Since Gartner sees most AI trends are just about reaching the “plateau of disappointment” this year, I thought it would be a good opportunity to give you an overview over what the different genres of AI are and why they made people realize they’re aren’t magic bullets (but Gartner cautions not to write AI off either).
Machine Learning AI
Machine Learning (ML) is a cutting edge form of AI. It’s the type of AI that is generally credited with driving your Tesla down the highway and promises to yield the biggest impact for everyday workplaces one day.
Machine Learning is about finding patterns in big data where commonplace statistical analysis doesn’t see any, and taking such pattern to predict results without much human interpretation.
ML however requires a few key ingredients to become effective:
ML requires lots of data
To teach the AI new tricks it requires bucket loads of data into its model input to make reliable output scoring. Tesla for example has deployed an auto steering feature to its cars which simultaneously sends home all the data points it collects, interventions by the driver, successful evasions, false alarms, etc. to learn from the mistakes and gradually sharpen the senses. A great way to produce a lot of input is through sensors: Either your hardware has built-in ones like radar, cameras, steering wheel, etc. (if it’s a car) – or you lean on IoT (Internet of Things). Bluetooth beacons, health trackers, smart home sensors, public databases, etc. are just a small fraction of the ever growing number of internet-connected sensors that can generate much data (too much for any normal human to process)
ML needs good discovery
To make sense of your data and cut through the noise Machine Learning puts algorithms at work that can sort, slice and translate a data chaos into comprehensible insights. (If you want to weird out your colleagues listen to sound of different sorting algorithms at work: https://www.youtube.com/watch?v=kPRA0W1kECg)
There are two ways for algorithms to learn about the data, unsupervised or supervised.
- Unsupervised ones deal with figures and raw data only, so there are no descriptive labels or dependant variables established. The aim for the algorithm is to find an intrinsic structure where humans didn’t think there would be one. This is useful for gaining new insights into market segmentation, correlations, outliers, etc.
- On the other hand, supervised algorithms have knowledge about relations between different data sets through labels and variables and use their power primarily to extrapolate and predict future data. This might come in handy for anything from climate change models, predictive analytics, content recommendations, etc.
Challenges for widespread business use
ML has proven its effectiveness under lab conditions as well as having a huge commercial track record with companies like Google or Tesla. So it’s surprising that it has nevertheless found it hard to make its way into bread and butter business applications, like CRMs, Analytics or Helpdesk software.
Even if such implementations don’t fail due missing big data or good discovery, then oftentimes it turns out, that on smaller scale deployments AI tends to be a quite an expensive tool to develop and maintain. Meanwhile, the benefits for users seldom push up productivity up enough so that it can recover these costs.
Of course app providers could create AIs that use data from multiple customers, but there are again some complications. Compliance or user privacy for one, and, indeed, customers don’t want their potentially commercially sensitive data to be fed into AIs they don’t control, just to pay extra to get the benefit of them.
Deep Learning AI
If Machine Learning is cutting-edge, then Deep Learning is the bleeding edge. It’s the kind of AI that you would send to Jeopardy! It combines big data and analytics with unsupervised algorithms. The applications usually center around gigantic unlabeled data sets that need structuring into inter-connected clusters, inspired by nothing less than neural networks in our brains – therefore fittingly called artificial neural networks.
Deep Learning is the basis for many modern speech and image recognition approaches and benefits from a much higher accuracy over time than non-learning approaches offered in the past.
It is hoped that in the future Deep Learning AIs can autonomously answer customer enquiries and fulfills orders over chats or emails. Or they could assist marketing in suggesting new products and specifications based on their enormous data pool. Or maybe they can one day be Omni-present assistants at the workplace that entirely blur the line between robots and humans.
AIs live and improve through the scale of data that is thrown at them, which means that we see better AIs over time but also that their development will center around those organizations that can tap into the largest sets of data.
Cognitive computing is the most popular form of AI and responsible for all interactions that are meant to feel human-like. Cognitive AI must be able to handle complexity and ambiguity with ease, while also continuously learn from its experience through data mining, NLP (natural language processing or understanding) and intelligence automation.
NLP, being the most prolific application, allows a chatbot to discern between entities, intent, timeframes and even locations. This allows chatbots to understand an entire sentence or question, whereas a regular search engine would just accept the query in its entirety, and thus would be unlikely to yield any useful results.
For example, “all documents by Tim in SharePoint this week” would be translated seamlessly into the search query:
filetype:(doc|ppt|txt) author:(Tim) location:(Intranet) upload_date:(Monday–Today)
The chatbot would arrive at this query, despite a certain degree of ambiguity in the verbiage, through a combination of cognitive AI, pattern matching and if-else logic.
Interesting for businesses
Cognitive AI is a low hanging fruit for enhancing the user experience of business applications. Since it’s not as much of a black box, it is easier to control the experience and detect when best to call for human backup. For example a chatbot (the most popular form of cognitive AI) is basically just a machine that makes “best bets” at the user’s intention, and if not successful, can quickly page in a real support agent – thus helping to reduce their load overall.
AI for internal business use
AI for business means applying algorithms, such as Machine Learning or Deep Learning, to improve specific operational systems concerned with running a business. Examples include better ways to qualify sales leads, improving ad spend for online marketing or using knowledge graphs to surface contextual information, etc.
This is different to enterprises selling AI products. The latter would encompass organizations like Tesla using AI to teach cars how to drive or IBM Watson using AI for X-ray interpretation. While those applications are AI, they aren’t made for running a back office more efficiently – and that’s a key difference.
Budgets for bread-and-butter line of business applications are traditionally pretty tight, therefore Machine Learning is seldom found in them. However, with cognitive AI, such as Natural Language Processing, there are more feasible options to enhance the user experience of a product.
In addition, there is a growing area where formerly consumer AI offerings make their way into enterprises, for example Amazon Alexa branching off an office version with Alexa for Business.