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

Dynamic adaptation in single-neuron response plays a fundamental role inneural coding in biological neural networks. Yet, most neural activationfunctions used in artificial networks are fixed and mostly considered as aninconsequential architecture choice. In this paper, we investigate nonlinearactivation function adaptation over the large time scale of learning, andoutline its impact on sequential processing in recurrent neural networks. Weintroduce a novel parametric family of nonlinear activation functions, inspiredby input-frequency response curves of biological neurons, which allowsinterpolation between well-known activation functions such as ReLU and sigmoid.Using simple numerical experiments and tools from dynamical systems andinformation theory, we study the role of neural activation features in learningdynamics. We find that activation adaptation provides distinct task-specificsolutions and in some cases, improves both learning speed and performance.Importantly, we find that optimal activation features emerging from ourparametric family are considerably different from typical functions used in theliterature, suggesting that exploiting the gap between these usualconfigurations can help learning. Finally, we outline situations where neuralactivation adaptation alone may help mitigate changes in input statistics in agiven task, suggesting mechanisms for transfer learning optimization.

Advantages of biologically-inspired adaptive neural activation in RNNs during learning