Development of adaptive activation functions with curvature and range modulation for abstract image feature learning in complex CNNs for cross-domain applications
Ali Raza, Akhtar Ali, Sami Ullah, Basit Rehman, Zafar Ali · 2026
Read Development of adaptive activation functions with curvature and range modulation for abstract image feature learning in complex CNNs for cross-domain applications by Ali Raza, Akhtar Ali, Sami Ullah, Basit Rehman, Zafar Ali (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/plos-pone-0355613.txt.
Description
Ali Raza and four colleagues redesign a small but crucial component of neural networks in this 2026 paper: the activation function, the curve that decides how strongly each artificial neuron fires. Standard curves are fixed, which limits how well a network can adapt to varied data. Their proposal lets the function's slope, range, and curvature adjust, in three flavors: preset by hand, learned globally by the network, or learned separately by each layer.
Across handwritten digits, fashion images, skin lesion photos, and COVID-19 scans, the adaptive versions generally converged faster and classified better than fixed functions. A compact network using the layer-wise version beat two famous deep architectures on the fashion benchmark and handled imbalanced medical classes more robustly. The authors argue that letting each layer tune its own nonlinearity is a cheap upgrade for image models that must transfer across domains.
Catalog record
| CREATOR | Ali Raza, Akhtar Ali, Sami Ullah, Basit Rehman, Zafar Ali |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 113480 |
| LICENSE | CC_BY |
| RIGHTS | CC BY |
| SOURCE | PLOS |