AI Series - Machine and Deep Learning, how it works and what it can do for your business
As part of my “Introduction to AI Series” which is designed to cover the five most relevant aspects of AI for Business, I am focusing on Machine and Deep Learning.
In an earlier post, I covered Natural Language Processing (NLP), and my next post will be on Computer Vision and robotics.
Machine Learning is the primary driver of AI applications for most businesses. Machine Learning is basically an incredibly versatile algorithm that uses NLP and feature programming to assess virtually any type of data source to support the interaction for which it has been designed.
Once they have been designed and deployed, they require very little or no human supervision. They learn by analysing data that they have access to and the interactions that they have in its use.
Machine Learning (ML)
This is how AI computer programs learn patterns from data. As it interacts within the confines for which it has been designed, it builds up an approach to align or adjust its future interactions. As stated earlier, once deployed it can do this without ongoing human supervision.
How Visa has been using this since 2019
An example is Visa and its use of ML in its AI platform called “Advanced Authorisation”, which monitors suspicious activity in real time. This initiative has been reported to have saved them close to $25 billion in fraud in 2019 alone.
It does this by considering key factors occurring within a transaction such as:
Whether the customer has purchased something from this store previously.
The transaction type being such things as online or in person if the customer used contactless or swiped their card.
Then it assesses the customers' buying patterns to ascertain if the purchase conflicts with their normal buying patterns, such as the amount being aligned to normal expenditure, it really gets interested if it’s for a high amount, time of day and where the geographical location of the place of purchase.
It also assesses all this against the wider set of data it has access to ie has the source has been identified for fraud previously, assessing tenure of operation based on the number of successful transactions with it in the past, and will get particularly interested with sources that mask their location or if the operate in known hot spots for fraud.
Their AI platform does all this in milliseconds, producing a score on the probability of transaction fraud. Taking into account that Visa is processing, on average, six million transactions an hour, imagine if this was done manually or if it was limited by a defined set of rules only. Then, think about the operating costs that would be needed to support this volume of transactions in people, technology and the sites to operate from.
Whilst speed is a huge benefit, the other benefit of ML is that it can comprehend unique or complex patterns that a human would be unable to do or would be lagging to an extent that the exposure would be deeply felt within the area of operation.
How does ML work?
Similar to how you learn through repetition and access to information AI is doing the same once it has been programmed for the service it is designed for.
Eventually, over time the computer will learn and become more accurate and certain about the outcomes it supports. The time is determined by the amount of accurate data it has access to and its interactions in the use of that to action the outcome.
The accuracy and certainty is built on its familiarity with the data it is shown and its application of this to the task(s) it performs. In the end, the goal is for the AI program to go beyond the original data sources and apply what it has learned to future interactions without human supervision or the need to create new rules in the case of fraud detection.
All of this is made possible by the algorithms; these act as a set of instructions that define how the AI will act, learn and adapt based on mathematical formulas, probability and likelihood it can compute, and the prediction parameters for the output.
The three types of ML Algorithms
We have outlined how algorithms support AI Machine learning, but these come in three different types. Which I go through as it will help guide you on the AI Business Strategies you undertake and the approach it will take to achieve them.
Supervised Learning
These algorithms learn from the data by being shown what the “correct” answers are, which are called “labels”. If we maintain the Visa example, it would have been built out on the successful outcomes of applying traditional rule-based methods for detection and treatment where the AI tool would have been provided with the correctly identified fraud intervention data. Other areas that this could be applied to are:
Predictions in Customer adoption, demand and churn rates
Diagnostics for Machine failure or screening for health concerns.
Forecasting staffing levels, costs of materials and building out to costs to serve
Unsupervised Learning
These algorithms also learn from the data but do not require the “correct” answers. Instead uses clusters or grouping and recommended problems (which can also be supervised). Examples of grouping could be applied to customer support requests, identification of duplication in records, and applied to the recommendation of products and services to customers. Other areas that you could consider the application of this are:
Document management
Search results
Support management
Customer behaviour or segmentation
Reinforced Learning
The most uncommon today are algorithms that learn through their actions rather than data. The program tries to take a certain action, and if it is correct for the given problem, it is rewarded. If you think about your early use of ChatGpt, you would have been prompted if the answer was correct or not. Whilst today this is not a model that most businesses would apply, as algorithms mature, and trust in them increases this type of learning may become more common.
