Convolutional Neural Generative Coding: Scaling Predictive Coding to Natural Images
This work addresses the challenge of applying predictive coding to complex image data for researchers in computational neuroscience and machine learning, offering an incremental improvement by generalizing existing methods to convolutional architectures.
The authors tackled the problem of scaling predictive coding to natural images by developing convolutional neural generative coding (Conv-NGC), a brain-inspired algorithm that refines latent feature maps for reconstruction and denoising. The result showed competitive performance with convolutional auto-encoders on tasks like reconstruction and outperformed them in out-of-distribution reconstruction on datasets such as CIFAR-10 and SVHN, including tests on the full 90k CINIC-10 set.
In this work, we develop convolutional neural generative coding (Conv-NGC), a generalization of predictive coding to the case of convolution/deconvolution-based computation. Specifically, we concretely implement a flexible neurobiologically-motivated algorithm that progressively refines latent state feature maps in order to dynamically form a more accurate internal representation/reconstruction model of natural images. The performance of the resulting sensory processing system is evaluated on complex datasets such as Color-MNIST, CIFAR-10, and Street House View Numbers (SVHN). We study the effectiveness of our brain-inspired model on the tasks of reconstruction and image denoising and find that it is competitive with convolutional auto-encoding systems trained by backpropagation of errors and outperforms them with respect to out-of-distribution reconstruction (including the full 90k CINIC-10 test set).