With everyday applications in language, voice, image, and automatic driving cars, deep learning has become a popular concept to the general public in the past few years. However, many of the concepts of deep learning started as early as the 1940s. For example, the binary perceptron classifier, invented in the late 1950s, uses a linear combination of input signals and a step activation function. This is the same as a single neuron in a modern deep learning network that uses the same linear combination of input signals from neurons at the previous layer and a more efficient nonlinear activation function. The perceptron model was further defined by minimizing the classification error and trained by using one data point at a time to update the model parameters during the optimization process. Modern neural networks are trained similarly by minimizing a loss function but with more modern optimization algorithms such as stochastic gradient descent and its variations.
Even though the theoretical foundation of deep learning has been continually developed in the past few decades, real-world applications of deep learning are fairly recent due to some real world constraints: data, network structure, algorithm, and computation power.
We are all familiar with all sorts of data today: structured tabulated data in database tables or CSV files, free form text, images, and other unstructured datasets. However, historical datasets are relatively small in size, especially for data with accurately labeled ground truth. Statisticians have been working on datasets that only have a few thousand rows and a few dozen columns for decades to solve business problems. Even with modern computers, the size of the data is usually limited to the memory of a computer. Now we know, to enable deep learning applications, the algorithm needs a much larger dataset than traditional machine learning methods. It is usually at the order of millions of samples with high-quality ground truth labels for supervised deep learning models.
The first widely used large dataset with accurate labels was the ImageNet dataset which was first created in 2010. It now contains more than 14 million images with more than 20k synsets (i.e. meaningful categories). Every image in the dataset was human-annotated with quality control to ensure the ground truth labels are accurate. One of the direct results of ImageNet was the Large Scale Visual Recognition Challenge (ILSVRC) which evaluated different algorithms in image-related tasks. The ILSVRC competition provided a perfect stage for deep learning applications to debut to the general public. For 2010 and 2011, the best record of error from traditional image classifications methods was around 26%. In 2012, a method based on the convolutional neural network became the state of the art with an error rate of around 16%, a dramatic improvement from the traditional methods.
With the prevalence of the modern internet, the amount of text, voice, image, and video data increased exponentially. The quality and quantity of data for deep learning applications enabled the use of deep learning for applications such as image classification, voice recognition, and natural language understanding. Data is the fuel for deep learning engines. With more and more varieties of data created, captured, and saved, there will be more applications of deep learning discovered every day.
Lacking high-quality high-volume data was not the only constraint for early deep learning years. For perceptron with just one single neuron, it is just a linear classifier. Real applications are nearly always non-linear. To solve this problem, we have to grow one neuron to multiple layers with multiple neurons per layer. This multi layer perceptron (MLP) is also referred to as a feedforward neural network. In the 1990s, the universal approximation theorem was proven and it assured us that a feedforward network with a single hidden layer containing a finite number of neurons can approximate continuous functions. Even though the one-layer neural network theoretically can solve a general non-linear problem, the reality is that we have grown the neural network to many layers of neurons. The number of layers in the network is the “depth” of a network. Loosely speaking, deep learning is a neural network with the many layers (i.e. the depth is deep).
The MLP is the basic structure for the modern deep learning applications. MLP can be used for classification problems or regression problems with response variables as the output and a collection of explanatory variables as the input (i.e. the traditionally structured datasets). Many of the problems that can be solved using classical classification methods such as random forest can be solved by MLP. However, MLP is not the best option for image and language-related tasks. For image-related tasks, pixels for a local neighbor region collectively provide useful information to solve a task. To take advantage of the 2D spatial relationship among pixels, the convolutional neural network (CNN) structure is a better choice. For language-related tasks, the sequence of the text provides additional information than just a collection of single words. The recurrent neural network (RNN) is a better structure for such sequence-related data. There are other more complicated neural network structures and it is still a fast-developing area. MLP, CNN, and RNN are just the starting point of deep learning methods.
In addition to data and neural network structure, there were a few key algorithm breakthroughs to enable the widespread adoption of deep learning. For an entry-level neural network structure, there are hundreds of thousands of parameters to be estimated from the data. With a large amount of training data, stochastic gradience decent and mini-batch gradience decent are efficient ways to utilize a subset of training data to update the model parameters. Within the process, one of the key steps is back-propagation which was introduced in the 1980s for efficient weight update. There is a non-linear activation for each neuron in deep learning models, and sigmoid or hyperbolic tangent functions were often used. However, it has the problem of gradient vanishing when the number of layers of the network grows large (i.e. deeper network). To solve this problem, the rectified linear unit (ReLu) was introduced to deep learning in the 2000s and it increases the convergence speed dramatically. ReLu is so simple (i.e. y = x when x >= 0 and y = 0 otherwise), but it indeed cleverly solved one of the big headaches in deep learning. We will talk more about activation functions in section 12.1.4.
With hundreds of thousands of parameters in the model, deep learning is easy to overfit. In order to mitigate this, dropout, a form of regularization, was introduced in 2012. It randomly drops out a certain percentage of neurons in the network during the optimization process to achieve more robust model performance. It is similar to the concept of random forest where features and training data are randomly chosen. There are many other algorithm improvements to get better models such as batch normalization and using residuals from previous layers. With backpropagation in stochastic gradience decent, ReLu activation function, dropout, and other techniques, modern deep learning methods begin to outperform traditional machine learning methods.
With data, network structure and algorithms ready, modern deep learning still requires a certain amount of computation power for training. The entire framework involves heavy linear algebra operations with large matrices and tensors. These types of operations are much faster on modern graphical processing units (GPUs) than the computer’s central processing units (CPU).
With the vast potential application of deep learning, major tech companies contribute heavily to open-source deep learning frameworks. For example, Google has open-sourced its TensorFlow framework; Facebook has open-sourced its PyTorch framework, and Amazon has significantly contributed to the MXNet open-source framework. With thousands of software developers and scientists behind these deep learning frameworks, users can confidently pick one framework and start training their deep learning models right away in popular cloud environments. Much of the heavy lifting to train a deep learning model has been embedded in these open-source frameworks and there are also many pre-trained models available for users to adopt. Users can now enjoy the relatively easy access to software and hardware to develop their own deep learning applications. In this book, we will demonstrate deep learning examples using Keras, a high-level abstraction of TensorFlow, using the Databricks Community Edition platform.
In summary, deep learning has not just developed in the past few years but have in fact been ongoing research for the past few decades. The accumulation of data, the advancement of new optimization algorithms and the improvement of computation power has finally enabled every day deep learning applications. In the foreseeable future, deep learning will continue to revolutionize machine learning methods across many more areas.