11 Jan AI for Business: Deconstructing AI (Part 2) – Cognition
As we saw in the previous installment, AI cognition refers to the ability for a machine to “learn” and that machine learning (ML) is the commonly used term for the family of AI techniques typically used to “perform cognition”.
In this installment, we’ll more closely examine cognition to better understand its different capabilities and how these can be used. For brevity, we’ll use ML as shorthand for both the concept of cognition and its related AI techniques.
Read the AI for Business series:
PREDICTION, CLASSIFICATION, CLUSTERING, AND THE HARD STUFF
ML is an amazingly versatile collection of capabilities that, under the right conditions and with the right data, can predict or forecast the future, classify or sort data into (impossible for humans to discern) categories, discover hidden patterns within data, and solve highly complex challenges such as facial recognition and language translation. As the preceding sentence suggests, ML’s capabilities naturally fall into four categories: prediction, classification, clustering, and recognition.
Prediction
Prediction provides the ability for ML to predict or forecast an outcome after learning to identify patterns in historical data. Applications of AI prediction are found all around us, from weather forecasts to product demand predictions to forecasting a sales lead’s likelihood of closing.
Consider how an AI temperature forecasting model works. To train the model, we provide it with historical weather observations for a specific location, preferably going back decades. In addition to the location’s historical temperature readings, contributing weather data could include other inputs such as day-of-year, wind speed, wind direction, cloud cover, precipitation, prior hour temperature, etc. All of these data points are fed into the model. As the model processes each of the tens or hundreds of thousands of historical observations, it “learns”, becoming more and more accurate with each iteration. Once training is complete and validated, we can then feed it a new set of input data (observations) and expect the model to generate an accurate prediction of the temperature.
If you take this example of predicting a single temperature reading at face value, it’s easy to see the unimaginable complexity of forecasting a region’s weather for tomorrow (not to mention the next ten days), yet we take this for granted every time we check our phone’s weather app.
Model – In the AI field, you’ll see the terms model, algorithm, and application used interchangeably. These typically refer to the core technology component that provides “AI” to a solution. In the above example, the temperature forecasting model refers to a ML technique (probably some form of a neural network).
Prediction is frequently used in combination with other AI techniques to solve larger, more complex AI use cases such as autonomous vehicle navigation (e.g., “learning” that a yellow light indicates that a red light will quickly follow), sports betting, and day trading (from an AI perspective, sports betting and day trading are remarkably similar – a philosophical peculiarity and perhaps the basis for a future blog installment).
Classification
While classification is closely related to prediction, its objective is different. Rather than predicting or forecasting a discrete value (e.g., a temperature), a classification model sorts or classifies data into one of several predetermined categories. Two examples should make this distinction clear.
We are all inundated with email spam, to the point that we could spend an inordinate amount of time each day just wading through it. Fortunately, AI provides a solution for much of our pain. A spam classification model (commonly known as a spam filter) uses historical information (email that has already been classified as spam or not spam) as training data, and then applies the trained model to assign newly received messages to one of these two categories. While these models occasionally slip up, I find them remarkably accurate and trustworthy.
Credit card companies use similar techniques to identify potentially fraudulent transactions, giving customers the opportunity to decline charges before they are processed. While criminals constantly look for and sometimes find workarounds to the classification models, these are quickly identified, and the models are updated to deal with the new threats. Imagine trying to navigate the financial world without classification.
Supervised Learning – Prediction and classification solutions use a technique called supervised learning to train a model. In this scenario, the model uses historical data to “learn” like an infant does, through repeated observation and imitation. And, as new observations are presented to the model, it continues to improve its “understanding” over time.
Clustering
Clustering differs from prediction and classification in that it doesn’t learn from historical data. Instead, clustering is used to examine a collection of seemingly unrelated data to discover hidden patterns or data groupings without the need for human intervention[1].
While applications of clustering aren’t as common as those for prediction and classification, for some use cases it can be powerful. In marketing, accurate personalization and targeting are vital. Clustering provides the ability to identify the specific characteristics of a customer and then share marketing campaigns that have been successful with similar people. Interestingly, Palantir Technologies, a leading technology supplier in the defense intelligence domain, holds at least one patent[2] focused on clustering and touts that, “With one coherent model, users can discover previously unseen links across their entire universe of data.[3]” Coincidence?
Recognition
Our final category deals with recognition. Recognition is an advanced form of machine learning that uses deep neural networks to solve highly complex problems such as facial recognition and automated language translation. In one sense, recognition is just a refinement of unsupervised learning, but given its power and complexity, it’s worth calling out separately.
In the past, recognition technology was highly specialized, very expensive, and difficult to implement. But with the advent of pre-configured models and common source AI toolkits such as TensorFlow, these techniques have democratized ML, moving it into the mainstream and making recognition-based use cases accessible to the broader business community. Today, it’s often difficult to detect the boundary between prediction and recognition. An example that focuses on object recognition (just one of several recognition categories) should illustrate this breakthrough.
We recently received a client inquiry about developing a mobile solution that would take a customer-provided image of an industrial valve and then identify it in the client’s product catalog, returning that product’s technical specifications, pricing, and related information. While much of the solution is somewhat mundane systems development, at its core there is the need to recognize an object (the valve) in an image. Using the models and toolkits described above, this functionality requires less than 20 lines of code.
Recognition’s power and potential benefits to businesses cannot be overemphasized. If you get nothing else from this blog series, consider this: With a little effort and imagination, every organization has the ability to leverage recognition to uncover value, sometimes enormous value, in the information that they already have or can easily obtain. Technology is no longer the limiting factor.
FINAL THOUGHTS
Before we shift our focus to AI’s other high-level capabilities, it’s worth considering machine learning’s importance to the AI domain. While language, vision, and automation are important and valuable concepts in the human and AI worlds, they pale in comparison with the ability to think (cognate). Not surprisingly, in future installments we’ll see numerous examples of AI use cases/solutions where ML is at the core.
Now that we have a handle on cognition and machine learning, in our next installment, we’ll focus on AI’s other high-level capabilities: language, vision, and automation.
References:
[1] As you might deduce, this technique is known as unsupervised learning.
STAY TUNED
Stay tuned for Glen’s next blog in the AI for Business series to learn about another one of AI’s core capabilities.
STAY TUNED
Stay tuned for Glen’s next blog in the AI for Business series to learn about another one of AI’s core capabilities.
