- 17 Aprile, 2019 - 8 ’ read

Artificial Intelligence applications and use cases. A look into the future.

Artificial Intelligence, Deep Learning, Machine Learning

Simone, 26, entered Imagicle R&D dept. this year while writing the last lines of his degree thesis. In an innovative environment, he is the ultimate innovation: he's studied artificial intelligence! If you want to discover the main applications of this new magic technology to the tools you use every day in your tech company (and find out how we make it real in Imagicle!), you really need to read this article. And now, just because any reference is purely coincidental...let's go back to the future!

Artificial Intelligence (and me) in today’s world.

What led me to artificial intelligence? Honestly, the desire to cross borders!

I know it may sound odd, but AI is simply the sound science for curious people who want to extend the technology to new and different fields. In short, for people who do not accept borders.

Of course, after curiosity, there is a long, in-depth study.

Courses like Data Mining or Intelligent Systems, held by the computer engineering University of Pisa, have helped me in broadening the vision of data from mere statistical descriptors to real protagonists of the revolution taking place. (Well, also living in an age where data are considered as “the new gold” has played a part).

During my studies, I was lucky enough to use my knowledge to develop projects involving the use of ML algorithms to create motion classifiers and sentiment analysis, and finally, when I joined the Imagicle troop, I discovered the best way to apply machine learning to something really handy. The stuff I’m going to talk about, though, can be extended to Unified Communications products in general, so, basically, there’s a little something for everyone.

Talking about artificial intelligence is never easy, as it never is to talk about things still little known.

In recent years, AI dynamic nature and the fantastic results that can be obtained through it, have consecrated artificial intelligence as the must-have technology for all companies working in the IT field (…’cause hey!, it does magic!). Once you get it running, everything seems to work smoothly without “so much” effort.

The truth behind the curtains is very different and much more complex than common sense suggests. In this blog post, we’ll try to give a very entry-level introduction to AI and Machine Learning (ML) through the study of some common Unified Communications use cases. Moreover, we’ll briefly introduce common problem companies struggle with when trying to implement AI in their products.

So, guys, if you’re ready to take a nice trip through the real possibilities of the near future, fasten your seatbelts. Let’s roll!

Artificial intelligence: what it is and why it matters.

A great artist once wrote that “To define is to limit”. Well, we could really use some limitations, here, at least not to get lost in the extreme vagueness of a “still in progress” technology.

Artificial Intelligence applications and use cases. A look into the future.

A standard way to define AI can be: “The theory and development of computer systems able to perform tasks normally requiring human intelligence”.

As far as we are concerned, AI is a technology that allows making our daily actions smarter, faster and more effective, especially through the use of Machine Learning (ML), or “The study of computer algorithms that allow computer programs to automatically improve through experience”.

In fact, AI and ML are closely related: the first one represents the purpose, what we want to obtain (WHAT); the second one represents how to achieve it (HOW).

Of course, ML is only one of many ways to obtain AI. For instance, Deep Learning (DL) is an ML subset that takes advantages of deep neural networks architecture to extract many representations level and obtain much better results than ML in particular scenario (e.g., computer vision).

Machine and Deep Learning are, in their turn, AI subsets that allow software applications to be more accurate in predicting results through algorithms, without being explicitly programmed. Therefore, they may also be defined as forms of automatic learning. This can be categorized as supervised or unsupervised. The first category requires the presence of one (or more) membership classes, input data, desired membership classes, and a training phase to obtain a predictive model. Once obtained, it is possible to apply it on a new set of data to obtain a prediction of its class of belonging. The typical example of supervised learning are classifiers (e.g. decision trees, k-Nearest Neighbors).

The second category (unsupervised learning) does not include a “strict” training phase; there are no predefined classes of belonging, only raw input data. In this case, the goal is to search into raw data entities that are similar to each other and dissimilar with the rest. In this way, it is possible to extract classes of belonging without the aid of supervised techniques, perform outlier detection and conduct a behavioral analysis. The typical example of unsupervised learning is clustering (k-Means, DBSCAN, Hierarchical clustering).

“Predicting the future isn’t magic, it’s artificial intelligence.”

Dave Waters.

General applications.

And now let’s break-in on the action. Let’s see what we can actually do with all this future.

Of course, due to its nature and versatility, ML has multiple application domains. The evolution in the technological field has recently allowed having more and more specialized and miniaturized hardware. As a consequence, the application of ML algorithms has become pervasive, starting with wearables and ending with data warehousing.

Let’s examine a few domains which are currently “hot topics”.

Classification.

Here is the typical problem we can solve and the most general application you can find in real-life ML. Given a dataset with a feature set and its classes, you can create a model that allows you to classify new data by associating them with the class they belong to (where classes are discrete values).

Regression.

