ELEMENT HUMAN - Matt Celuszak, Founder & CEO. What's next for the attention economy and emotion AI? Strengths & weaknesses of facial coding data; introducing measurement for the influencer economy; the importance of great co-founders.

From stingrays and hot coals (listen to find out) to building one of the leading companies in facial coding, emotion AI and behavioral leading models, Matt Celuszak of Element Human has had quite a journey. Matt’s always astute, engaging and  in this interview  we get into the issues above and more, including 

• Strengths and weakness of facial coding data
• Evaluating the intersection of claimed and behavioral data sets
• Understanding effectiveness of content for the BBC - rating vs reaction
• Using sensor date to create measurement for the influencer economy
• Measuring what’s valued, rather than just valuing what you can measure
• How to derive granular human understanding from big data sets
• How privacy works in relation to sensor data
• How behavioural learning models work
• Lessons from building a business from scratch. 

All episodes available at  https://www.insightplatforms.com/podcasts/

Suggestions, thoughts etc to futureviewpod@gmail.com

Transcript

Speaker 2 0:09

Emotions happen before people think . So if you ask somebody how they feel , you're actually asking them how they think they feel , because whatever they felt has already happened . Being able to get landmarks on a face and be able to look at unique features on a face , that works , that side works . I'm going to separate facial coding from emotion .

Speaker 2 0:27

AI because emotion is a judgment call as to whether somebody is feeling happy or sad , or it's a probability as to whether somebody is feeling happy or sad . That's where a lot of these models fall down . I would be careful and reticent as a user to be putting a lot of weight behind the emotional labels .

Speaker 1 0:52

Facial coding data in the insight space has

Episode introduction

Speaker 1 0:54

been around for quite some time . In simplistic terms , it's a technique that involves reading the signal or landmarks on a respondents face to understand what they're paying attention to , and the reaction are listed . It's widely used by a variety of companies , often to gauge reaction to advertising content . Many brands and agencies have bought into the theory . Facial coding data will give you a true , unfiltered reaction to content understanding system . One type reactions based on immediate or emotional rather than stated response . However , there's also been some skepticism around this type of data too , not least around a number of key questions how accurate is it ? To what extent can you really categorize emotion , self-selected samples , who are engaging with the content , questions around consumer privacy and a whole range of other issues .

Speaker 1 1:37

Well , matt Saluzac addresses all those topics and many other very relevant points , including how Element Human , the company where he's the founder and CEO , takes a considered approach to this type of data , the use of behavioral learning models and how human-sensitive data can empower the creator economy . I must give a disclaimer that I'm on the board of Element Human , so I have some existing ingoing perspectives and viewpoints . That said , I hope you'll agree that we have a really good in-depth discussion around my facial coding data and emotion . Ai may go next and the impact on insights and marketing . So onto the interview . So , matt , it's good to track you down wherever you are in the world . They , west Coast of Canada , I've caught you on , I believe .

Speaker 2 2:20

Yeah , thanks , Henry , for having me here .

Speaker 1 2:22

Not at all . It's very , very good to have you on Now . I wanted to start off with a traditional icebreaker , which is something from your deepest , darkest past . It doesn't have to be deepest , darkest past , it's just something that most people wouldn't know about you , something that might be a little bit surprising .

Speaker 2 2:38

Well , I don't think you'd actually find it anywhere written anywhere , but

Stingrays and hot coals

Speaker 2 2:43

I was actually stung by a stingray and almost lost my leg down in Costa Rica .

Speaker 1 2:48

Wow , how long did that happen .

Speaker 2 2:50

Oh , it was in my 20s .

Speaker 1 2:51

Oh , so two years ago .

Speaker 2 2:54

Yeah , I wish it was two years ago . No , my other leg reminded me just recently how old I am with an Achilles rupture . So yeah , unfortunately , being active certainly gets you out there , but getting you out there certainly gets you exposed . So yeah , no , I guess that's something that most people wouldn't know , but I spend a lot of time in the ocean and I really enjoy the ocean , and people know about my fishing , some of them know about my kayaking , but a big part of out there is just getting out there and just being in the wild . And a stingray stumbled across my leg and almost lost that thing .

Speaker 1 3:33

So what are the consequences of a stingray bite , then ? I mean clearly it does more than just really hurt . I mean , if you almost lost your leg .

