thelearning dynamics of Generalized Hopfield networks, which permit avisualization of internal memories. These networks have been shown to proceedthrough a 'feature-to-prototype' transition, as the strength of networknonlinearity is increased, wherein the learned, or terminal, states of internalmemories transition from mixed to pure states. Focusing on the prototypelearning dynamics of the internal memories we observe a strong resemblance tothe canalized, or low-dimensional, dynamics of cells as they differentiatewithin a Waddingtonian landscape. Dynamically, we demonstrate that learning ina Generalized Hopfield Network proceeds through sequential 'splits' in memoryspace. Furthermore, order of splitting is interpretable and reproducible. Thedynamics between the splits are canalized in the Waddington sense -- robust tovariations in detailed aspects of the system. In attempting to make the analogya rigorous equivalence, we study smaller subsystems that exhibit similarproperties to the full system. We combine analytical calculations withnumerical simulations to study the dynamical emergence of thefeature-to-prototype transition, and the behaviour of splits in the landscape,saddles points, visited during learning. We exhibit regimes where saddlesappear and disappear through saddle-node bifurcations, qualitatively changingthe distribution of learned memories as the strength of the nonlinearity isvaried -- allowing us to systematically investigate the mechanisms thatunderlie the emergence of Waddingtonian dynamics. Memories can thusdifferentiate in a predictive and controlled way, revealing new bridges betweenexperimental biology, dynamical systems theory, and machine learning.

A Waddington landscape for prototype learning in generalized Hopfield networks