Artificial Intelligence: Teaching Machines to Learn Like Humans

AI brain

Using the human brain as a model, machine learning teaches AI computers how to learn new things, recognize patterns and make decisions.

We learn by doing, Aristotle once said. Today, experiential learning doesn’t apply just to humans — machines are increasingly able to sense, reason, act and adapt based on learned experience.

It’s unlikely the ancient Greek philosopher ever dreamed artificially intelligent machines would also learn by doing — to improve precision medicine and self-driving cars, and even analyze data using processes similar to his own logic systems.

It’s taken more than 60 years for computer scientists to figure out how to make machines smarter, and this work still continues today. It turns out that doing versus telling makes a huge difference in machine learning.

In early artificial intelligence (AI) attempts, researchers tried to tell computers everything they needed to know. When these direct instructions failed, scientists tried machine learning strategies, allowing the AI programs to directly analyze and learn from data.

That was a smart move, according to Andy Hickl, former chief product officer at Intel’s Saffron Cognitive Solutions Group.

“Learning techniques took off with the convergence of three things,” said Hickl.

High performance computer power, access to more data online (including text, vision, images and sensor data), and annotating or adding notes to data — together, these three factors dramatically improved the ability of AI systems to infer knowledge, he said.

Today, a combination of software, hardware and the internet helps software developers program powerful AI applications that can do almost anything from detecting financial fraud, to winning at Jeopardy, to even chatting with customers online.

How Smart Machines Learn

Used in many industries today, machine learning relies on computer programs that can learn from data and improve from experience without directly being programmed, according to Niven Singh, the program and community manager for Intel’s Student Program for AI. Part of the Intel Nervana AI Academy, the program showcases innovative work done by graduate and postgraduate students at universities worldwide.

“The biggest advantage to machine learning is that it allows us to do things much more quickly than we’d be able to do otherwise,” said Singh.

“It can’t solve problems that a human being couldn’t also solve, but it can take in a huge amount of data and very quickly build connections and predictions based on it.”

In gaming, for example, machine learning algorithms can analyze competitive play to anticipate moves and create more challenging enemies.

In manufacturing, understanding data can help companies anticipate repairs and improve preventative maintenance.

And in healthcare, AI can analyze medical databases of multiple diseases, helping providers make diagnoses more quickly and accurately.

Machine learning is based on one of the most powerful computing machines around — the neural networks in the human brain.

Artificial neural networks (ANN) use algorithms to learn things, recognize patterns and make decisions. They mimic the way the human brain solves a problem — by taking in inputs, processing them and generating an output.

Machine learning identifies people and object
Through machine learning, computer programs learn how to identify people and objects.

The ANN knows which information is most important by using weights, said Hickl.

Artificial neurons connect to each other throughout the network, weighting the connections to show how strongly one neuron influences another. The weights are adjusted through a training program that teaches the neural network how to properly respond to its inputs.

Deep learning programs perform many layers of computations. Results from one layer feed into the next layer’s analysis to build up a more complex understanding of the data.

For instance, an ANN that analyzes images of buildings might detect edges in one layer, then recognize that the edges form a rectangle in the next layer. In the following layer, it would recognize the rectangle as a building and finally, in the last layer, it would determine that the building was a skyscraper or a barn.

Software developers train the ANN to learn using huge data input sets. Raw data by itself isn’t that useful, said Hickl, so developers annotate or add notes to the input data with the “right” answer.

Annotating the data is one critical component for success, said Hickl.

Self-Taught Computers

Through machine learning, computers learn without actually being programmed. Software developers create learning algorithms, allowing the ANN to improve as it’s exposed to more data over time, said Singh.

Learning algorithms come in three subsets: supervised, semi-supervised and unsupervised learning. The first two require large sets of training data to guide the desired results.

For example, to create a face detection algorithm, developers might provide images of landscapes, people and animals with their respective labels until the machine could reliably recognize a face in an unlabeled image, said Hickl.

AI facial recognition
Artificial neural networks use algorithms to recognize facial patterns and find matches.

“The third and the holy grail, for most of the work that we do in machine learning, is unsupervised learning,” said Hickl.

In this unsupervised approach, which Intel Saffron performs, the program simply figures out what the data means on its own. Instead of providing labeled images, developers allow the machine to cluster images based on shared characteristics that humans may not see, said Hickl.

Tools for Building AI

Despite the mystique surrounding AI, tools for building smart systems are actually widely available. Modern development doesn’t use old, specialized AI languages like Lisp and Prolog. Developers now rely on general purpose languages like Python, Java and C++.

And there are great toolkits for supervised learning, said Singh. Some programs, such as Caffe, Theano, MXNet and TensorFlow, are optimized to run on Intel hardware. High-powered chips like Intel’s Xeon processor can do high speed matrix multiplication that machine learning algorithms require.

Singh is excited by the progress of AI’s growing ability to make decisions based on analyzing and fusing different kinds of data.

Just like humans, machines will continue to improve as they learn … by doing.

 

iQ newsletter signup

Share This Article

Read Full Story