Speaker 2 3:39

Yeah . So I mean it really hurts . Yeah , Not the most painful thing , but close to the challenge that you have is that the poison goes in and it starts to create necrosis and so , and unfortunately , hit my lymph system . So the whole leg kind of started to swell up after three or four days . I couldn't really figure it out and we just had to absolutely nuke myself with antibiotics , try and recover , and it was . We ended up getting it back . It didn't help that we were kind of remote Costa Rica at the moment , and so at the time it was very much . They said put heat on it , and , or at least the doctor did , and the people there I was out with a bunch of kind of local guides and they were like , oh they , just it was hot , is what they heard . And so they put these burning hot coals on the back of my leg and I got third degree burns and so great , just to make it even better Complicated the whole issue .

Speaker 2 4:41

It's the first time I've ever bit down on a stick to handle the pain .

Speaker 1 4:44

It was quite interesting Anyway well , man , that was probably , that was probably very good preparation of burning hot coals , biting down on sticks , extreme pain for getting into the startup journey in the startup world which we're going to get on . So we'll get maybe into your background a little bit , which is interesting in itself . But could we just start with element human

Entering the world of human sensor data

Speaker 1 5:06

and it's various sort of iterations , and why did you found the business and what were you looking to do ?

Speaker 2 5:12

Yeah , so my co founder , Diego , and I founded the business based on a kind of one simple principle we were realizing that people were starting to spend more time interacting with technology in the outside world than they were with human beings , and I think we can all say that even our interactions with humans were going through technology , but technology was still in a very binary stage I mean , we founded this in 2013 .

Speaker 2 5:37

So this is still what people were in the marketing industry were all like , oh , mobile phones are going to take over , and so we had heard about kind of artificial intelligence and machine learning , and we thought , well , if people have kind of cameras and microphones and all this really rich data that we get out of , say , a focus group or out of one to one interviews that contextualizes everything , could we create a nice feedback loop that complements this big data , but very kind of broad data set . That that's . That's not very granular , and so we wanted to work on generating human behavior signals and emotional signals out of sensor data , and we just believe that sensor data offered this kind of body language feedback loop that helped us really unpack kind of why and how people are doing things , rather than just what they're doing .

Speaker 1 6:29

So that's kind of where we got started . And , matt , when you were saying sensor data , we're kind of diving into the nitty gritty bit now , and so obviously it was always survey based , I believe . Was it the in terms of elements , humans solution ? Oh , it wasn't okay .

Speaker 2 6:42

No , no , no , no , no . So our first iteration was actually an API where you could just send a video to us and we would translate what the expressions and emotional signals were , and that was using early day support vector machine models . So it was really good . And then we teamed up with the University of Nottingham and and then we started . It was BBC who asked us to bake it into a survey to demonstrate the value , because what we were learning is that people would express differently than they would claim , so we would have the , the expression , and then we'd have what they claim . And then those two they were contradicting where they contradicted ended up being where the insight was . So that actually ended up being the leading indicator as to where whether a piece of content would perform or not at the BBC . But that's where we were incubated and that's when we ended up putting it into survey .

Speaker 2 7:31

The fourth sorry , the third generation of the tool was actually an embeddable , so it was again being able to flip on a webcam , do eye tracking , emotion detection and implicit testing and and as a researcher , you could just plug that into a survey . And then it was . It wasn't until this fourth generation that we ended up taking over the survey , the sample , all that stuff , because we realized quality control for rich data was was important . So part of that is trying to design out the human variability enough while creating a natural enough human experience .

Speaker 2 8:07

So and by owning the survey . It allows you to do that .

Speaker 1 8:11

Yeah , yeah got it . It's really interesting . I mean going back to that point that you made around the difference between claims data and behavioral data

Understanding the interaction of claimed and behavioral data

Speaker 1 8:20

. You may have used different phrase . Now I forget exactly what it was .

Speaker 2 8:23

But then but then ?

Speaker 1 8:24

but then how did you start to discern ? You said the truth was somewhere in between or the intersection gave the truth . But how did you start to discern where the truth in inverted commas would lie ?

