2D Image Reconstruction using Differentiable Plasticity
Shashidhara B. Vyakaranal, Akshata Hiremath, Inzamam Sayyed, Kshitij Ijari, S. M. Meena, Sunil V. Gurlahosur, Uday Kulkarni · 2021
The greatest boon on mankind is the adaptive nature of the brain. Apparently this roots in the exceptional flexibility possessed by the brain while learning. This trait of human learning bids an inspiration to Artificial Intelligence researchers to incorporate the human brain as a new design mechanism. The research opens a new perspective to embrace Artificial Intelligence (AI) with neural networks adding a key to mimic the functioning of the human brain. In contrast, recurrent neural networks (RNNs) and LSTMs can learn from ongoing experiences which will require additional neurons to store previous experiences. Here, the paper describes a simple methodology by taking the inventiveness from biological brains and recreating it as differentiable plasticity to adapt the features into learnings of neural network connections. Thereby, showing that plasticity participating with neural connection along with the Hebbian plastic connection rule can be optimized in recurrent neural networks. On the basis of a plastic neural network, a unique varied solution to image reconstruction tasks is experimented with and also affirmed that the traditionally used recurrent neural networks fall short to resolve these tasks. In this paper, an approach to novel problems to solve reconstructing and memorizing tasks is described on the sets of neoteric colored images by tracing plastic recurrent networks. In addition, this method is evaluated against existing models of RNN's to endorse the results. Hence, concluding that the plasticity method provides a firmer way to solve meta-learning situations in the new era of technology.