This chapter presents various cloud platforms that are available in market offerings from different vendors. IBM provided a machine learning (ML) platform “IBM Watson Studio” (formerly “Data Science Experience”), and this is considered here for the field of study. An overview of artificial intelligence, ML, and deep learning (DL) with their relationship is deliberated. Discussion on popular DL architectures with elementary comparison is also considered.
A lot has changed in the world since the inception of the so-called deep learning (DL) era. Unlike conventional machine learning algorithms, where human learns more of feature extraction than machine, which just crunches numbers, DL algorithms are quite promising as they have revolutionized the way we deal with data and have become adept in taking human-like decisions. Based on the powerful notion of artificial neural networks, these learning algorithms learn to represent data or, if we rephrase, can extract features from raw data. DL has paved way for end-to-end systems where one learning algorithm does all the tasks. Convolutional neural networks have earned a lot of fame lately, especially in the domain of computer vision where in some cases its performance has beaten that of humans. A lot of work has been done on convNets and in this chapter we will demystify how convolutional neural networks work and will illustrate using one novel application in the field of astronomy where we will do galaxy classification using raw images as input and classify them based on its shape. In addition to this we will investigate what features the network is learning. We will also discuss how DL superseded other forms of learning and some recent algorithmic innovations centered on convolutional neural nets.
Optical character recognition (OCR) systems have been used for extraction of text contained in scanned documents or images. This system consists of two steps: character detection and recognition. One classification algorithm is required for character recognition by their features. Character can be recognized using neural networks. The multilayer perceptron (MLP) provides acceptable recognition accuracy for character classification. Moreover, the convolutional neural network (CNN) and the recurrent neural network (RNN) are providing character recognition with high accuracy. MLP, RNN, and CNN may suffer from the large amount of computation in the training phase. MLP solves different types of problems with good accuracy but it takes huge amount of time due to its dense network connection. RNNs are suitable for sequence data, while CNNs are suitable for spatial data. In this chapter, a CNN is implemented for recognition of digits from MNIST database and a comparative study is established between MLP, RNN, and CNN. The CNN provides the higher accuracy for digit recognition and takes lowest amount of time for training the system with respect to MLP and RNN. The CNN gives better result with accuracy up to 98.92% as the MNIST digit dataset is used, which is spatial data.
Deep learning techniques have had a huge impact on artificial intelligence research. They have improved upon the traditional machine learning techniques where human expertise was required for feature engineering. By removing one human factor, they have moved us one step forward in the field of artificial intelligence. They have not entirely removed humans, though. They are required for designing the architectures and cleaning the data. Deep learning techniques have managed to achieve breakthrough results in domains such as speech recognition, machine translation, image recognition, and object detection. This chapter gives a brief overview of various deep learning techniques being used today. Techniques that make deep learning more effective have been described. Some interesting applications have also been covered.
Nowadays, handwritten document analysis using intelligent computing technology is a demanding research area, considering its usefulness in identifying a person and human characteristics, particularly that of persons having typical disabilities such as dyslexia, dysgraphia, and Parkinson’s disease. Analysis of handwriting, falling under the broad purview of graphology, helps us understand the writer’s psychology, emotional outlays, and noticeable disorders as well. Since there prevails a broad spectrum of cursive nature and high inconsistency of handwriting styles, the techniques for modern handwriting analysis need to be more robust and sensitive to different patterns compared to the traditional graphological techniques. Herein lies the necessity of computing technology, which should intelligently analyze handwritten texts to find out the similarity of finer aspects of handwritings of children or adult with some kind of learning/writing disability. Deep learning technology is chosen as the technical tool to identify and classify common features of handwriting of children with developmental dysgraphia. Variational autoencoder, a deep unsupervised learning technique, is presently used for this purpose. This chapter reports successful extraction and interpretation of significant number of distinguishable handwriting characteristics that are clinically proved to be symptoms of dysgraphia.
Audio signal processing and its classification dates back to the past century. From speech recognition to speaker recognition and from speech to text conversion to music generation, a lot of advances has been made in this field using algorithms such as hidden Markov models, recurrent neural networks with long short-term memory layers (LSTM), deep convolutional neural networks (DCNNs), and the recent state-ofthe- art model for music and speech generation using WaveNets. These algorithms are applied after the audio signals are processed and effective feature extraction techniques are applied on them. Nowadays, devices come up with personal assistants with which they can interact either through text inputs or voice inputs. Most applications have also come up with voice search features, while some can generate transcripts from videos and recognize the song title when played. The constant urge for attaining perfection has also led to hybrid models combining supervised and unsupervised learning techniques for better feature extraction. The ability to deal with spectral and temporal data makes DCNNs different from deep neural networks and also makes it the appropriate choice to deal with speech data because correlation between words and phonemes are a characteristic of such data. The potentials of convolution neural networks are huge and being extended in areas like environmental sound classification, music and instrumental sound processing and classification, and large vocabulary continuous speech recognition. Therefore, this chapter gives detailed explanation about what an audio signal is and how it is processed. It will also cover the various feature extraction techniques and the classification algorithms. Finally, the presentday applications and the potentials of deep learning in this field will be explored.
In deep learning, data is transmitted through a number of layers in the feedforward network between input and output layers. In a recurrent network, data may propagate through a layer several times. Backpropagation through time (BPTT) technique is used to train recurrent networks (RNN). The underlying idea of BPTT is to transform a recurrent network into an unfolded feedforward network (multilayer network) where conventional backpropagation learning is used for gradient calculation. Here, each layer of the unfolded network represents a time step. The objective of this study is to integrate concept of BPTT in the framework of fuzzy time series prediction. The model takes sequence of previous values as input (fuzzy inputs) to the different layer of the unfolded network and produces fuzzy output. Temperature dataset is used to evaluate the performance of the model and prediction accuracy of BPTT is better than that of backpropagation neural network model.