The Engineering behind Machine Learning

3 min read >

The Engineering behind Machine Learning

Advanced Technologies & Embedded

How Computers Learn Natural Language

Google is making a big impression right now with the launch of the Pixel Buds, the earphones that can translate conversations between 40 languages. This is just the latest display of remarkable machine learning power, applied to human language.

After years of trials and small steps, machine learning has given computers the power to understand and use words and sentences at a good level for a wide range of practical applications.

As a result, computers have partially overcome challenges such as language understanding, language generation, communication context, and dialogue. These achievements rely on a rather complex process that “teaches” the machine to place words, sentences, and even phrases in a context, creating its ability to process language.

Repetition, the mother of machine learning

Practice is at the heart of learning, whether we talk about humans, animals, or computers. This is why machine learning algorithms rely on repeated experimentation with a wide range of data, leading to the creation of context around various bits of information.

Here is an example: let’s say we want to teach a computer to act as tech support.

It means that customers will call in to explain their problem and the computer will process the information to suggest a probable solution. The tricky part is to teach the system to understand the customer’s problem.

This is where the machine learning magic begins:

  • read a lot of requests from many different customers. It will vectorize each sentence, word by word, forming a new dimensional representation in the shape of a matrix.
  • This new representation will be served as input to the algorithm, which will transform it furthermore, using several layers of representations. This will be re-run several times through the algorithm, knowing the expected solution associated with the input.
  • During this process, a very large number of parameters (weights) is adjusted in order to learn the context. Adjustment takes place gradually, at each re-run, in a process known as “gradient descent learning”. It means that the system makes small changes, over and over, for each parameter, until it gets things right. The process is repeated until the machine cannot improve furthermore
  • Eventually, the system should be able to decide that “My mom’s phone has no signal” and “I can’t call anyone” might be about the same problem – no SIM card in the phone.

This is just one of many different approaches to machine learning. Based on them, many functional, real-life applications have emerged on the market.

Read more in our next story about the most disruptive uses of machine learning in natural language processing.

For 12 years Tremend has successfully delivered end-to-end solutions ranging from complex banking platforms to eCommerce solutions and embedded software. We use advanced technologies like Artificial Intelligence, machine learning, IoT, blockchain, and microservices for clients from industries such as banking, finance, telecom, and automotive.

Feel free to contact us at hello@tremend.ro for support in developing your own software projects.

Unlock innovation and digital transformation for your business.

Get in touch

We are always happy to talk

Phone

+40-21-223-7700

Email

hello@tremend.com

Address

165 Splaiul Unirii, Timpuri Noi Square,
TN Office 2 building, 4th floor,
District 3, Bucharest, Romania, 030134