This article was originally published in The New Stack
on May 8, 2020
In part two of this series, we explored how
computers can learn from data using machine learning (ML) — without explicit programming or instructing the flow and logic of learning processes. We
also explained how computers can discover and learn patterns and correlations from any data — no matter where it comes from or what it is about.
In this final article of the series, we’ll focus on deep learning (DL), artificial intelligence (AI) and explore how computers can make use of human brain structure
to perform natural language processing, image recognition and much more, in some cases surpassing human expert benchmarks! Also, we will learn the capabilities and
limitations of modern AI.
On a high level, DL is a subset of methods within machine learning. It is another algorithmic approach for implementing ML and it brings us a step closer to an
understanding of AI.
Deep learning is based on Artificial Neural Networks (ANN) and feature learning. To understand DL better, it’s important to understand what lies in its core.
The Core of Deep Learning: ANN
ANN is inspired by the structure of a human brain and our current understanding of how a brain processes data through the interconnections between neurons.
Compared to a human brain, the current implementation of ANN has a more simplified structure with discrete layers and a defined number of connections and directions
that the data can flow in. ANN can be trained with or without a teacher (supervised, unsupervised, or semi-supervised), like other ML models.
For example, you can break down an image into a matrix of consumable data that will serve as the input data for the first layer of ANN. From there, each neuron of
the first layer will pass the data to a second layer, and so on, until it reaches the final layer where the resulting output is produced. For instance, the output
may suggest that a cat is in the image. For each image that the ANN consumes, each neuron of each layer assigns or adjusts the weight to its input which indicates
how correct or incorrect it is relative to the task. The total of the weights will determine the result.
The science behind ANN is not new and it has been advancing for more than seven decades. You may guess that ANN requires powerful computational resources and,
indeed, the revolution in DL occurred when the hardware advanced so much that it enabled a renewed interest to this field. Nvidia has played a huge role in the
so-called “big bang” of deep learning as the ANN models were trained with Nvidia GPUs (graphic processing units).
The Role of Deep Neural Networks
Deep learning is a set of different architectures of ANN called Deep Neural Networks (DNN), such as Recurrent Neural Networks, Deep Feed Forward,
Deep Boltzman Machine and many more. Each DNN has its unique set of strengths and areas of application where it fits best. It also uses the feature learning
methods that we described in the earlier machine learning section.
Deep learning has been applied to speech recognition, natural language processing, social network filtering, image analysis and more fields, where results
comparable to or even sometimes surpassing the human expert have been produced.
Deep Learning vs. Machine Learning
Deep Learning is a subset of ML, but the new capabilities that it brings to the ML field take it to an entirely new level.
For example, deep learning does not require you to research, develop, and use new features. Instead, the neural network learns how to select the most critical
features. The training time is also longer for DL since it requires a massive amount of training data. Furthermore, with DL, it is more difficult to explain the
results. You can find out which neuron was activated, but it is harder to interpret what neurons were doing collectively.
Unlike ML, DL may provide near-human performance and, in some cases, surpass human performance. Machine learning still requires some guidance from a data scientist,
such as manually identifying and creating new features. However, in most algorithms, ML provides a clear set of rules which can be used to explain the results.
Deep Learning vs. Data Science
Data science and DL differ in the following ways:
Data science is an interdisciplinary field that involves many processes for collecting, cleaning, analyzing, visualizing the data and developing predictive
or prescriptive models to solve a particular task or problem.
Deep learning (along with its broader ML field) techniques are often used in data science to deal with a vast amount of data, find critical patterns and new
features from the data, and create predictive/descriptive models for analytics.
You’ve probably seen a science fiction movie about AI at least once. It is one of the hottest technology topics. AI has many flavors and has been heavily
popularized by the media and entertainment industry — sometimes close to reality and in other cases highly exaggerated.
AI is the capability of any device or software (agents) to display human-like intelligence. AI-capable agents have cognitive functions that allow them to observe,
learn, take action, and solve a problem or task autonomously.
However, AI today is quite different from the humanoids we see in movies that can walk, talk, think, and act like humans or even better. While that’s just science
fiction, there are great advances in AI which fuel our anticipation that this may soon be a reality.
Here are a few examples of popular AI-enabled devices and services:
Personal voice assistants, like Siri or Alexa, use sophisticated voice recognition to translate, interpret, and act on customers’ questions or requests.
Self-driving vehicles created by Tesla are transforming the data received from 12 integrated cameras into a vector space to recognize objects,
road conditions, and specific situations that enable them to autonomously drive and activate safety measures in case of emergency.
E-commerce platforms are getting smarter in predicting what we are interested in purchasing, by analyzing past purchases and our behavior on their website.
Nest, a smart thermostat that learns from your heating and cooling preferences, autonomously adjusts the temperature in your home.
Robots, such as Neato vacuum cleaner, use lasers to draw a floorplan for smarter navigation and avoidance of obstacles on their way.
Artificial Intelligence vs. Machine Learning vs. Deep Learning
AI, ML, and DL come together to provide powerful data and predictive analytics capabilities. In summary, AI is a general area of automation of any task that
requires cognitive decision-making and problem-solving.
ML is a subset of AI, but not all AI is created using ML. Instead, ML is a set of algorithms and methods that enable AI to process, learn, and analyze the data;
perceive from the environment (image, voice, video stream and other sources); and solve a particular problem or task. AI heavily uses ML and DL techniques
to autonomously process and learn from data, identify critical patterns and features and make decisions or classifications.
We are closer than ever to creating human-like AI, but there are still many unknowns. The progress in technologies, computational power, and the constantly
improved understanding of how the human brain works are contributing to the exponential growth of AI.
Many believe that the AI of the future is a threat to humanity due to its potential to overwhelmingly outperform the brightest minds on Earth. We believe that the
progress of AI is unstoppable, but while it is hard to predict the future, we will continue contributing to a better, smarter and safer AI of the future.