Machine Learning Use Cases: iTexico’s HAL
The startling reality of most major society-affecting products runs along the concept of creeping normality; that a subject will not notice changes in its environment until the change is normalized. Such advancements, like the ever evolving study of medicine, the prevalence of touchscreen technology, and the shifting landscape of popular music, are onset cases that rely on hindsight to notice their rise in modern culture. Change is inevitable, and tends to happen before we’re aware of it.
AI technology has been the focus of large-scale attention for decades, both as science-fiction theory and conclusive scientific performance. Now, with all the grace as a leaf on the water, AI has become that backbone behind which the world works. We’ve been inundated with mundane AI usage, such as smart replies that Google has implemented in their Gmail service since 2015. The smart reply function utilizes machine learning to automatically suggest three different brief, customized responses to quickly answer any emails you may receive.
Another example, coming from the retail industry, comes from Lowe’s as a method of effective store management. Through AI use, the company keeps track of their inventory in real time, as well as detecting sales patterns that help them guide business decisions. Small cameras, placed on top of shelves, monitor and stream real-time information on shelf-stock levels. This technology detects when a hole appears in any section, such as the light bulb aisle, and the system then sends a real-time notification to the store’s devices so staff can quickly head to the stock room and replace the item.
Amazon, Apple, Facebook, and countless other organizations have already taken the step toward AI incorporation into every functional aspect of their business. Aware or not, chances are you’ve already been making full use of AI technologies for years now, with more promised on the way.
What Is Machine Learning?
Machine learning is a subset of AI, and while it falls under the umbrella, it’s designed with a function that other AI programs are not capable of: learning. In machine learning, you need to program the machine to do what you want, and it will “learn” by repeating the same task repeatedly until it does exactly what you want. Just like the human learning process, the program learns through trial and error while developing its own unique solutions to problems and stimuli that may not have been pre-programmed into it.
AI and machine learning, while similar, are not the same concepts, and it’s an important distinction to make. You’ve already got an idea of what artificial intelligence is - that is to say, a collection of algorithms designed to mimic human intelligence - but what is machine learning, specifically? According to SAS, machine learning is the scientific study of algorithms and statistical models that computer systems use in order to perform a specific task effectively without using explicit instructions, relying on patterns and inference instead.
Just about as clear as mud.
As it stands, AI and machine learning still sound like the same basic idea: programs that simulate intelligence. It is, as you may have guessed, a bit deeper than the surface explanation. The most important distinction, as well as the easiest to understand, is the fact that AI is the parent body, but is generally has a pre-set collection of algorithms in which they can process data and respond to stimuli. Machine learning, by comparison, does all this while learning and extrapolating on the information its exposed to.
Artificial intelligence means training a computer to mimic human behavior through these algorithm collections, which are executable mathematical functions that process data through a series of rules and step-by-step instructions. Impressive, yes, but AI isn’t designed to develop beyond the parameters that are established for it.
Types of Machine Learning
The AI umbrella has a tendency to trickle down into progressively extensive, particular subsets of the same concept. In the same way that machine learning is a subset of AI, there are also three different types of machine learning. These discrete subsets are defined by the methods by which they learn from the data they process: supervised, unsupervised, and reinforcement.
The main premise here is “learn through example.” For instance, if you’re planning on teaching a baby their colors, you use real life examples to walk them through it: red apple, yellow banana, orange orange. Supervised learning follows a similar method, wherein established labeled examples are used as the baseline for inputs with known desired outputs. So long as the algorithm has the correct outputs to compare the inputs to, it will be able to find errors and learn how to use patterns applying historical data to predict future events, such as credit card fraud or insurance claims.
Where supervised learning has access to and applies historical labels, an unsupervised learning method does not. Without having the correct outputs to compare against, the algorithm will self-determine from the processed information by exploring and determining structure within data inputs. These are often used to relegate similar structures of customers to help organize marketing campaigns directed at appropriate groups using transactional data. Using the baby example, this method of learning is like a baby learning language through listening to their parents.
Reinforcement learning is the trial and error method for algorithms. The goal for this machine learning subset is to achieve the greatest reward through the most efficient method possible. In fields like robotics, gaming, and navigation, this type of machine learning is critical for developing algorithms that are capable of assessing and interacting with their environments. In these instances, the agent (being that algorithm) will choose actions that result in the greatest expected rewards in the fastest time possible. This process is repeated over and over, reinforcing the best policies for the machine.