Speaker 2 8:36

The first kind of big project where it became quite clear that there was a whole different type of signal you can get from from interacting with people and sensor devices was when we had people rate trailers of the BBC in multiple markets and then we had them react to those trailers or just watch it naturally and see how they reacted , and you'd get people . It ended up creating kind of a two by two where you'd have kind of a rating that was high low and a reaction that was larger or small , and we could very clearly see that there were people who would rate high and react high and that was high engagement , true engagement . There was people who would rate high and react low and those are people that were more they fill out surveys and they want to say that they would watch that show because their friends do , but actually when you went deeper they didn't know much about the show itself and they actually wouldn't watch it . Then , when they would kind of rate low and react low , that was just not a very good piece of content for them . It was not good audience fit . And then the final one was the really interesting was where it would rate low but it would react high and where it was rating low and the reaction was high , we were actually able to .

Speaker 2 9:52

Those were always the shows that outperformed prime time average , and so we found out that we could actually anticipate beating prime time average about which shows and which markets would do well , and so that held the BBC price their inventory for those markets and sell it in . So it was these . We called them the guilty pleasure . It's these shows that you would never admit that you would watch , but when they're on , you're never going to turn them off . So the Jerry Springer's , the talk TV show , some of those ones where you wouldn't want anybody , you wouldn't want your friends knowing that you actually watch this , but when you watch it , you love it .

Speaker 1 10:31

Yeah , I think we're all aware of all sorts of guilty pleasures .

Speaker 2 10:33

We're all aware of which ones those are from the press .

Speaker 1 10:37

But that is interesting in that it's so . It wasn't really one or the other . It genuinely was an intersection of saying by putting the two core data sources together , you get a much more nuanced picture as to what the reaction is . Yeah , kind of interesting , man , I realized . I know what element human does very well , but you should probably actually describe what element human does for people who are less familiar with the business .

Speaker 2 11:05

Yeah , sure . So I mean , in its current articulation element , human is very much . We work on trying to take sensor data

Using sensor data to help the influencer economy

Speaker 2 11:15

and learn how people behave or react to any interaction they have with a device . Today that articulation is in ads and ad measurement and media measurement . The media measurement industry , or the media industry in particular during COVID kind of hit this new bubbling up influencer and creator economy which has over 200 million creators conversing with 4.2 billion people 365 days a year . It's a whole different form of marketing . It offers the largest untapped market in the world and there's absolutely no way to measure across platforms at the moment and the old traditional TV metrics are really hard to apply because it's a nonlinear format . There's a lot of user interaction control . The media formats are changing or expanding , so how do you compare a Snapchat with a Facebook media buy and how do you know whether your content's better on YouTube or Instagram , youtube shorts or Instagram real ? So those kind of common questions became real challenges A for the platforms to prove their value , their unique value and their qualitative value , but also B for the media buyers and then C for the creators themselves . So how do I stand out in these platforms If I can make money from this .

Speaker 2 12:31

Early days , there were very few creators and a lot of ad dollars available for experimentation . Now , total opposite , there's creators coming on a dime a dozen . There's a lot of content that's being pushed . People are kind of tuning out these feeds . You've heard of the doom scrolling and that type of stuff , so attention is really important . So the question for a creator is how do I capture attention and how do I get them to stay with me ? That also ladders up to the platform level and then further ladders up to the brand performance level . So we're working with pioneers in this space Whaler , twitch and the BBC to understand what is really good quality content that's going to grab people's attention in the context of the environment . And how does that allow each of those type of stakeholders creators , brands , agencies or platforms be able to articulate their unique proposition and the value of that proposition in a measurable way ? So instead of measuring clicks , likes and shares which are okay for audience growth but not great for brand performance instead of valuing what's measured which is what the whole programmatic industry is kind of based off of that's flipping it on its head to measure what's valued .

Speaker 2 13:46

And I think that's where sensor data is really interesting because it's a super rich data source where surveys can't give you that context . You get this really rich data source and you could kind of it creates a machine learning feedback loop where you can kind of learn that against whatever metric that matters . So first and foremost , it's an ad testing tool for campaign measurement . It's been used by over kind of 200 brands . We've road tested this for the last four years . We've got over 30 billion data points and benchmarks across four big platforms . It's pretty cool from that perspective , but it's got a long way to go too . There's a lot of the question is what else can body language teach us ? What else can we learn it against ? It's a hyper contextual source of information and I think , just to put this in context , for every like , like a thumbs up that you get on a YouTube video , we're getting over 10,000 data points . So for every single data point you're getting on a YouTube video from a like perspective , we're getting over 10,000 .