In summary, for businesses today the application of Supervised and Unsupervised will be the most common used and the easiest to benefit from today, as long as you have large amounts of high-quality data. This should be an early consideration in your AI Business Strategy, with many organisations’ primary focus being the need to build its data sources to see the value of AI Machine Learning.
The other factor is training, as along with high-quality data a lack of training will provide unexpected or poor behaviour or impact the value of your AI program, which is why you need to feed it with the correct training data, and then use feature development for it to be successful. Which we cover in the next section.
So, if you starve your model of data, don’t show it what good looks like and don’t educate it on how to apply this to generating an outcome it will fail.
Feature Development
Machine Learning is limited to the data it has access to and how it has been programmed to interpret and use it. To set ML up for success, it needs to also understand the features that will support it in its decision-making. Features are a set of signals that support the training model learn and interpreting the data it has been provided; it tells the AI program the attributes that will support it in making its decisions. Data Scientists, Feature Engineers, or AI Engineers usually perform this.
During Covid, yeah remember that… people started to bake. I noticed heaps of people talking about sourdough and the complexities that existed in creating their starters and their successes and failures in this endeavour. The “features” talk about the factors, methods and tips that create the perfect sourdough starter and the options to get there. Features could be understanding things as basic as water choices where you need to avoid water that has high chlorine or chloramine as it can impact fermentation. This can be impacted by where you are in the world and the disinfectants that some cities use to clean tap water. It could be determined by testing your water with an over-the-counter kit, or using bottled spring water and how to tell if it is okay to use. Or don’t worry about buying the kit, and just try to use tap and then bottled water and see which performs best. This is why I never tried to make sourdough but shows the mental models that need to be applied in the generation of features. This was just for the water choices, think about all the other variables….
Feature engineering is a critical part of ML development and is especially important when it comes to Deep Learning.
Deep Learning (DL)
This subclass of Machine Learning uses neural networks behind the scenes. These networks try to simulate how the human brain works by running multiple interconnections. A key difference is that the AI is trained to uncover its own features within the data and its interactions to use for learning and to make optimal predictions and matures over time.
The challenge is in the time it may take to perform as expected, and that comes down to how it is programmed, the training provided, the data it has access to from the network and the inputs through its interactions that feed its maturity through the completion of tasks.
I think this is what most people may think AI is doing for many applications today; it is Machine Learning that is driving the outcomes. In areas where these have been used once, performing as expected, they are reported to be more accurate than traditional ML. It doesn’t mean that ML is inferior it just means that they can be more accurate, but it all comes down to the task and the factors we have discussed already. Just think about what Visa has achieved through its AI tool supported by ML.
Risk and Compliance Concerns
The challenge for most businesses will be the audibility of how the AI arrived at the decision-making process and the course of action that it took. Deep Learning is basically a black box, which in a financial regulatory landscape or health treatment application, I believe it would present too many challenges based on how we operate today in a controls-based ecosystem. As of today, I understand that you may never be 100% clear as to why it took the actions it did.
Data Computation Challenges
The next challenge will be the cost, yes, even though you take out the need for Data Scientists or like to create the features. Until recently, gaining access to powerful enough computers, we are talking about a shift from GPUs to what Google has created in TPUs to run a DL algorithm. My thoughts in the near term, this will still prove to be too expensive, and then the other consideration is if you use these cloud services, you need to think about where the data is stored geographically.
Conclusion and Next Steps
The reason why I share this is the ultimate goal for me is to apply what I know from creating business strategies and the industries that I have partnered with to provide my point of view and thought process. In the hope that it supports organisations in preparing their own AI Business Strategies. In addition, it’s my learning process to understand the opportunity and consider the impacts, benefits and confidence in realising that strategy.
So, there you have it, a dive into how AI uses Machine Learning and Deep Learning to achieve its outcome. My next blog will round out this foundational knowledge with Computer Vision and Robotics.
If you have any questions or could use my services, feel free to reach out.