As technology continues to advance rapidly, the machines we use are getting smarter. Machine learning is the technology of constructing “learning” algorithms that drive a broad range of smart technologies — and the new generation of this discipline, called deep learning, has the potential to power more advanced artificial intelligence capable of everything from sophisticated speech and image recognition, to self-driving cars.
What is deep learning?
Deep learning, also called deep structured learning or hierarchical learning, is a type of machine learning that uses high-level data abstractions, nonlinear transformations, and layered cascades applied to learning representations of data, in order to help machines “learn” tasks through observations and examples.
Algorithms with deep learning applied are often inspired by communication patterns found in neuroscience — the study of the human nervous system. For example, a deep learning algorithm might be based on the relationship between a stimulus and a neural response, which registers as electrical activity in the brain. This type of machine learning attempts to create neural networks for machines that “think” in ways similar to humans.
Following are a few of the applications currently being developed with deep learning algorithms.
Automatic speech recognition
Technologies such as Apple’s Siri are built on machine learning algorithms that work to recognize speech, including words and sounds. Deep learning has led to the advancement of automatic speech recognition using the TIMIT data set — a limited-sample database using 630 speakers and eight major American English dialects, each with 10 different spoken sentences — to large vocabulary speech recognition through DNN models that rely on deep learning algorithms.
Deep learning differentiates from other forms of machine learning through the use of raw features at a learning level, rather than pre-constructed models. With deep learning, speech recognition can be highly accurate using the true “raw” form of speech — waveforms, or visual representations of sounds using curves.
Similar to speech recognition, a limited size data set called the MNIST database has been the popular model for powering image recognition applications. This database includes 60,000 training examples and 10,000 test examples, composed of handwritten digits. However, MNIST relies on shallow machine learning for image recognition — and deep learning allows for more large-scale image recognition at a higher accuracy rate.
One practical example of deep learning algorithms applied to image recognition can be found in the automotive industry. A car computer trained with deep learning may enable cars to process and interpret 360-degree camera views, allowing for heightened “awareness” in self-driving or assisted-driving vehicles.
Many in the tech industry view deep learning as a strong step toward realizing truer artificial intelligence. In 2013, Google hired three DNN researchers tasked with not only dealing with the search engine giant’s constantly growing stores of data, but also to improve Google’s existing machine learning products, such as semantic role labeling and search results.
Facebook has also created an artificial intelligence lab, largely dedicated to the development of deep learning techniques that will improve the user experience. Automatic image tagging was developed in Facebook’s AI lab — a technology that is still being refined for greater accuracy using deep learning.
As machine learning continues to increase in sophistication, more companies will look to hire IT professionals interested in developing deep learning algorithms and improved artificial intelligence applications. Machine learning is an exciting field with a wide range of possibilities ahead.