When people hear I work with artificial intelligence, they often imagine that I create software that does magical things. Their minds go to humanistic androids or programs that predict distant future events. But the reality of AI is much closer to home; it enables many tasks that seem mundane, and even when it can feel like magic, that effect is really coming from a combination of techniques and programs that have a lot in common with the more familiar category of analytics. When we talk about analytics, people usually find it much clearer that we mean interpreting some set (or sets) of data. AI is also very much about interpreting data, but it goes beyond a direct and immediate calculation on the data: It generalizes, predicts, or interprets information in a way that we might associate with a human, typically because it involves some kind of approximate judgment or draws on some broader knowledge.
In analytics, we talk about descriptive analytics, which finds facts about the data such as correlations or outliers, and predictive analytics, which estimates unknown values such as the next value in a sequence, or a category that isn’t represented in a data set. For example, suppose we have a data set that represents information about visitors to a museum for each day: the number of visitors overall, the number of visitors at each exhibit, and some demographics about those visitors. Descriptive analytics might find a correlation between the day of the week and the number of visitors (perhaps more people visit on the weekends), or between visiting one exhibition and another (perhaps the same people are interested in both). With predictive analytics, a system might predict how many people will visit next Saturday, or might “predict” an unknown value, such as how many people visited one particular exhibition on a day that wasn’t measured. In these predictive cases, the system is recognizing patterns in the data and generalizing them, applying them to new cases, and using that to reach an estimate or hypothesis about something that is not definitively known. This brings things into the realm of supervised machine learning, which is the most common—although not the only—type of machine learning, a common building block of AI. Machine learning means learning and generalizing from data.
There are two types of machine learning: unsupervised and supervised. Unsupervised machine learning operates without training to examine data and identify patterns within it. It can do things like find clusters of similar data, or find anomalous instances that look different from almost everything else. From our hypothetical museum visitor data set, for instance, it might find clusters of weekend days with high traffic and heavy concentrations in certain exhibitions, or of typical weekdays with lower traffic and heavy concentrations in other exhibitions, or perhaps another set of weekdays with higher traffic and yet another set of exhibition concentrations. (This last set of weekdays might turn out to be school holidays with more family participation than the other days.) But while unsupervised machine learning can find these clusters; it cannot tell us what they mean or estimate something new.
For that, you need supervised machine learning, which studies a training data set where the right answers are already provided and learns patterns behind them. Using these patterns, it can then look at new data sets and predict answers that are not provided. Similar to predictive analytics in our hypothetical example, it can look at the number of visitors for each day in the training set, then predict a value like how many visitors there will be next Saturday. Or, from a training set where the number of visitors per day to a given exhibition is provided, it can estimate how many visitors there were to it on a day the number wasn’t recorded. Many applications of machine learning are much more complex in the values they consider, of course. Consider, for example, machine vision when it uses machine learning to identify specific types of objects in images. Rather than a relatively small data set, like days and attendance numbers, machine vision considers individual pixels and/or features derived from those pixels and their combinations (edges, transitions, and the like), which creates an enormous number of values, and it looks for patterns in subtle variations among them. This is where machine learning goes beyond the capabilities of conventional predictive analytics, though it is based on exactly the same principles.Skip over related stories to continue reading article
Machine learning is part of AI, but it is far from the only part. AI also includes systems that make meaningful interpretations of data based on knowledge, which may be expressed in rules or in resources that are developed for the purpose, or both. This means that you’re often using AI when you may not realize it: when a machine recognizes your speech and takes some action in response, for instance, or even when you scan a document and get searchable text from OCR, rather than just an image. It also means that we can use AI to build systems that do very subtle and complex things, some of which feel more like the “magic” that people associate with AI. These systems generally combine a variety of AI pieces with significant engineering to execute them. Consider, for instance, a realistic avatar that converses with a user on some topic. There is AI in recognizing the words that the human speaks. There is AI in interpreting those words, to identify what meaning the avatar needs to respond to. There is AI in working with resource information about the subject matter domain, for the avatar to be able to respond meaningfully within that domain. There is AI in generating the avatar’s response. There is AI in turning the avatar’s response into something that sounds like natural speech. There is AI in creating the avatar’s facial movements and expressions that accompany the sounds and content of the response. And all of these pieces—and more—need to work together for a single effect.
