“A Cognitive Reasoning Bot and Machine Learning Bot Walk into a Bar” – by Matt Buskell
“A Cognitive Reasoning Bot and Machine Learning Bot Walk into a Bar” – by Matt Buskell
September 14, 2016
![](https://eu-images.contentstack.com/v3/assets/blt6b0f74e5591baa03/blt2b43dffc16f1b042/6319624d3d195f4e71fe1a2a/Chatbots.jpg?width=850&auto=webp&quality=95&format=jpg&disable=upscale)
This week, AI Business expert guest writer Matt Buskell, who is also Head of Sales and Business Development at Rainbird, takes an alternative look at two key branches of artificial intelligence that are finding transformative applications in business.
“What’s the difference between Cognitive Reasoning and Machine Learning?” I get asked this question a lot, so how about we have some fun and use a story to illustrate…
A Machine Learning Bot and Cognitive Reasoning Bot went out one evening to a bar. Having never been to a bar before both decided to teach themselves what drink to order.
Machine Learning reviewed all the data it could find, digesting lots of articles and pictures of drinks. This took some time because it had to find many articles on drinks. Eventually though it found the features of what it thought makes a good drink.
Cognitive Reasoning didn’t read anything. It called a couple of friends who explained what they considered were the important factors in choosing a suitable drink. This also took some time because the friends had to explain the basis of their decision making. “It depends on who you are with, the weather, the setting and what kind of evening you want.”
Machine Learning and Cognitive Reasoning finally meet at the bar, and both stare at the available drinks menu. The bar tender comes across and asks them what they would like to drink.
Machine Learning asks for a beer, with plenty of ice in, ice being the feature it believed was the best indicator of a good drink based on lots of pictures it had seen. The bar tender asked, “Why ice with beer?” and Machine Learning replies “I don’t know why, but I’m sure it’ll make a great drink.”
Taking a sip it promptly spat the beer out in disgust. You see, no one had told it bots don’t like cold bitter drinks, or that ice ruins beer. Machine Learning makes a note to not drink beer with ice again.
Laughing at Machine Learning’s reaction, Cognitive Reasoning asked the bar tender for a “mojito”. Confused by the choice the bar tender asks “Why a mojito in the middle of winter when it’s raining and cold outside?”. Cognitive Reasoning explained is has been taught that mojitos were good in warm conditions and the bar was especially warm. It also explained that as it had an umbrella and it intended to sit on the bar’s sandy deck outside.
When this didn’t work out so well, Cognitive Reasoning decided to reduce its certainty about the factors that led it to this conclusion so it didn’t happen again.
So there are a few morals to this sorry tale.
Firstly, AI Bots are not that smart unless you expend effort, although both have the potential to learn from training or mistakes. Machine Learning relies on data which must be extensively tagged by a human and Cognitive Reasoning needs to know the human logic that, in aggregate, might lead to a good outcome.
Secondly, the two technologies are really good for different things, and each has strengths and weaknesses.
Machine Learning learns on data. In our example, this may comprise lots of data on drink ingredients, locations, properties, individuals, ethnicities, ages but also could include less spurious data like skirt lengths and hair colour and any other data you provide it. However, this data needs to be tagged, and any outcomes of previous decisions need to be marked up as “good” or “bad”. Building this training set is exceptionally time-consuming and requires data scientists to achieve a result. Until you get it entirely trained, it will continue to return nonsense, so in that way, building a machine learning model can feel a little like climbing Everest: it is an amazing view from the top but you might die on the way up.
Cognitive Reasoning is built from a model of existing human knowledge. This also takes time to make because you need to design a symbolic model. You can encode the individual rules in isolation and the cognitive reasoning engine can stitch it all together to make a conclusion. It is possible with tools like Rainbird to also encode uncertainty which enables the Cognitive Reasoning engine to draw conclusions in the absence of data or where the accuracy of data may be uncertain. It can carry forward any uncertainty in the conclusion. Also, because it is based on a model of knowledge, it can explain clearly why a particular decision was made.
We should be clear at this point that neither Cognitive Reasoning nor Machine Learning is the same a Natural Language Processing. A simplest explanation to differentiate these technologies is to draw a parallel to humans. So Natural Language Processing (which is sometimes used alongside speech recognition) might be the eyes and ears. You might consider Machine Learning to be the unconscious brain. Machine Learning looks at correlations between large amounts of data and therefore may consider something to be instinctively correct but cannot explain why. Cognitive Reasoning may be comparable with our “deliberate” brain where conscious decisions are made by applying well understood, learnt logic to the data available. Because a decision is made in this way, the rationale for that decision can be explained.
So which is best, Machine Learning or Cognitive Reasoning?
That depends on what you are trying to achieve. The latest thinking from academia in fact is that some of the best possible outcomes can be reached by using both approaches together.
"Deep learning is a very powerful technology that is having a huge impact in both industry and academia. But I believe it can be made even more powerful if conjoined with search and inference techniques from symbolic AI that have been under continual refinement for decades.”
