Research also opened a world of new applications. Neural networks can learn from huge amounts of data, and because they’re more prone to high variance than to bias, they can take advantage of big data, creating models that continuously perform better, depending on the amounts of data you feed them. However, you need large, complex networks for certain applications (to learn complex features, such as the characteristics of a series of images) and thus incur problems like the vanishing gradient.
In fact, when training a large network, the error redistributes among the neurons favoring the layers nearest to the output layer. Layers that are further away receive smaller errors, sometimes too small, making training slow if not impossible. Thanks to the studies of scholars such as Geoffrey Hinton, new turnarounds help avoid the problem of the vanishing gradient. The result definitely helps a larger network, but deep learning isn’t just about neural networks with more layers and units.
In addition, something inherently qualitative changed in deep learning as compared to shallow neural networks, shifting the paradigm in machine learning from feature creation (features that make learning easier) to feature learning (complex features automatically created on the basis of the actual features).
Big players such as Google, Facebook, Microsoft, and IBM spotted the new trend and since 2012 have started acquiring companies and hiring experts (Hinton now works with Google; LeCun leads Facebook AI research) in the new fields of deep learning. The Google Brain project, run by Andrew Ng and Jeff Dean, put together 16,000 computers to calculate a deep learning network with more than a billion weights, thus enabling unsupervised learning from YouTube videos.
There is a reason why the quality of deep learning is different. Of course, part of the difference is the increased usage of GPUs. Together with parallelism (more computers put in clusters and operating in parallel), GPUs allow you to successfully apply pretraining, new activation functions, convolutional networks, and drop-out, a special kind of regularization different from L1 and L2. In fact, it has been estimated that a GPU can perform certain operations 70 times faster than any CPU, allowing a cut in training times for neural networks from weeks to days or even hours.
Both pretraining and new activation functions help solve the problem of the vanishing gradient. New activation functions offer better derivative functions, and pretraining helps start a neural network with better initial weights that require just a few adjustments in the latter parts of the network.Advanced pretraining techniques such as Restricted Boltzanman Machines, Autoencoders, and Deep Belief Networks elaborate data in an unsupervised fashion by establishing initial weights that don’t change much during the training phase of a deep learning network. Moreover, they can produce better features representing the data and thus achieve better predictions.
Given the high reliance on neural networks for image recognition tasks, deep learning has achieved great momentum thanks to a certain type of neural network, the convolutional neural networks. Discovered in the 1980s, such networks now bring about astonishing results because of the many deep learning additions.
To understand the idea behind convolutional neural networks, think about the convolutions as filters that, when applied to a matrix, transform certain parts of the matrix, make other parts disappear, and make other parts stand out. You can use convolution filters for borders or for specific shapes. Such filters are also useful for finding details in images that determine what the image shows.
Humans know that a car is a car because it has a certain shape and certain features, not because they have previously seen every type of cars possible. A standard neural network is tied to its input, and if the input is a pixel matrix, it recognizes shapes and features based on their position on the matrix. Convolution neural networks can elaborate images better than a standard neural network because
- The network specializes particular neurons to recognize certain shapes (thanks to convolutions), so that same capability to recognize a shape doesn’t need to appear in different parts of the network.
- By sampling parts of an image into a single value (a task called pooling), you don’t need to strictly tie shapes to a certain position (which would make it impossible to rotate them). The neural network can recognize the shape in every rotation or distortion, thus assuring a high capacity of generalization of the convolutional network.