dc.contributor.author |
Chauhan, Arushi |
|
dc.contributor.author |
Vatsa, Mayank (Advisor) |
|
dc.contributor.author |
Singh, Richa (Advisor) |
|
dc.date.accessioned |
2021-05-21T09:39:50Z |
|
dc.date.available |
2021-05-21T09:39:50Z |
|
dc.date.issued |
2020-06-02 |
|
dc.identifier.uri |
http://repository.iiitd.edu.in/xmlui/handle/123456789/893 |
|
dc.description.abstract |
Convolutional neural networks are being extensively used in real world applications like image
and video classification, natural language processing, medical image analysis, recommender systems etc. where previously machine learning algorithms and hand-crafted approaches were used.
Some of the well known CNNs are LeNet5, AlexNet, VGG, GoogleNet, ResNet, DenseNet etc.
Over the years, the architecture of CNNs is successively becoming deeper (using more layers)
and more complex with the introduction of new types of layers. These CNNs are designed by experts who have rich domain knowledge of both datasets and CNNs. There is a great demand for
algorithms which can build CNNs with the best architecture that can work for a given dataset.
This would reduce the dependence on researchers to use hand-crafted networks for every new
task. The algorithms should also be able to work with small datasets and use efficient computation techniques to use GPU effectively. This project explores the various techniques used
in neural architecture search, and works on developing models to predict the performance of
the CNNs. We propose the Network Epoch Accuracy Prediction Framework (NEAP-F) which
predicts the accuracy achieved (in an epoch) of a sample network on an image dataset, and a
dataset of network architecture training curves on image datasets CIFAR-10 and MNIST. In the
current scenario where fast and efficient computing is the norm, NEAP-F reduces the resources
required in neural architecture search systems by eliminating the need to train architectures,
which is major bottleneck, thus, cutting down computation time and resources needed to examine architectures heavily; current computation time of NEAP-F is in the order of miliseconds.
The dataset released augments existing datasets with networks having residual connections to
reflect the state-of-the-art architectures. The results have been computed on dataset consisting
of 50670 data points for prediction, distributed across 3 image datasets - CIFAR-10, SVHN and
MNIST. |
en_US |
dc.language.iso |
other |
en_US |
dc.publisher |
IIIT-Delhi |
en_US |
dc.subject |
convolutional neural networks, datasets, genetic algorithm, memetic algorithm, cross entropy loss, vector representation, performance prediction, ResNet, architecture search, sequence-to-sequence model, dataset, description, regression models, epoch prediction, image datasets, architecture vector representation, dataset vector representation, ”ease of classifying” dataset |
en_US |
dc.title |
Predicting deep learning architecture |
en_US |
dc.type |
Other |
en_US |