So, what does this mean if you’re looking at ways to use AI? It means the range of things that it can bring to the table is extremely broad, though its success depends heavily on having data available that is suitable for the purpose. For supervised machine learning, this includes having labeled data, or ground truth, the data where the “right answers” are provided, so that the system can learn how to estimate those right answers on new data that does not provide them. Beyond this, with any AI, the data that is used for developing, training, and tuning the system needs to be representative of the data that will be encountered when the system is deployed. This is often more challenging than it sounds: for example, suppose we were to take only our most recent data, perhaps the last two months, to train our hypothetical system with museum visit information. If the patterns of visits to the museum are highly seasonal, our system will learn only about the current season, but it won’t know that. As the season changes, the predictions will be less accurate. This means that it’s important both to select data that represents the range of variations you expect, and to check regularly on how the system is doing. You may find that there are variations you didn’t expect to matter that do, or you may simply find that your data changes over time. Data isn’t the only thing you need, of course; you also need a good problem that makes sense for AI, i.e., one that requires some interpretation and judgment and follows some discernable patterns that the machine can discover in order to learn that interpretation and judgment—or that is amenable to the development of explicit knowledge and rules in lieu of machine learning. So, to summarize what makes up a good use case for AI, it has:
- A task that is well-defined; requires some judgment, interpretation or estimation; and has likely indicators for that judgment, interpretation, or estimation.
- Data that is available, clean (which is to say accurate, represented consistently, and reasonably complete), and representative of the data that is likely to be encountered in production, with ground truth labels if the task is one for supervised machine learning
- Use that fits into a workflow so that people can find a way to take advantage of it, and that has value to your organization’s mission—of course!
The use you pursue can be small and simple, or large and complex. I find that in general it is best to start small and add incrementally. Most uses of technology are more complex than they appear on the surface, and this is at least as true in AI as everywhere else. Starting small also gives you an opportunity to experiment, to examine the strength of your data, and to course correct if needed before you expand to bigger uses. A different way to get started, of course, is to use an AI tool that is already built and applicable to the data you need, such as sentiment analysis on social media, which is available in many forms.
With all the above in mind, here are a few sample ideas, both small and large, about where AI might provide value for museums’ needs.
- Sentiment analysis on social media postings that mention a given museum, or perhaps a given exhibition. This may be an easy place to start with AI, using existing sentiment analysis tools on public social media and selecting content that mentions the museum or exhibition of interest, to find whether people’s postings on social media that mention the exhibition or museum tend to be positive or negative, and whether there are identifiable trends in this association. Even this straightforward use can build in complexity over time, adding things like looking for whether the sentiment is directed at the museum or exhibition, or simply expressed in conjunction with it.
- Interactive assistance for a lightly personalized tour based on an individual’s interests and questions. This could use a combination of existing AI pieces trained and configured for the particular purpose. It could include speech-to-text, text-to-speech, and language processing AI for dialogues and conversations, to create a conversational interaction that might share additional information on topics where a visitor expresses questions, recommendations for what to view next based on the visitor’s expressed interests, and more.
- Cluster similar items to assist with managing collections and archives. AI could cluster objects in collections based on available metadata or textual descriptions, or even use image processing to cluster based on recognized visual similarities.
- Deeply personalize a visitor’s experience with AI interactions and recommendations based on their past responses to elements of visits. This kind of application could combine elements of all of the above ideas, and try to interpret the individual’s point of view. In addition to making recommendations of what to visit based on past responses, it might also introduce novelty with suggestions that are not similar to what an individual has enjoyed before, and perhaps even apply machine learning to recognize how frequently, or under which circumstances, it is most useful to inject these kinds of suggestions in order to expand the horizons of the visitor.
More generally, any task that goes a bit beyond the mechanical but is currently highly demanding for people—because of time, repetitiveness, or the number of people required—is usually something where the right AI system can help. AI is not magic, and it does not replace human judgment, but if you have the right data, it can help with almost anything.
About the author:
Kristen Summers is the Technical Delivery Leader for Data and AI implementations in North American Government at IBM. She has been working in Artificial Intelligence and Human Language Technologies, with a primary focus on applied research, for the past 15 years. Her experience in this area includes leading projects on question answering, entity recognition, entity co-reference, machine translation, Optical Character Recognition (OCR), and other related topics. Before joining IBM Watson, she was the Technical Director in the Knowledge and Information Management Division Group at CACI. She holds a PhD in Computer Science from Cornell University and a BA in Computer Science and English from Amherst College.