We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:


Current browse context:


Change to browse by:

References & Citations

DBLP - CS Bibliography


(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Machine Learning

Title: A Differentiable Transition Between Additive and Multiplicative Neurons

Abstract: Existing approaches to combine both additive and multiplicative neural units either use a fixed assignment of operations or require discrete optimization to determine what function a neuron should perform. However, this leads to an extensive increase in the computational complexity of the training procedure.
We present a novel, parameterizable transfer function based on the mathematical concept of non-integer functional iteration that allows the operation each neuron performs to be smoothly and, most importantly, differentiablely adjusted between addition and multiplication. This allows the decision between addition and multiplication to be integrated into the standard backpropagation training procedure.
Comments: ICLR 2016 extended abstract
Subjects: Machine Learning (cs.LG); Machine Learning (stat.ML)
Cite as: arXiv:1604.03736 [cs.LG]
  (or arXiv:1604.03736v1 [cs.LG] for this version)

Submission history

From: Sebastian Urban [view email]
[v1] Wed, 13 Apr 2016 12:38:27 GMT (2204kb,D)

Link back to: arXiv, form interface, contact.