Implementation and Math¶

Complex convolutional networks provide the benefit of explicitly modelling the phase space of physical systems [trabelsi2017deep]. The complex convolution introduced can be explicitly implemented as convolutions of the real and complex components of both kernels and the data. A complex-valued data matrix in cartesian notation is defined as $$\textbf{M} = M_\Re + i M_\Im$$ and equally, the complex-valued convolutional kernel is defined as $$\textbf{K} = K_\Re + i K_\Im$$. The individual coefficients $$(M_\Re, M_\Im, K_\Re, K_\Im)$$ are real-valued matrices, considering vectors are special cases of matrices with one of two dimensions being one.

Complex Convolution Math¶

The math for complex convolutional networks is similar to real-valued convolutions, with real-valued convolutions being:

$\int f(y)\cdot g(x-y) \, dy$

which generalizes to complex-valued function on $$\mathbf{R}^d$$:

$(f * g )(x) = \int_{\mathbf{R}^d} f(y)g(x-y)\,dy = \int_{\mathbf{R}^d} f(x-y)g(y)\,dy,$

in order for the integral to exist, f and g need to decay sufficiently rapidly at infinity [CC-BY-SA Wiki].

Implementation¶

Solving the convolution of, implemented by [trabelsi2017deep], translated to keras in [dramsch2019complexsoftware]

Complex Convolution implementation (CC-BY [trabelsi2017deep])

$M' = K * M = (M_\Re + i M_\Im) * (K_\Re + i K_\Im),$

we can apply the distributivity of convolutions to obtain

$M' = \{M_\Re * K_\Re - M_\Im * K_\Im\} + i \{ M_\Re * K_\Im + M_\Im * K_\Re\},$

where K is the Kernel and M is a data vector.

Considerations¶

Complex convolutional neural networks learn by back-propagation. [Sarroff2015] state that the activation functions, as well as the loss function must be complex differentiable (holomorphic). [trabelsi2017deep] suggest that employing complex losses and activation functions is valid for speed, however, refers that [Hirose2012] show that complex-valued networks can be optimized individually with real-valued loss functions and contain piecewise real-valued activations. We reimplement the code [trabelsi2017deep] provides in keras with tensorflow , which provides convenience functions implementing a multitude of real-valued loss functions and activations.

[CC-BY [dramsch2019complex]]

 [DC19] Jesper Soeren Dramsch and Contributors. Complex-valued neural networks in keras with tensorflow. 2019. URL: https://figshare.com/articles/Complex-Valued_Neural_Networks_in_Keras_with_Tensorflow/9783773/1, doi:10.6084/m9.figshare.9783773.
 [DLuthjeC19] Jesper Sören Dramsch, Mikael Lüthje, and Anders Nymark Christensen. Complex-valued neural networks for machine learning on non-stationary physical data. arXiv preprint arXiv:1905.12321, 2019.
 [HY12] Akira Hirose and Shotaro Yoshida. Generalization characteristics of complex-valued feedforward neural networks in relation to signal coherence. IEEE Transactions on Neural Networks and Learning Systems, 2012.
 [SSC15] Andy M. Sarroff, Victor Shepardson, and Michael A. Casey. Learning representations using complex-valued nets. CoRR, 2015. URL: http://arxiv.org/abs/1511.06351, arXiv:1511.06351.
 [TBZ+17] (1, 2, 3, 4, 5) Chiheb Trabelsi, Olexa Bilaniuk, Ying Zhang, Dmitriy Serdyuk, Sandeep Subramanian, João Felipe Santos, Soroush Mehri, Negar Rostamzadeh, Yoshua Bengio, and Christopher J Pal. Deep complex networks. arXiv preprint arXiv:1705.09792, 2017.