Speaker 1 14:45

And Matt , we'll get into that , to get in a second as to what you mean by those 10,000 data points and exactly what you're measuring , but playing it back a little bit , so it sounds like there are multiple propositions in that . You've got , I guess , the old sort of insights propositions being around for quite a long time or surveying , gauging the effectiveness of your advertising , whether from a creative perspective , in terms of you're making the advertising you want to make it better . That's a use case . But it also sounds like there's a really interesting use case too in terms of areas in the digital world , particularly the influencer world , whereby there isn't really a lot of measurement at the moment , there's no measurement , to be honest , and the measurement that does exist , it's usually cost prohibitive .

Speaker 2 15:33

So our goal with the creator economy is to create something that was super fast , fast enough for influencer , reliable enough for brands and for platforms and agencies , but also cost effective , so that it could scale a lot , but also accessible to these influencers , who might actually be pre-revenue in their own personal journeys . And so how do you attract brands in a language that they speak , but in a way that's useful to you ? So you can . You have a feedback loop , so it's worth asking your followers to review your content .

Speaker 1 16:12

So the use case in that side of things is one of which is fine . You get your followers to review the content , you show it to them , you get the feedback in a survey environment and also , using facial coding and potentially other metrics , you can make better content . But then if a brand is looking to buy from , say , I was an influencer , then I can actually start to give them some validated metrics around recall or purchase intent or some type of brand on list .

Speaker 2 16:41

Yeah , so exactly . So it's brand uplift , brand recall , full funnel analysis . It provides all that .

Speaker 1 16:50

And so that , yeah

How the process works, strengths & weaknesses

Speaker 1 16:51

, on that side of it , could you just talk through in simple terms ? So people who aren't familiar so how does the methodology work ? At the moment it's mainly a survey based methodology , but then what happens ? Just very high level , if you're a respondent and you go to that survey , and then what does the client get at the end of it ?

Speaker 2 17:11

Yeah , so , client , we handle the sample our partners sent . So they and they're one of the main industry providers for participants . So you basically just go into the platform , you upload your creative , upload a few kind of brand goals and objectives , and then you select the audience you want to go to . Then we then a link gets sent out to the our partner sample partner . Some people , like Twitch , bring their own samples , so some some of them plug us into their own customer databases and communities and then the link goes out . They agree , they consent to do it . They turn on their webcam . That's part of it . It is entirely privacy based we'll get into privacy in a second but they turn on their webcam and as they're filling out the survey they're being observed and then they get to the points of where they're popped into a mock feed or they're exposed to the content . It was just a simple creative test and then they're .

Speaker 2 18:07

Then we also do test and control . So there is a group of 100 people who will be done as the control group to be able to compare all the results and see what the differences are . So that's what creates our benchmarks , that's what allows us to do uplift and recall and awareness , consideration , all that stuff , and then you can see how much you actually would compare to it , either a gen pop population or a more targeted control population , if you want . So from a brand , there's essentially three products create , measure and learn .

Speaker 2 18:37

From a create perspective , you understand where people are paying attention , what's catching their attention , which objects and scene and elements inside the creative is capturing their attention . You get to understand where the emotional key moments are . So what are the parts of the creative that are actually driving their engagement , continuous engagement within the creative , and you can get some emotional context as well , kind of some of the labeling around , kind of happy , sad , that type of expression that's being reviewed . And then , on the measure , it's very much your traditional brand metrics photo analysis , all that type of stuff . We also have implicit traits . So you can , I understand is it authentic , are you trustworthy , are you sustainable , whatever your brand goal is , are you actually getting there ? And that's a subconscious kind of yes , no . And then the final bit , of course , is learning . So once you have all that data across a bunch of your data sets , you can then learn your own bespoke metrics for you as a customer , or you can start to pipe in , say , your sales data and machine learn against that .

