Dynamic Learning Networks PDF Download

"In a continual learning system, the network has to dynamically learn new tasks from few samples throughout its lifetime. It is observed that neuromodulation acts as a key factor in continual and dynamic learning in the central nervous system. In this work, the neuromodulatory plasticity is embedded with dynamic learning architectures. The network has an inbuilt modulatory unit that regulates learning depending on the context and the internal state of the system, thus rendering the networks with the ability to self modify their weights. In one of the proposed architectures, ModNet, a modulatory layer is introduced in a random projection framework. This layer modulates the weights of the output layer neurons in tandem with hebbian learning. Moreover, to explore modulatory mechanisms in conjunction with backpropagation in deeper networks, a modulatory trace learning rule is introduced. The proposed learning rule, uses a time dependent trace to automatically modify the synaptic connections as a function of ongoing states and activations. The trace itself is updated via simple plasticity rules thus reducing the demand on resources. A digital architecture is proposed for ModNet, with on-device learning and resource sharing, to facilitate the efficacy of dynamic learning on the edge. The proposed modulatory learning architecture and learning rules demonstrate the ability to learn from few samples, train quickly, and perform one shot image classification in a computationally efficient manner. The ModNet architecture achieves an accuracy of ∼91% for image classification on the MNIST dataset while training for just 2 epochs. The deeper network with modulatory trace achieves an average accuracy of 98.8%±1.16 on the omniglot dataset for five-way one-shot image classification task. In general, incorporating neuromodulation in deep neural networks shows promise for energy and resource efficient lifelong learning systems."--Abstract.