Few would argue that artificial intelligence is completely changing the business landscape across most industries, while pushing these industries to gather greater volumes of data than
ever before. An emerging focus area of artificial intelligence, called “cognitive computing”, shows particular promise for applications where a computer shouldn’t necessarily replace the decision-making of humans but should instead interact with humans in a way that augments and supplements their decisions. Let’s dive in to what this means.
Cognitive Computing vs Traditional AI/Machine Learning Approaches
Cognitive computing is all about interaction with humans, often providing insights that are discovered or “learned” autonomously (we call these insights “unsupervised” – more on that later) which can then be used by humans to make better decisions. These could be broad-reaching business decisions, such as choosing an emerging market to increase portfolio investment; or decisions related to safety, such as locking down accounts that appear fraudulent; or even life-saving decisions, such as choosing the right treatment for a cancer patient.
Traditional approaches to machine learning and artificial intelligence (which do not involve cognitive computing) learn from labeled training data. Humans must manually add these labels and then test whether the model can use them to accurately predict the labels on new unlabeled data. We call this type of learning “supervised”, because humans are supplementing the data provided to the machine, rather than the machine supplementing the data a human needs to make a decision. In the case of fraud analysis, humans would supervise the model by manually labeling transactions as either fraudulent/nonfraudulent, and then test whether the model can accurately predict the label for new transactions. A recent, example of this in the medical field involved researchers labeling thousands of female breast biopsy slides with benign and malignant cancerous tissue to successfully train a neural network to recognize them on new patient biopsies. While this traditional approach may help detect cancer earlier and lead to better treatment, it is also tedious. New cognitive computing approaches are proving to be even more helpful because they don’t require labeled data.
“Unsupervised” Neural Networks Shouldn’t Make Decisions
Cognitive computing models are commonly implemented using neural networks because such systems learn to perform tasks by considering examples- similar to the way the human brain would learn to perform a task. Complex data sets are fed in, and these models must automatically learn certain characteristics of the data which may be too complex, tedious, or costly for humans to do themselves. Without labels the neural network discovers information about the underlying structure of the data. We call this learning “unsupervised”.
Going back to our fraud detection example, it might look like this- the neural network learns the underlying structure of data that represents “regular” behavior in a dataset of transactions. When abnormal behavior occurs, the neural network has no ability to recognize it. Since these abnormal transactions are not understood by the neural network, it automatically labels them as fraud (or suspicious to be reviewed). In the case of biopsy tissue analysis, a neural network would learn the “structure” of healthy tissue slides, so when it sees tissue it doesn’t recognize, it automatically labels that tissue as cancerous (or irregular to be reviewed).
As you can imagine, this approach is not foolproof. Accuracy of the labels, predictions, or classifications that a neural network can assign to data in an unsupervised way is often lower than supervised methods- prompting the need for human “review” or decision-making. While removing the need for tedious human-labeling of thousands (or even millions) of datapoints is enticing with unsupervised approaches, the outcome of their analysis can often only be a suggestion to the user, who must then make a decision. This is where cognitive computing comes in.
Cognitive Computing Brings It All Together
Let’s look at a new example of cognitive computing, using anomaly detection in physical spaces. Consider a large oil refinery, where hard hats are worn around most spaces of the site. Cameras placed around the refinery could use machine vision models to detect whether people within each space are wearing hardhats, and then feed that information into a neural network which learns whether each space requires a hardhat or not. Then, when people are recognized not wear hardhats in spaces that generally require them, the neural network alerts users, who decide how to ensure the regulations are met within those spaces. Once again, the key here is that the neural network does not make a decision, but simply suggests to the end user that workers should probably be wearing hard hats in a certain area. This is what cognitive computing is all about: using unsupervised machine learning techniques which supplement the human decision-making process.