The Simplest Way to Explain Hybrid Intelligence (Machine Learning + Human Understanding) for Consumer Insights

What I learnt breaking down data analysis that involves machine learning and social sciences for a TED talk and how this helped me find our purpose.

(If you wish to get the article in your inbox, sign up here.)

Jump on the video instead: Hybrid intelligence: rise of the human side of data in the A.I. era.

 

Article Index

The challenge.

  • Introducing machine learning for consumer insights to a non-expert audience
  • Solution: make it simple and relevant for everybody while going beyond buzzwords

The Simplest Way to Explain Hybrid Intelligence (Machine Learning + Human Understanding) for Consumer Insights

  • How machines learn: the simplest way to explain Machine Learning
  • The limits of machines and the limits of humans
  • The future is hybrid: machine learning + human understanding
  • Conclusion. Benefits of using hybrid intelligence to get actionable consumer insights

Wrapping up. Fear is not a solution, but you can leverage it to create impact.

 

 

The challenge.

Introducing machine learning for consumer insights to a non-expert audience

When they asked me to give a TED talk I almost panicked.

I mean, it’s amazing, don’t get me wrong, I love TED talks. I love them because they make accessible complex topics out of my expertise and create inspiration across industries and information silos.

But, it’s a huge challenge I realised.

Debunking your own work and knowledge for people that have never even heard of it while being accurate and scientific about it is extremely hard. In 10 minutes or so? Almost impossible. :O

To make it even harder is the question: would people care? Very few people outside of market research talk about market research. People outside the industry don’t even know what consumer insights are.

So no, they wouldn’t — I thought.

But, we live in a time of amazing opportunities and great challenges and

  1. Most of us have heard of how machines are learning to do our jobs. Being through sci-fi stories, news or simply overhearing a chat about technology.
  2. Some have already faced this kind of disruption and have had to learn how to upgrade themselves and their companies to be competitive in the market.
  3. Some have even seen their own job becoming obsolete.

Solution: make it simple and relevant for everybody while going beyond buzzwords

So yes, there’s a reason why AI and ML are buzzwords. In the era of artificial intelligence, machine learning and automation, we increasingly ask ourselves if machines will end up replacing us.

That is the core of the curiosity, passion and fear involved in all the conversations about AI and ML. The real fuel.

So, once I got that after hundreds of conversations back and forth between our CTO, some industry experts and the “normal people” (with zero knowledge on tech and implications of AI/ML) plus the TEDx team, I felt I’d unlocked the superpower of TED. This story IS relevant even for people outside of the industry. Yay!

Now, the second part of the challenge: what value am I providing to the audience?

I need to create inspiration, I need to get them to learn something that they can apply to their own job, industry and path, I said to myself.

I need to provide them with a solution.

So, what if there was a solution to our fear of machines taking over? What if the future was a man-machine hybrid?

How can machine learning augment human minds and give birth to a hybrid intelligence, as opposed to just artificial intelligence?

Boom.

The next step was to make it simple and tangible. So we sat down with my partner, expert of hybrids systems, and explained step by step how a data analysis complex process actually works — how we leverage machines AND humans power for marketing.

We took a philosophical issue (quoting the Humanism cultural movement) to the very simplest way to explain tech and data analysis: Hybrid Intelligence (Machine Learning + Human Understanding).

The Simplest Way to Explain Hybrid Intelligence (Machine Learning + Human Understanding) for Consumer Insights

(in case you prefer to save the article for later, you can download it here. If you want to watch the TED talk instead, here is the link to Hybrid intelligence: rise of the human side of data in the A.I. era)

How machines learn: the simplest way to explain Machine Learning

So what is this Machine Learning?

ML is basically machines imitating and adapting human-like behaviour.

How exactly?

Let’s start with a quiz.

How did you come to 81???

That’s exactly the kind of behaviour that we, humans, are trying to teach machines today. We’re trying to teach them to “learn from experience”.

More specifically, machine learning algorithms use computational methods to

  • Learn information directly from the data
  • Find natural patterns within the data
  • Get insights
  • Adaptively improve their performance as the number of samples available for learning increases

In the talk, I couldn’t go into the types of Machine Learning techniques for time reasons, but I would like to briefly mention them here. Just so you have a tiny, super simple digest of machine learning techniques.

You can find Machine Learning in two forms:

  1. Supervised Learning
  2. Unsupervised Learning

1) Supervised Learning

Supervised Learning is when you train the machine with both input data and output data. The machine learns by finding patterns and is able to predict by itself.