Speaker 1 19:43

Got it and to ask the million dollar question , hopefully maybe literally the million dollar question . So , facial coding data and eye tracking . Does it really work ? Inside agencies have looked at this for a long time and I think in some cases they've had trouble with it in terms of shaky cameras , self-selected respondent groups , problems with beards , whatever the issues may be . So I know that you're going to say it's really , really improving , but what's your kind of honest evaluation at the moment ? The strengths or weaknesses of facial coding data ?

Speaker 2 20:27

Yeah , so facial coding data , or facial recognition data . I mean , frankly , it's been around for yonks and I mean does it work ? I mean , some governments are using it to do automated border control .

Speaker 1 20:41

Okay , yeah , true .

Speaker 2 20:42

Yeah . So , quite frankly , the facial coding stuff works and the computer vision side behind the facial coding , more specifically , facial recognition , works . So being able to get landmarks on a face and be able to look at unique features on a face , that works . That side works , the side that I'm going to separate facial coding from emotion .

Speaker 2 21:03

AI because emotion is a judgment call as to whether somebody is feeling happy or sad , or it's a probability as to whether somebody's feeling happy or sad . And that's where a lot of these models fall down , and I would be careful and reticent as a user to be putting a lot of weight behind the emotional labels . What the emotional labels are doing is they're mathematically taking a probability of somebody , say , smiling , and saying , well , we have labeled this smile as happiness . And that is highly dependent on two things . One is your training data set . So who you've trained against ? So if you've only trained against non bearded people , bearded people are probably going to fool the system because of the way that the machine learning supervised machine learning works . The other thing , the other challenge that you have , is a universal agreement on what emotions are and how to label them . There's a reason why body language goes largely unlabeled , even in human construct the word love and the emotion . Love in English is summarized in one word . In Sanskrit it's 96 words , so so , so there's 96 different variations that you can get in Sanskrit to describe the same emotional feel . You and I would have entirely different perspectives of what happiness means to each of us , and I think that has been the biggest misnomer is actually less on the technology itself and more in the theory of emotional modeling . And so what I would , what I would caution any user on in this space , because of course it's always improving , because the technology is always improving . Mathematically you can get facial points off a face pretty accurately . You can deal with the human variability movement now with the high quality webcam . So if you tested this four years ago , you would have been experiencing human variability from mobile webcam movement . Today that's pretty much gone because it's it's pretty easy to lock onto a face on a moving and moving , so so two moving objects back and forth , so that stuff's pretty much gone . The racial issues that were happening a lot of those are gone now because of the way that cameras have improved .

Speaker 2 23:20

But the part that still I would say be very careful is the labels that you're using and your interpretation as a user of those labels . So we offer the labels , but we offer the standard Ekman labels and we only offer it in the create function because they're only useful in saying were you intending to get people to smile here ? If you weren't , then they are smiling . That's what we can tell you If we are saying were you intending to make people disgusted here ? Well , we did a trial on disgust and just the different ways that people express and react to disgust . We had Bear Grylls eating an exploding bug and people would grimace , people would laugh , people would cover their face , people would turn away from the screen , people would lean into the screen . Everybody had a very different reaction . So you can't really model that . What you can model is there's a lot of expressional movement going on and therefore that's a key moment , and so the key moment stuff is very solid , very expressional .

Speaker 2 24:17

Emotional interpretation is very psychological .

Speaker 2 24:20

The one thing I would say and this is what a lot of people get wrong is that they think that emotions don't work . The reality is that you're not going to get a better emotional read than you're not going to get a better emotional read anywhere else . That is at least objectively validated , because a psychologist will give you their view and that opinion and a focus group is just too small , and so you have , and the ability to anticipate truly anticipate an emotional level , depending on how you ground truth . It is just so difficult that , even at 50% accuracy rate of emotion or whether it's 80% accuracy of emotion , depending on which one you're trying to express . That's still a lot better than a 20 , 30% that you're going to get in a human . So the here I would say the where it has been massively misused in the market research industry is people think that emotions and facial coding is to scale up to quantify a qualitative element , and it's not . It is to scale call , not to enrich , quant , and I think you have to be really careful about that .

Speaker 1 25:30

Yeah , that's a really interesting distinction , and so that's a what can you quantify ? Then you can quantify the most .

