A dive into the deep waters of machine learning

Can an algorithm mimic the behavior of the brain? This is the premise of the concept of deep learning. While it is far from matching the cognitive performance of humans, its architecture – artificial neural networks with three or more layers – makes it possible to learn from large amounts of data and issue extremely accurate and optimized predictive analytics. Not only that: it improves over time with its own performance.

At the core of machine learning (ML) are deep neural networks with multiple layers of interconnected nodes. Each node relies on the data received from the previous layer to refine and optimize the prediction or categorization. These calculations through the network are called “direct propagation”. In the so-called “input layer” the model ingests the data and in the output layer, it delivers the prediction or classification to be used by the next layer. The back-propagation, on the other hand, calculates the errors in the predictions and then adjusts the function weights and biases, training the model itself. 

Machine learning and deep learning: similar but different

Deep learning is often confused with machine learning. The reality is that the first one, is just a subset of the second one. The main difference between the two relies on the type of data they use as raw material and in the learning methodologies they apply. While both work with supervised learning, deep learning also excels in unsupervised learning: it detects patterns and groups data by any distinctive feature.

Thus, in general, ML uses structured and labeled data, with characteristics that are defined based on what the model needs to work and provide results. Even when unstructured data -images, videos, social media posts, geographic locations, among others- is used, it is preprocessed to be placed in a table so that the algorithm receives all data with some kind of structure.

Deep learning, on the other hand, can ingest and process unstructured data and automate the extraction of functions, which reduces (even to zero) the dependence on human involvement. For example, if a deep learning algorithm is set to categorize photos of cars by brand, it will be able to determine the distinguishing feature to identify each one, whereas in machine learning, it will probably require a human expert to manually define the recognition pattern (such as the logo on the front).

Deep learning everywhere

We are in continuous contact with deep learning solutions, frequently used for analytics and automations related to artificial intelligence for processes where there is no human intervention. 

Digital assistants capable of carrying on conversations in natural language, intelligent translators or algorithms capable of detecting financial fraud, are just a few examples. In healthcare, image recognition apps are key to anticipating illnesses or researching drugs or treatments. Every time we give an instruction to Siri or Alexa, we are interacting with a deep learning solution.

This is just the beginning

Depending on the specialization, deep learning uses different types of neuronal networks. Convolutional (CNN) are used for image classification and computer vision, while recurrent (RNN) are used in natural language and speech recognition applications.

Of course, this is a technological concept that is still a long way from realizing its full potential: its presence is key in the development of autonomous vehicles or smart cities, among other emerging technologies. And, of course, to help businesses improve their processes, predict behaviors and take customer experience to the next level.

Deep learning: a paradigm that is worth starting to deepen.