Our Very Own HAL
You may already be familiar with this image, Dave. This is HAL 9001, an AI program from the movie 2001: A Space Odyssey, directed by Stanley Kubrick. In the movie, HAL controls the system of the Discovery One spacecraft and works during an interplanetary mission to Jupiter. HAL was capable of many activities such as: Speech, Speech Recognition, Facial Recognition, Natural Language Processing, Lip Reading, Art Appreciation, Interpreting Emotional Behaviors, Automated Reasoning, and could even play chess.
AI technology is actually part of our daily lives at iTexico, where we have been able to adapt and transform manual activities to sophisticated AI algorithms that make our professional life a lot easier. Our HAL is a facial recognition and control system based on Azure, that works on machine learning and a neural network, built by our AI experts at iTexico.
HAL, named after the Stanley Kubrick’s Odyssey HAL system, was born as a solution to a very common problem within mid-growth organizations, which is having efficient communication and control systems as the workforce keeps growing year by year.
The way HAL works as an attendance device for the employees is by identifying the faces of our employees through cameras and sending over the information to Microsoft Cognitive Services for ID. Once the person is correctly identified, the system then returns basic details of the person such as their full name and the exact time when they entered the iTexico premises. It uses Machine Learning to learn from a large amount of data from our ERP to predict business scenarios and human behavior.
This technology has allowed us to increase productivity with a 93% monthly reduction of attrition, and also works as a chatbot contacting on average 25 of our employees a day to remind them of basic information such as meetings or data needed for their daily activities.
This system has already been sold to other companies, allowing them to implement this control mechanism within their processes.
How Does iTexico’s HAL Work?
Our AI software, HAL, can do many things, but what’s important to us is how it maintains its focus running on a human business model. That is to say, everything HAL does is designed to ensure that our lives, as humans, are a little bit easier. To meet this demand, we’ve programmed HAL with two key algorithmic classifiers: Naïve Bayes and KNN.
The Naïve Bayes algorithm is a commonly used family of algorithms in machine learning circles. We say family, of course, because it is not a single algorithm, rather a collection of related code based on Bayes’ Theorem that share a common principle. Bayes’ Theorem’s intended use is to find the probability of an event occurring given the probability of another event that has already occurred. Using this, Naïve Bayes is able to assert that all pairs of features are independent of each other, which allows machine learning software to compare pre-existing data and determine an outcome based on those independent variables.
KNN, otherwise known as k-nearest neighbors algorithm, is another popular machine learning code, much like Naïve Bayes. As a classification model, KNN relies on identifying votes of neighboring data points to a predetermined K property to classify information.
Source: Antti Ajanki
This example is simplified, but by assigning a voting number to the K property while attempting to sort red circles from green squares, KNN can separate and classify distinct data points. In this example, the green circle must be assigned to either blue squares or red triangles. If K is 3, then it will be assigned to the red triangles. If K is 5, it will be assigned to blue squares. While this information is sorted, KNN is a type of lazy learning, which means its data is generalized until a query is made. Until then, the actual classification of any data is left deferred.
Use Case: Employee Attendance
So where can HAL be effectively utilized? Let’s set up a probable scenario. For context, iTexico generates revenue from the time worked on a specific project. Obviously, it’s precisely due to this that employee attendance records are critical, ensuring that the proper allotted time that an employee put in to work on the project is compensated for. However, there’s been an issue: employees haven’t been filling in the timesheet on time, meaning iTexico can’t send the invoice to their clients.
Filling out the timesheet may be important, but it’s one of those simple, repetitive, and tedious tasks that tends to be preferred to be forgotten about. So, AI software like HAL are designed with the solution in mind to learn from the employees and remind them about this task when they are more receptive to the message.
As a result, 94% of the employees are promoters and 72% fill their timesheet on time.
We’re proud of the work we’ve put into creating HAL and, quite frankly, excited to see where it will go in the coming years. As machine learning technologies continue to develop, we’re going to find a greater and greater presence in our daily lives as they make things better, faster, and easier. Don’t worry, while they may share a similar name, our HAL only shows off its best behavior.
Machine learning, as we know, does not come in just one flavor. Proposed with three distinct types of learning - supervised, unsupervised, and reinforcement - it’s a highly versatile algorithm for nearly any sort of practical use. HAL, running on KNN and Naïve Bayes classification algorithms, is an excellent addition to any workplace’s environment.