Speaker 2 25:39

Yeah . So what's ? What's hyper accurate at the individual level is whether we've detected a face , whether we have actual landmarks on that face and all the vectors that go in between . So all these vectors , you might have like 256 to 1000 data points on a phase and then those will ladder up into what we call classification . So happiness , sadness , and that's looking at certain muscle groups predetermined by an emotional model .

Speaker 2 26:10

Our view in this whole world is that that's in this whole space , is that that's actually the wrong way to approach a problem . Our job is not to go in there and tell you what emotion they're feeling . Necessarily . Our job is to see if body language and behaviors are indicative or a leading indicator of a metric that matters to you and your business . So by having this really rich data on the face , you then , as Coca Cola , might say well , when we see brand lift , go by X , that means we've hit our happiness moment . Okay , then your version of happiness can be machine learned against that metric and we will tell you . It turns out vector three , four , 75 and 989 are all indicators , strong indicators or predictive indicators of that performance . And that's a much better mathematical way to do things in the emotion space than trying to label somebody's emotions , because our language is just frankly too limited . So , and it's the wrong question .

Speaker 2 27:15

Emotions . Emotions happen before people think . So if you ask somebody how they feel , you're actually asking them how they think they feel , because whatever they felt has already happened .

Speaker 1 27:25

What did you mean by the vector example you just gave that ?

Speaker 2 27:28

you have a bunch of landmarks on a face and then vectors are the space between the landmarks . So we track how how much they move back and forth .

Speaker 2 27:36

Okay , so if you can imagine , a face with a face mask that has a bunch of dots on it , we'd be tracking those dots and as the person moves their face or moves to and from the camera , we're tracking the distances . And those distances are the things that tell you whether there's expression movement or not and whether that expression movement is a large enough , a relevant deviation from the norm .

Speaker 1 27:58

Interesting and again turning kind of quite deep here , but I like to geek out on this stuff , as you know . What exactly are you tracking ? It's the face you're tracking . It's not broader body movement , it's just the face .

Speaker 2 28:11

So the way that we built our system is you can kind of you can have it recognize any part of the body if you wanted it to . So if it was , if the interaction was broader , body movement great . But television media at this point in time , and that's where a lot of the consumption is for people it's a pretty passive activity . So face is probably the most expressive and most telling . Hand occlusion you'll see time to time . Hand occlusion is good , and eye tracking and head gaze . So eye tracking , head movement , those fall within the vectors of the face . So that's so we would track those as well . Who knows , we might find out that a nose flare twitch might be indicative of , you know , high sales for the toothpaste category . But we just don't know that . The key is to get the really rich data first and then be able to link it to whatever metrics you want .

Speaker 1 29:01

Which probably actually links to the next question . I was about to ask Things like audio and strength of their reaction when they're speaking . I imagined companies must be exploring that as well . I mean , there's a big rise in open ended , open end responses through audio . Is that something you've been looking at as well , or do you feel that there's enough to work on with ?

Speaker 2 29:24

broadly with the face , no , so multimodal has always been our end game , and the models that we build are kind of input agnostic they need to take on any sensor data , so sensor data to us is anything that will have any features that could be extracted from an input on a time series . So for us , that's kind of sensor data , so the data structures are the same . So then , okay , let's put audio in it and see if that works . Let's put galvanic skin response in it and see if that works . So it's about .

Speaker 2 30:00

I mean , we really the core product , that , the one that isn't seen on the website and that , but the core product that powers all this is what we call the human behavior data lab . It's a very ugly , long name , but it's . That's what it is , and it's purely meant for experimentation , because we don't know where these sensors are going to be relevant . What I do like about that approach , though , is that a client can come in . We can say great , all this survey stuff ticks the box for your measurement , the learning element . We're going to tell you now what's actually relevant , and then we could fine tune the data capture around what's relevant , and so that allows them . Again , this optimization process allows any business or any creator to really scale up their measurement . And if they want to scale up measurement , then they get that constant feedback loop , and that's that's how we expect to standardize the industry .

Speaker 1 30:51

One of the other big factors

How does human sensor data work from a privacy perspective?

Speaker 1 30:53

we touched on this very briefly earlier is privacy . I mean , I'd imagine that must be a pretty big consideration within this space .