All Supervised Learning techniques area form of either Classification or Regression.

  • Classification is used for predicting discrete responses. An everyday example you came across is whether an email is spam or genuine.
  • Regression is used for predicting continuous responses. Some common examples: trends in stock market prices, weather forecast, etc.

2) Unsupervised Learning

Unsupervised Learning finds patterns based only on input data. This machine learning technique is useful when you’re not quite sure what to look for. It is in fact often used for Exploratory Analysis of raw data.

Most Unsupervised Learning techniques are a form of Cluster Analysis.

In Cluster Analysis, you group data items that have some measure of similarity based on characteristic values. At the end what you will have is a set of different groups.

(example from Blaze of Inspiration — where we tracked millions of posts from communities, groups, tribes, influencers, brands, organizations, events, official and unofficial channels that constitute the ecosystem of Burning Man and transformational festivals)

Real-life examples of machine learning

If you’re a user of Spotify or Amazon, you have directly experienced the results of machine learning. Both companies use machine learning algorithms to recommend products based on your listening and purchase history. And as you feed them more data, the recommendations get even better.

The limits of machines and the limits of humans

Now, all this sounds amazing.

But, is that all? Can machines actually replace us — given the great results achieved in tech? Will machine learning make our job obsolete? Will the industry where we work be completely automated and run by AI, making us humans redundant?

To answer this, let me start from my industry, by giving some examples of AI/ML applied to consumer insights.

Simon Chadwick in this article brought great answers to the widespread feeling that marketing and business insights going forward will be a technology-driven industry and that the human element will become increasingly unimportant.

Simon argued — backing up with useful case studies — that the best tool for understanding the human being is the human brain and given the growing excitement around AI and ML, these forms of tech-led processes will not replace the ability of the human brain, human collaboration, human curiosity and the capacity to accept ambiguity and contradiction as pointers to insight.

I totally agree with Simon. Another case that I can think of on top of my mind is NLP vs NLU.

NLP (Natural Language Processing — a part of computer science and artificial intelligence which deals with human languages) is making our job much more accurate by giving us the chance to scale research and understanding. It’s getting better and better, even though it carries — still — limitations if compared to human language processing.

But the gap human-machine is huge when we face NLU.

NLU (Natural Language Understanding) is much harder than NLP, and of course extremely easy for humans, even for non-developed humans, see: babies. It concerns things like text planning, sentence planning, text realisation.

So, why is NLU so hard for computers?

1) Lexical Ambiguity (also called semantic ambiguity): the presence of two or more possible meanings within a single word. I.e.:

  • She is looking for a match (Is she looking for a match or a partner?)
  • The fisherman went to the bank (A bank where he withdraws money or a bank where he parked his boat?)

2) Syntactic Ambiguity (also called structural ambiguity or grammatical ambiguity): the presence of two or more possible meanings with a single sentence or a sequence of words.

  • The chicken is ready to eat (is the chicken ready to eat something or is the chicken ready for us to eat?)
  • I saw the man with the binoculars (Did I see the man by using my binocular or did I see a man holding a binocular?)

3) Referential Ambiguity: this ambiguity arises when we refer to something using pronouns.

  • The boy told his father the theft. He was very upset (was the boy upset or his father?)

But let’s step outside of the market research/consumer insights industry.

Perhaps one of the most famous examples of machine learning was when a supercomputer from Google research made headlines. When it was fed 10 million thumbnails from Youtube videos, the computer was able to learn how to identify a cat with 75% accuracy.

That seems impressive, right?

Until you remember that a 3-year-old can do this with 100% accuracy.

This further shows the differences between humans and machines, when machines can beat the smartest mathematicians at some tasks and can’t beat a 3-year-old at others.

My point is: humans and computers are fundamentally good at different things.

We can’t make sense of an enormous amount of data, but machines can do that efficiently.

We can form plans and make decisions in complex situations whereas a machine can’t make a basic judgement that would be simple for a 3-year-old.

We experienced this first hand analysing consumer insights: we faced the limits of both humans and machines.

Let me explain to you why, exactly. Our work at Trybes Agency requires us to gain a deep understanding of people.

We look for answers to questions like “What makes people join a topic, a brand, or a movement?”. For that, we need to understand their cultures and beliefs, why they make specific choices, what influences them, and how their motivations and perception supports the way they see their world.

No human-only or machine-only solution can gain the level of understanding we aim to achieve on people.

To do this, we have to look through millions and sometimes billions of data points. A huge portion of which is qualitative.

