Abstract
Construction of a neural network is the cardinal step to any machine learning algorithm. It requires profound knowledge for the developer in assigning the weights and biases to construct it. And the construction should be done for multiple epochs to obtain an optimal neural network. This makes it cumbersome for an inexperienced machine learning aspirant to develop it with ease. So, an automated neural network construction would be of great use and provide the developer with incredible speed to program and run the machine learning algorithm. This is a crucial assist from the developer's perspective. The developer can now focus only on the logical portion of the algorithm and hence increase productivity. The use of Enas algorithm aids in performing the automated transfer learning to construct the complete neural network from the given sample data. This algorithm proliferates on the incoming data. Hence, it is very important to inculcate it with the existing machine learning algorithms.
TopBackground Study
Neural Networks and Domain Specific Language
Neural networks a set of logic and knowledge based approach sequenced, modelled based on the brain of human (Ahamed and Yuvaraj, 2018). They recognize by interpreting patterns, sensory data through machine perception. Commercial applications technologies generally focused on solving complex pattern or signal processing problems such as speech, handwriting, oil exploration data analysis, facial recognition and weather predictions. A Domain Specific Language (DSL), computer programing restricted expressiveness focused on a specific domain. DSLs provides significant gain in Application productivity, creativity developers, portability and performance. DSLs offer pre-defined abstractions represent concepts from application domain. A programmer uses one or more of the DSLs write the programs using specific domain constructs and notations. In additional ability to use domain knowledge to apply static & dynamic optimizations to a program (Cessac et. al, 2016). DSL specifically targeted, machine learning, abstractions to define neural networks, capture high level information used to increased productivity, performance and expose parallelism. It moves the programmers approach from a low level detailing that are not the focus of the system in development, lets them target on work the solution for the problem at hand.
Key Terms in this Chapter
Neural Architecture Search (NAS): The method of searching for the optimal architecture in the neural network search space.
Net2Net: System that allows the user to transfer knowledge to wider and deeper networks with methods called Net2WiderNet and Net2DeeperNet, respectively.
Domain-Specific Language (DSL): A computer programing-restricted expressiveness focused on a specific domain. DSL specifically targeted, machine learning, abstractions to define neural networks, capture high level information used to increased productivity, performance and expose parallelism.
Meta-QNN: The method for meta modelling of neural networks using reinforcement techniques.
Structural Learning: Networks that learn the appropriate network architecture for the learning task instead of having a fixed architecture design and parameters.
Deep Learning (DL): A part of machine learning with its algorithms, to the structure and working of the brain called artificial neural networks; a knowledge process and a way to automate Predictive Analytics.
ProxylessNAS: The method of modelling the architecture search space as a neural network, with each choice being represented as a path in the network.
Deep Architect: A domain-specific language that allows the user to define the search space for NAS. The DSL allows any choice based NAS algorithm to be implemented in a modular and extensible fashion.