Speaker 2 31:03

Yeah , my , I mean my early work . Right at the beginning we were helping the digital catapult feed into GDPR . So , coming from market research , you're kind of yeah , I mean even straight out of academia into market research . Your kind of privacy is kind of hammered into you .

Speaker 2 31:19

From an ethics perspective , what I love about facial coding data is is that it's not facial recognition . It's built on the same principles of facial recognition mathematically , but it's actually not facial recognition . So I actually don't need to know who you are to know how you're expressing and and it's because I literally just start to retrospectively normalize how you're reacting to this . This , this environment so I kind of liken it to a comedian on stage really has a clear feedback loop . They know when a joke fails and when a joke wins right . They don't know the individuals in the audience , but they know if their content is working or not . So that's what I love about this technology is it offers at the individual level .

Speaker 2 32:01

Our data agreement as a data controller is with the individual . So no matter how much a client might want to get access to the face videos , nobody has access . Even I don't have access to those . Nobody has access to the face videos . It is highly protected , that there are two people in our company who can access the face videos to maintain or to train , and that's it . And it is entirely separated , even by like proper server design . It's entirely separated from the rest of the ecosystem . So even if you hacked through the client portal to get the client data , it wouldn't , you wouldn't be able to get through to the other side of it . Well , I mean you might , but you'd have to work hard to get out of it .

Speaker 1 32:45

Yeah , it's an interesting point you make as well around the fact . Oh sorry .

Speaker 2 32:49

There is just one more thing , I apologize . And then one more thing is on the individual side , Because our relationship with the individual we just said you control it . We capture enough data that we can give a strong enough result to a business and have an acceptable loss . So if people want to delete their data , they . We give them a code right at the end of the survey that they can just email to privacy and their record isn't deleted forever .

Speaker 1 33:15

So oh , that's a very nice touch . I haven't heard about other companies doing that . I don't know if they are or not . I was just thinking about it . That's an interesting point that you make as well , that it's in some ways , you're not interested in who the individual is at all . In fact , the last not interesting . Yeah , the last thing you want is individual responses . What you want is aggregate responses .

Speaker 2 33:40

Yeah , we need what we need . For us , our focus is to make the individual stuff , the small rich data , as accurate as possible so that the aggregate data on top fulfills its need , but so that you also need less and less people and you can kind of synthetically fill in the rest . So the key , the key is then winning the trust of the individual and in that case it's giving them full autonomy and power over what they've provided and giving them the ability to delete it right away and , frankly , once we have our metadata , we're not carrying anything that's personally identifiable .

Speaker 1 34:13

Okay , matt , you're continuing my education . Blm's behavioral learning models . I think I've

Moving to behavioral learning models

Speaker 1 34:19

just about getting my head around LLMs . So what are behavioral learning models ?

Speaker 2 34:25

Well , just then , I would argue , the natural extension . Once you've got LLMs in play that can kind of accurately predict what the next word's going to be , or can actually generate the next word and synthesize kind of a thought , then BLM's are , but they work on , I would say , language is just relatively limited . Again , I use the love example is a very good example of this . But language is relatively limited . It's a , it's a synthesis and a summarization of our thought , whereas body language tends to be much more contextual . So BLM's is very simple , I think of it as simple as the body language physics engine for the future . So , taking the principles of LLMs and applying it to human behavior and emotion .

Speaker 1 35:13

And then how does that work then in practice ? So for this , like you know , an LLM is looking at what a combination of words and it's predicting the next words , and then it optimizes over time . And so for LLM at human , for BLM , what you're looking at , let's just stick with the face . You're looking at the physical reaction and then you're calibrating it against a certain data set and then so , again , calibrating , yeah , so you want to calibrate against the metrics that matter .

Speaker 2 35:48

The whole principle of the business is to kind of bring empathy to digital interactions or the internet and those interactions . So I would say nothing's ever going to be done in isolation of just the face making an emotional call . What you want to do is you want to make sure you have some sort of feedback loop in there , so some sort of pre cognitive , post cognitive feedback . So here's the emotional data set . And then , did they click a button ? Did they move through the website ? Did they leave the website , whatever it might be , whatever that interaction might be , did they leave the app ? Did they , you know , click on something ? That's the feedback loop . So you want to link it into something that that's relevant .