The future is hybrid: machine learning + human understanding

So, our solution was to create a hybrid intelligence, a part-human, part-machine solution. Our hybrid intelligence approach couples AI-driven research methodologies with social sciences.

How exactly does hybrid intelligence work?

Let me give you an example of a full data science project in market research.

A few months ago, the tourism board of Sardinia reached out to us to conduct market research in order to design a winning marketing strategy.

Sardinia is an Italian island with Maldives-like beaches and ancient, unique culture. Considered by us Italians one of the most incredible places on Earth. If you haven’t been there yet, you should really consider spending some holidays there.

So, here are the steps we took, and so you can see which step is human, which is machine-led, and which is both.

1) Understanding the business problem

Sardinia wanted to have a deep understanding of who the traveller is and why she wants to travel to Sardinia as opposed to Puglia or North Africa.

The thing is: a deep understanding of a user involves a lot more than their demographic data.

Think about it: can we define this person by saying that she is 35, female from Nancy, living in Paris?

No, we can’t. Sardinia won’t benefit much from information like she is 35, female from Nancy, living in Paris.

We have to go deeper.

[human or machine?] — human

2) Data Exploration and Strategy

In this step, we try to understand what matters for French travellers interested in Sardinia as a destination.

Once we have understood what matters, we now have to understand where to gather the data.

To get this done, we have to ask ourselves questions like:

  • Where do French travellers gather to make their travel decisions?
  • Where are the conversational territories where they gather to talk about Sardinia?
  • Where do they find information?
  • What else has an impact on the perception of Sardinia?

[human or machine?] — human

3) Data Mining

Here we gather hundreds of thousands of data from multiple sources, such as:

  1. Web servers
  2. Logs
  3. Databases
  4. API’s
  5. Online repositories

In simple, non-technical words, here we gather a bunch of data from multiple sources. In this travel case study, this includes information on hotels, historical records on reviews on travel agencies like Booking.com (over 342 thousand reviews) and Airbnb.com (almost 20 thousand reviews), online forums and communities, and over 200 other online destinations.

[human or machine?] — machine. It would take us years to do that manually ;)

4) Data Preparation/Wrangling

  • Data Cleaning: since the data gathered is obtained from numerous sources, which are of varying kinds, the data is prone to be noisy, to have inconsistent data types, misspelt attributes, missing and duplicate values. Thus, we apply Data Cleaning techniques to remove noise and correct inconsistencies in these data.
  • Data Integration: In this step, we merge the data from multiple sources into a coherent data store, and we store them in a database (sometimes available to the client for further inquiries).

[human or machine?] — part human, part machine. (the machine cannot possibly identify the data we don’t want. But on the other hand, we need the machine to automate the cleaning as we can’t remove the data manually for millions of records).

5) Data Understanding

Here we need to simply understand what the data is telling us by finding patterns, correlations, and insights in the data.

In the case study, since the majority of the data was qualitative and we couldn’t possibly look into each review or conversation (only Booking.com had over 340.000 reviews!!!), we applied the branch of Machine Learning called “Unsupervised Learning”. That allowed us to find patterns based only on input data, as explained earlier.

Our goal here was to allow machines to dig into the data, group data items that have some measure of similarity based on characteristic values.

This makes such a large qualitative dataset manageable for our team of researchers, psychologists, anthropologists, analysts, and ethnographers. And it surfaces parts for us to dig further and give context to the data.

[human or machine?] — part human, part machine

6) Data Visualisation and communication

Here we communicate our findings followed by actionable insights (which means insights that can be implemented in the business and marketing strategy, we talked about it here in the Consumer Insights Roadmap — you can download the whole roadmap at this link)… simply and effectively to the stakeholders.

I want to stress here the importance of this step. Communicating insights effectively is so crucial for a company that this step generated in some cases the need of a data translator: being relevant and speaking the business language with the stakeholders seems to be quite rare among researchers (according to CEOs and CMOs).

I think that being an entrepreneur and having a marketing background helped me a lot in the past to communicate consumer insights that are aimed at business results. I think that all research agencies and suppliers should try their best to study companies’ strategies and revenues, in order to make their information valuable.

[human or machine?] — mainly human, with a bit of machine in the automated data viz tools.

Conclusion. Benefits of using hybrid intelligence to get actionable consumer insights

So, what are the benefits of hybrid intelligence for marketing a travel product or destination?

With the process explained above, we were able to identify opportunities and threats for Sardinia in the market, for example, which price segment was less satisfied (thanks to machine learning).