Professor Murray Shanahan
At Rainbird we are currently involved in a fun project where both technologies are being used together. We are creating a system that simulates a music artist so their fans can feel like they are interacting with them on social media. Machine Learning is being used to create a model of the “voice of the artist” i.e. the tone of their language and the slang they use. Cognitive Reasoning is being used to interview the artist and create a model of their music knowledge. The result is an interactive and insightful conversation in the style of the artist.
Why is this useful? Well, in the past eight months over 70% of music recommendations have moved from the public web to private discussions on social media. This shift leaves Digital Marketing Managers blind. So while fun, this sort of interaction with the fans using a Bot may also become critical to driving music sales, while operating at a scale impossible for the artist to achieve personally.
So when should I use Machine Learning vs Cognitive Reasoning?
Generally speaking, machine learning is an analysis technology capable of finding patterns in vast amounts of data that are not immediately obvious to humans. So it’s great in data analysis situations and also useful in natural language applications where we are trying to determine things like emotions and intent. Remember, machine learning cannot articulate the rationale for a decision, so it requires an expert to interpret the output, especially in regulated industries.
Top 3 Simple Use Cases to start to use Machine Learning
Classification: By analysing a large number of classified examples of data, machine learning techniques can be used to predict the likely classification of a new, unseen piece of data.These techniques can be used to perform tasks such as image recognition or document classification. For example, by providing the algorithm with hundreds of photographs and a label for each indicating its contents, machine learning can (after sufficient training) describe the contents of a new photograph. This is great as a search technology.
Future prediction from prior data: Machine learning techniques are great at finding patterns in data that humans may not be able to see. This can be used to make predictions or recommendations about what a good course of action might be based on prior experience.We’re all used to movie or other product recommendations being made by Amazon or Netflix. These recommendations are coming from a machine learning algorithm that’s looking at the data collected from many hundreds of thousands of previous consumer interactions to make a prediction of what you might like. The rationale behind these predictions can’t be explained, so while such techniques can be powerful, predictions, recommendations or diagnoses often have to be verified by an expert, and this can be expensive.
Finding intent from natural language: Understanding what humans mean when they interact with a computer in natural language is a difficult task. Machine learning techniques can be used to find the intent behind what a human user is asking, matching that intent up to a pre-trained set of previous interactions. So for instance a user might ask “What time does your London branch open?”, or “When can I get into your London branch?” or any other number of variants. Because the question can be posed in an almost infinite number of ways, machine learning can be used to recognise the intent behind the question based on a training set, and direct the user down the appropriate path.
Cognitive Reasoning works well when a judgement needs to incorporate existing human knowledge and understanding, not just new patterns derived from data. This method works well when it is important to be able to articulate why a decision has been made, necessary in any regulated industry such as financial services, insurance or medical.
Top 3 Use Cases to start to use Cognitive Reasoning
SmartForms: This is when you take a complex form and you replace it with a reasoned conversation. Because Cognitive Reasoning is ruthlessly efficient at asking the necessary questions to get to a goal, it can dramatically reduce the number of questions users need to answer. If you have existing data, this can be taken into account making the interaction even more efficient. Consider how many government forms are retuned because they are incorrectly completed or how many credit card or insurance applications are abandoned because customers just lose interest in completing the form.
Directed Advice: Let’s say you are thinking of changing bank accounts or looking for a place to go on holiday or wondering which tyres to buy for your car. You have a specific goal in mind but you need the advice from an expert to help you make the right decision. In these situations, the knowledge of an IFA, travel advisor or mechanic can be encoded into a cognitive reasoning system and combined with external product data to consult with the customer and direct them to the best outcome for their specific situation. What’s important here is that cognitive reasoning can explain why it gave the specific advice, and because it is based on human knowledge and comes with an audit trail for every decision, it can meet the requirements of regulators.
Triage: You may be familiar with this term from a medical perspective. However the concept of triaging occurs everywhere, from insurance to legal. We are working with a leading law firm who are using Rainbird to analyse their customer base to identify if new legislation might impact their customers. In this use case the customers receive a link to an integrative Bot that is powered by Rainbird. This Bot has access to existing client data, and may ask questions until it has qualified them into one of five risk categories. Based on the triaged category, they are then provided with recommendations based on that profile.
What’s important here is the system is figuring out the fastest path to get to a risk category for the customer, but also provides a plain English rationale describing its reasoning to both the customer and the lawyer.
Conclusion
There are some great applications for both Cognitive Reasoning and Machine Learning technologies and together you can make some really interesting applications and Bots.
Contact [email protected] to continue this discussion. Or for more information about how we can help you better understand your customers and drive innovative and efficient solutions that both transform your business and keep the regulator off your back.
For the latest news and conversations about AI in business, follow us on Twitter, join our community on LinkedIn and like us on Facebook
About the Author(s)
You May Also Like