AI, Machine Learning & Deep Learning - What does this really mean anyway? AI Series #1

Any serious study starts with definitions. In our series about Artificial Intelligence (AI), we will also start by defining the main terms we’ll be using, then we’ll try to understand how Machine Learning led to Artificial Intelligence. We’ll discuss the ethical questions of artificial intelligence, the alignment problem, and security issues. Finally we’ll see how AI impacts different industries, with a special focus on e-commerce & retail as they are leading industry in term of data generated. 

Let’s get back to basics. Besides the buzz words, what is Artificial Intelligence and what do we mean by that? 80% of the startups our team are meeting claim to produce and use Artificial Intelligence. I’ve even heard someone at a tech meetup stating that it’s perfectly fine to use the concept of Light AI as soon as we are using algorithms (ie. ALWAYS). This abusive use of the word ? AI ? aims to attract investors and public or governmental funds. We tend to be much stricter with this terminology.

In this series and in all future articles, algorithm is not what I mean when using the term AI, and I’ll try to explain why basic algorithm and AI are different at their core. But instead of giving a Wikipedia definition, let’s put some context and understand why AI became so popular recently.

In 2008, the buzz word was Big Data. One of the first appearances was in a Gartner study, to nominate the new structure of data collection. With the democratization of internet at high speed and the rise of smartphones, more and more data has been created, and we could not analyze them with the old way.

With this amount of data, simple algorithms were not enough: too many data to treat, with too many options, and too many instructions were necessary to gather, store, and use this data. This led to the apparition of new techniques, and new jobs like Data Scientists. 

But as we mentioned before, too much data is gathered, and no set of simple instructions is enough to get the most value out of them. This is why Machine Learning became so popular.

What is Machine Learning and how it works?

Quoting McKinsey “Machine learning algorithms detect patterns and learn how to make predictions and recommendations by processing data and experiences, rather than by receiving explicit programming instruction.”

In other words, instead of creating algorithms with suite of “if” and “then”, we develop algorithms with the ability to understand relationships between set of data.

Now, we have a clearer understanding of what AI truly means. AI is the ability for a machine to create its own set of relationships between inputs and outputs, by analyzing large sets of data. Machine learning is not a new concept (it appeared in 1958), but its efficiency is highly correlated to the amount of data.

Different techniques are used in Machine Learning, and one becoming more and more popular is Deep Learning, combining different algorithms structure that are inspired by the human brains. (1)

One basic concept used in Deep Learning is the Neural Network: Interconnected Layers of software-based calculators known as “neurons” form a neural network. A Neural network is created with 3 major layers: Input Layers, hidden layers (where calculation takes place) and output layers.

Using a Neural network, software makes a huge amount of calculations, allowing it to understand the correlation and relationship between Cause (inputs) and Effects (Output). For example, using Neural networks, companies are able to identify humors and feeling depending on the wording used in online reviews or social media posts.

Revuze, an Israeli-based startup is helping brands understand how they are perceived by consumers. Using client interactions as inputs (reviews, social media, email with support teams etc.), Revuze software understands the humors of clients, and can give insights to brands or retailers.

A Neural Network can be simple or recurrent. In the case of recurrent neural network, layers of outputs & inputs can be interconnected, meaning the software is able to assess the probability of the next input value. This technique is widely used in Fintech to assess the likelihood that a credit-card used s fraudulent, like Riskified or Forter. Using wide-range of information concerning user-behaviors, fraud-prevention companies are able to assess the probability of a user to be a fraudster.

Part of my role at Keyrus Innovation Factory, is to understand very quickly what a startup does, what is unique about them, and for which scenario they provide a real value. It has been crucial for me to understand the different types of machine learning, and it will probably help you understand the difference between companies.

The three Major types of machine Learning

·       Supervised Learning: When algorithms use training data and human feedback to learn the relationships between inputs (comments) and outputs (mood). Continuing our example with Revuze, they define what is positive and what is negative, and the software is creating the link between words used in reviews/comment and feelings.

·       Unsupervised Learning: When algorithms explores input data without being given specific outputs.  For example, Optimove uses software to discover patterns concerning consumer behaviors and segment clients according to certain behaviors. Using past consumer interactions as an input, Optimove can predict user behavior and give Marketing teams insights on how to segment and interact with each segments to maximize all customers KPIs.

·       Reinforcement Learning: When an algorithms learns how to perform a task by trying to maximize the reward. This technique is widely used in Machine learning and enables programmers to define primary and secondary goals, and ask the software to maximize these goals. For example, Facebook algorithms are trying to maximize the number of interactions (like, share, comment) you have with your news feed. This is why, the more you use Facebook, the more interesting your feed becomes for you. This example forces me to balance the positive effect of reinforcement learning: if, like the majority of the planet, you react more easily to negative news (comment only posts you disagree on), then your Facebook News Feed will show more of that over time and your negative feelings will rise.

So now that we have established the core elements of Machine learning and AI, we know one thing for sure: the more data machine learning software has to process, the more precise it will be.

Today, with the penetration of Internet globally, the rise of connected objects and Internet of Things (IoT), we generate each day the same amount of data generated in 2005. 90% of the global data in the world was generated over the last two years. (2) Not only do we generate more data every day, but thanks to microprocessors companies we have the ability to treat this data at the highest speed in history. We are computing data 10 times faster than a few years ago, which allow software to use more and more data for their predictions.

In a previous article, we have discussed the Autonomous vehicles (AV) and the Ethical questions still not resolved. Since then, I’ve learned that only the camera necessary for the AV generates between 20-40Mb/Sec, while the light detection & ranging (to stay at a safe distance from other cars) are creating between 10-70Mb per seconds. In less than an hour, one autonomous vehicle is creating more data that I have on my personal laptop.

This figure explains what one means when speaking about Data-centric world. More and more decisions are based on data. Once upon a time, Marketing was based on consumer interviews, panels and other fastidious techniques. It is now based on well stored data. With each click, swipe, share, and like, a world of valuable information is created. The ability to make data-driven decisions has become crucial to any business.

With that in mind we understand why Machine Learning, Deep learning and AI have created the buzz, and the confusion it creates. In our next articles we will address the difference between AI as we have discussed today, and General Artificial Intelligence – what questions in regulations, ethics and creation of conscious minds. We will then see how fare we are from developing a new conscious minds and where we are standing now in term of AI applications. This will lead us to dive in Retail & E-commerce, one of the markets with the most advanced AI technologies. If you have ideas/insights on technologies or companies I should know about, don’t hesitate to contact me, in private message or comments.

1. McKinsey Global Institute: An Executive’s guide to AI
2. https://www.forbes.com/sites/bernardmarr/2018/05/21/how-much-data-do-we-create-every-day-the-mind-blowing-stats-everyone-should-read/#29076ad260ba