We were also able to manage qualitatively (thanks to human understanding) a large amount of data (machine learning powered it), and this resulted in the topicgraphic analysis of the French traveller.

Do you remember the French traveller at the beginning?

Here is what topicgraphic data looks like.

As opposed to saying she is 35, female from Nancy, living in Paris, we are now able to define:

  • How long her trip is
  • How she gets there
  • The cities and towns she wants to visit
  • The locations she prefers (the beach, not the countryside)
  • The activities she wants to do
  • The type of accommodation, restaurants, and tours she wants to book
  • The problems and concerns she has
  • The services/disservices that are important to her

If you have any questions or you want to dive into this case, contact us. I’m happy to share the consumer insights obtained there.

Other cases studies we worked on

We used the hybrid intelligence model to analyse the most different segments,

from mothers to the Burning Man ecosystem,

from healthy living lovers to Italian Americans,

from digital nomads to beauty products users,

from fitness enthusiasts to binge-watchers.

Our Burning Man ecosystem research was the first quant and qual ever conducted on Burning Man and Transformational Festivals.

Our goal was to be able to answer questions like “who is BM?” Is it the official channel, or is it all the tribes, communities, official and unofficial events that gather around the BM?

And even more complex questions “What does the BM and the transformational festivals mean for millions of people?”, “Why do the tech leaders go to such an event”, or “What is BM impact on society?

These are examples of how we were able to leverage machines as tools to look for signals in large amounts of data for us to dig deeper into and find answers to complex questions about anything.

But, we’re not the only ones. There are several cases of humans-machines hybrids out there.

Peter Thiel, for example, suggests that we should play complementarity as opposed to substitution.

In Zero to One, he explains how the human/machine solution allowed PayPal to defeat online credit card fraud.

They were losing $10 million per month. While humans couldn’t possibly process each transaction to check if it was a fraud, they tried to automate fraud detection with a designed software.

But, as turns out, thieves were smarter than machines, changing tactics and fooling them. So how could they stop this race machine-humans?

They created a hybrid human-machine that would flag the most suspicious transactions on a well-designed user interface, and human operators would make the final judgement.

They saved Paypal by fixing a major problem, but they didn’t stop there. This results caught the attention of the FBI and inspired Peter to create big data analytics company Palantir.

Another fascinating case is the one of Netflix, that is majestically revealed by Alexis Madrigal in The Atlantic article.

Everybody knows how Netflix wins over by turning up with the right content to the right person. That’s why we binge, right?

This implies, yes, tracking the behaviour of millions of users (which is a machine’s job), but first, it means understanding the content itself so that they can propose you similar content.

But how could they understand the content and classify it?

It turns out that a trained team, under the strategic supervision of VP of product Todd Yellin, watched all the movies and TV shows to understand and classify by micro-tagging all the content.

They decoded, deconstructed, all the movies and shows by micro-tagging characteristics as specific as, for instance, how socially acceptable are the protagonists, how romantic is the movie from 1 to 5, where is the location, and when is set, who are directors, actors and whether it has a happy, sad or ambiguous end.

So next time you think the Netflix recommendation algorithm is so great, remember it was a hybrid intelligence that powered it.

In the TED talk, I’ve listed a few other cases, such as:

Amazon cut the time it takes to prepare a product for shipment from 1 hour to 15 minutes by using collaborative robots as opposed to human-only or machine-only solutions.

Mercedes E-Class factory achieved dramatic levels of performance by de-automating the large scale robots and using instead smaller scale robots collaborating with people.

The Japanese startup Ory Lab launched a cafe where robot waiters are controlled from home by paralyzed people. This was my favourite… I mean, just imagine: there are millions of people whose lives will be forever changed by this kind of virtual presence workforce!

There were books, articles and resources that have been taken off from the initial TED draft, for — again — time reasons. They have been incredibly inspiring for me and great lessons in our own hybrid intelligence path, so I’d like to share them with you. Please let me know if you come across any other, I’d love to co-create knowledge on this topic.

Academically speaking, there is the International Journal of Hybrid Intelligence (IJHI) that focuses on the role of the hybrid intelligence paradigm in the modern context of rapidly evolving technologies, and Ece Kamar’s Directions in Hybrid Intelligence: Complementing AI Systems with Human Intelligence (Microsoft Research).

In this article, Jo Stichbury explores how the hybrid model of computer-human intelligence offers a way to build on our mutual strengths and deliver efficiency. In this other, Ryan Nakashima reveals AI’s dirty little secret (spoiler: it’s powered by people).