Given a data history up to a time t, it’s possible to create a model (linear or logistic) that allows the prediction of the data history at time t + T. Both classification and regression are predictive models. The main difference is due to the predicted values: in regression, they are not discrete. A typical example of regression application can be a stock market forecast.

Image Processing.

Given an image, you can perform object detection and segmentation in the scene, search for the most similar k-top images, perform face recognition and much more. This is possible thanks to convolutional neural networks and the leveraging of specialized shortcuts (e.g., region proposals, neural support networks).

SpeechToText and sentiment analysis.

Given an audio file, it’s possible to extract the text and use it for different purposes (e.g., speech recognition, sentimental analysis). New models can even recognize the participants involved in a discussion, their number and mood in real-time, combining multiple ML algorithms. Also, it’s possible to catch the feeling of the person who wrote the text, intercepted as a classification problem. For instance, possible classes of emotions can be: positive, negative. Finally, the neutral feeling class can be discriminated.

Continuous Learning.

AI, ML, DL are dynamic technologies by nature; consequently, even the generated models must be so. The installation of a continuous learning system allows the updating of previous models with the use of new data, in order to improve and adapt them to the specific context.

Machine Learning allows capturing the change of habits within a model continuously and dynamically, thus obtaining potentially faster, better, more relevant results.

The Data Problem.

Every tech company would like to leverage AI benefits; only a few succeed in it. The Data Problem is related to the intrinsic data value; the main goal is to extract knowledge from it, no matter if using Data Mining, ML, DL and so on.

Companies that want to introduce AI in their products should make sure they have actual useful data to start with, independently from the way they are collected. Running ML algorithms on data that are not valuable will not produce the magic tricks they are willing to get.

In addition, the data size required to train and obtain good AI models can be very large. Indeed, due to the Data Problem, many companies drop the AI hype train and does not implement ML technologies in their solutions.

[Spoiler alert!] How to improve your UC systems with AI.

After a far-reaching investigation about the possibility of introducing this new technology into our company, we’ve come to identify some application areas that we want to develop.

The first step will be to enrich and enhance our Call Recording solution.
By inserting additional modules, we will enable particularly attentional functions, thus enriching and making the application more personal through third-party AI SpeechToText algorithms (e.g., Google).
Then, based on the results, we can start thinking about additional modules to use internal ML algorithms that allow textual analysis to obtain significant business outcomes, and so on.

(By the way, don’t know our Call Rec solution yet? No prob! Just press play!)

https://youtu.be/ihvCFJXRi-0

And this is just the start: the Imagicle future holds many other surprises.
We’ll rattle off some of them through the use cases you find below – those that you’ll see integrated into our applications soon!

Wanna take a closer look to see if you find them as amazing as we did? All right!
Just make sure you keep it under your hat.

Queueing Time Prediction.

Tired of calculating and vocalizing queue time to your customers roughly or using algorithms that don’t reflect your needs? Don’t worry! The use of ML algorithms allows you to estimate and vocalize more precisely, in line with the typical behavior of your UC, reducing the effort. Given the state of your system, queueing time can be predicted using a supervised approach.

Recording Search Engine.

Another AI application within the UC context is related to call recording.

Has it ever happened to you to remember only the content of a conversation you want to listen to again, but not the day it happened?

Well, then I guess you’ll be happy to learn of the birth of a “Google-like” search to seek for past calls related to the topic being searched, listing them in order of relevance.

SpeechToText technology makes it possible to extract text from recordings and create a concrete search engine.

Yes, yes: I’m aware it’s jaw-dropping stuff.

Advanced Contact Search.

The use case is similar to the call recording search engine one, but applied to the contact search. Through Text Mining algorithms, creating structures able to provide a ranked contact search and fittin the address book of the single employee could be possible.

Behaviour Analysis.

Through unsupervised learning, it’s possible to identify abnormal behavior within a given set and highlight any emerging classes of action within one’s own company. More generally, it’s possible to use unsupervised learning to characterize the behavior of similar entities, dissimilar from all others. Knowing the rate of adoption of particular tools within your company, for example, can lead to an optimization of the management of resources and employees.

More generally, behavior analysis can be related to data analytics: it can be seen as a data preprocessing step to apply before the AI training phase.

Are you up to the challenge?

In order to keep up with the new emerging technologies, it is necessary to keep up with the times and offer customers increasingly competitive, smart and efficient solutions. The introduction of AI issues within products is a step that requires a critical analysis of the business value that these technologies can add.

Those you’ve read above are just a few examples of the concrete advantages deriving from the extensive use of this technology. Artificial intelligence makes it possible for computers to learn from experience, adjust to new inputs and accomplish specific tasks by processing large amounts of data (and just for the record: it’s not a replacement for humans

So, go through the article once more from the bottom up. If your company involves at least one of the use cases shown, it’s really time for you to get on the artificial intelligence hype train.

But first tell me, are you ready to revolutionize your business?