Page 215 - Machine Learning for Subsurface Characterization

Deep neural network architectures Chapter 7 185

1.2 Deep learning
Deep learning is a specialized form of machine learning that uses several
processing layers to learn complex abstractions (concepts) in data by building a
hierarchy/levels of abstractions, wherein each level of abstraction is created
using the lower-level abstractions learnt by the preceding layer in the
hierarchy. In other words, a model developed using deep learning techniques
learns complicated concepts using simpler ones. There are many computational
layers between the input and output resulting in multiple linear and nonlinear
transformations at each layer. Deep learning uses multiple layers of
hierarchical, sequential, and/or recurring computational units to extract features
from raw data at multiple levels, such that the collection of extracted features
at a certain level forms a specific level of abstraction. In simple terms, a deep
learning model is a chain of simple, continuous transformations that map the
input vector space into output vector space by learning from a dense sampling
of available inputs and outputs (referred as training data). Learning each
suitable transformation in the chain of transformations requires computation of
certain parameters, which are iteratively updated using various optimization
algorithms based on the model performance.
Unlike traditional machine learning, deep learning does not require manual
feature engineering prior to the model development because of their capability
to perform hierarchical feature learning, where higher-level features are defined
in terms of lower-level features. Learning of features at multiple levels of
abstraction allows deep learning methods to learn complex functions that
map the input to the output directly from data, without depending on human-
engineered features. Consequently, deep learning is popular method when
dealing with unstructured data, such as images, video, audio, speech, text,
language, analog data, health records, metadata and game play. Deep
learning model is randomly initiated and then generally gradient-based
optimization is used to converge the model parameters (weights and biases)
to an optimal solution, which might not necessarily be the global optimum.
This optimization process for multiple layers containing multiple
computational units requires a lot of data.
Deep learning came into prominence due to the following reasons:
1. Availability of large data volumes being generated at a fast rate by multiple
sources.
2. Access to computational infrastructure and advanced hardware for data
processing and model training.
3. Easily accessible, robust optimization and learning platforms.
4. Opportunity for large-scale training and deployment of the data-driven
models.
5. Need for real-time decision-making and precision engineering/operations.