I also loved the panel Artificial Intelligence: Augmenting Not Replacing People. Here is an extract of the synopsis:

Artificial intelligence (AI) technologies are rapidly maturing into tools that are impacting our everyday lives. However, contrary to popular conception, most of these tools will not be autonomous, stand-alone systems, but rather will manifest as human assistants and augmentations. While autonomous driving is featured in the headlines, the short-term impact of advances in this field will be increased safety, comfort, and convenience, with the driver still at the wheel. New technologies in healthcare will not replace doctors, but will leverage their skill and judgement by providing super-human augmentations for eyes, hands, and intellect. As more robots move onto the manufacturing floor, they are most likely to function as ever-smarter programmable tools, and will still require human coworkers to teach them new tasks and to do those elements that are simply too hard to automate.

I even created a Youtube playlist dedicated to Hybrid Intelligence. It includes Alan Turing’s Imitation Game, when he came up with the epic question: Can Machines Think?

If you’re into hybrid thinking, you might have watched Get ready for hybrid thinking by futurist Ray Kurzweil. But, have you ever heard of NLG technology? Basically, machines that write for us.

According to Professor Robert Dale, Natural-language generation co-authoring gives the best of both worlds — human and machines.

  1. Human authors bring their insights and nuance and their subtle understanding of audience
  2. Machines can do the grunt work that would otherwise take a human author endless amounts of time, if it’s feasible at all, delivering detailed and accurate descriptive narratives about the information that would otherwise be left buried in data.

And what about Pinterest’s ‘artificial’ artificial intelligence?

In this article Adrian Bridgwater explains how, in order to build what it calls ‘artificial’ artificial intelligence, Pinterest’s ‘Discovery Science’ team built a library that plugs into multiple crowdsourcing services, including CrowdFlower and Mturk.

Maesen Churchill, software engineer at Pinterest, talks with Adrian about the things humans tend to do better, such as evaluating content, and how they’ve automated human evaluation for the purposes of analyzing the relevance of search results and to filter out certain types of content. This ‘artificial’ artificial intelligence, might be why Pinterest quality of content is way better than the other platforms’ content.

On the other side of hybrid approach, a very interesting POV on augmenting humans with tech is David Eagleman’s on Data sensing: essentially, we can use technology to develop new sensory perceptions to supplement or complement our current capabilities.

And while Karthik Rajan argued that intuition is still in the hands of us humans, and machines self-learn only for problems where the goals are clear and quantifiable (he talks about the epic Lee Sedol vs AlphaGo here), Giorgia Lupi was launching the Data Humanism Manifesto that forever changed the world of data vis and, I’d say, the whole big data industry. She, in fact, showed us the way to humanise data, and that resulted in shaping some of our careers (including mine, Giorgia! ❤)

I mean, WOW, right? I personally feel excited just listing these pieces of knowledge here. I could go on and on. For hours, maybe days. :)))

Wrapping up. Fear is not a solution, but you can leverage it to create impact.

Fear is never a solution, even though it sells and it’s easier to leverage (see: cli-fi, Black Mirror, pretty much all the narrative around AI and data. Etcetera etcetera: look around you, fear is everywhere).

In my case, I used it as an excuse to make a point in front of a broad audience that would probably not care much about consumer insights technicalities. But amazingly, people care enough about their future that they picked up my provocation and started talking about it and acting differently.

What about you? Are you using machines as tools to take your skills to the next level, or are you playing machines? Are you making yourself indispensable or easy to substitute?

The truth is, we are not machines. Machines are not humans.

I believe there are great benefits to those that find creative and strategic ways to redesign their work with tech. We should capitalise on the best attributes of humans and robots, instead of fearing the advent of machines taking over.

I’m very proud that at Trybes we took this path. I’m willing to fight for the importance of human understanding in the data analysis/market research industry.

Also, I’m a big fan of making complexity accessible, ’cause another big cause of fear comes from not understanding. The TED talk constituted an incredible opportunity to explain some of the most buzzed words in our industry, while giving us the chance to matter in the actual world by showing the way to hybrid intelligence processes.

I hope this article can be of inspiration to marketers and researchers that are interested in applying hybrid intelligence, machine learning and human understanding to consumer insights. After reading this, you should feel more confident in talking about these topics and forming your own opinion, beyond buzzwords.

Please let me know what your takes are on this, and let’s keep the hybrid conversation going. This is literally just the beginning.

(If you wish to save this article for later, you can download it here)

Jump on the video instead: Hybrid intelligence: rise of the human side of data in the A.I. era.


|