Speaker 2 36:27

So I wouldn't look at . You could try to look for a signal that is predictive on its own right , I would argue . Behaviors are very contextual , but they are contextual to an event that people have thought about and then done . So we're trying to find leading indicators to events that are commonly executed on in a digital interaction . So a survey is one way to do that , but frankly , I actually see one of the bigger use cases for this two in general .

Speaker 2 37:02

One is in UX , ui testing , so every website could put it on kind of like hot jar . It's like 1% of people that visit the website could actually flip on their webcam , or they have the element human app and they just pop it in . That's where we'd want to go . Again , future scope not currently , and then further , though , if we are accurate enough , we do see enough kind of interesting states like boredom , depression , things like that . We can get into mental wellness tracking as well , which again would be somebody interacting with their device regularly and being able to say , hey , look , you seem to be excitable here after these events , you seem to be depressed after those events . So it's about that what's exciting , what's draining , and being able to optimize your life that way . If you wanted to , or teach the machine to actually recognize this stuff so it can say , hey , it's time to stand up and take a break .

Speaker 1 37:53

Yeah , obviously that will be the next stage of it . You can very much see where that could go , though . Equivocally , as you said , we're monitoring your behavior , but it is about the interaction .

Speaker 2 38:03

Yeah , it is about the interaction . The interaction today is a survey , and it's a survey with an embedded experience of a feed like a TikTok feed , and so that's the interaction today and that's where that's our learning ground to be able to identify these other signals .

Speaker 1 38:22

Is anybody doing that at the moment ? From a UX perspective , I could see it's a very compelling use case of , as you say , just asking a percentage of people just to be monitored . You incentivize them to do it , see what they're doing and then start to track what the emotional reaction is .

Speaker 2 38:44

So I'd say there's three barriers to preventing that from mainstream adoption . The first one is technical , in the sense that you have to have XY coordinates from the eye tracking against the XY pixel or mouse coordinates or what's on the page and your areas of interest , and then be able to summarize the emotional state of the areas of interest . That is all assuming that you have a very good probability model on the emotional state , which I don't agree has really been nailed yet . There are two companies out there that I think are doing some really good work in this area . Yours is probably the most raw data set , so we're going to be building our own model around that stuff . We'll see . The second barrier to entry is trust . Frankly , I don't know if I would go onto any website , particularly with the way the data privacy has been abused in the past . I don't think I would actually go on and even say a consent to any website unless I had full control of the data that went over there and could retrospectively remove that control of the data and have full and final and total deletion . That to us , we think that that's an area we can own and that's a very interesting area and we like that area . We've played a lot there . We've got a pretty good ethics committee for review on that . We've said no to some very big financial gain opportunities to specifically not work in that space . We want to trust the technology itself .

Speaker 2 40:18

The third is the variability . It's hard to take one model's assumption that works for one website and assume that it's going to work for the next website . That variability also translates into human variability of people use different devices on it for different types of sites and different types of workflows . I think it's a much broader problem , even more variant and more broad than marketing . Marketing is that happy medium . It's not as isolated as trying to detect yawning from drivers and autonomous vehicles or whatever it is , or enhanced vehicles . That's a much more isolated workflow . The variability in UX , ui design , is almost intractable , I would say .

Speaker 1 41:10

That does make sense broadly . I'm conscious of time and I wanted to talk about the business actually a little bit as well , Not necessarily the detail of your journey , but it would be good just to get a little bit of an overview of the evolution of the business and then also what you've learned along the way

Learning from different business stages

Speaker 1 41:29

as an entrepreneur doing this business since whenever . I'm sure there have been ups and downs and you've probably got all sorts of insights to pass on to others .

Speaker 2 41:41

Yeah , I mean I think wow . Let's just say there's been four iterations of the business . I touched on this before . First was the algorithms weren't really accessible outside of a university . The first was plugging those into an API . That business was great , but there just weren't enough people building products commercially to create enough feedback for us to be able to get the insight . Today that might be a different story . The second one was then something that we packaged up as a demo solution called MIMO . That one was good for the broadcast industry , but the broadcast industry deals were still being done predominantly Rights holder deals were still being done by Handshake Agreement . So great , but it wasn't a big enough uplift or certainty for sales .

Speaker 2 42:32

There was a lot of resistance from salespeople of having the technology tell them which show to sell for the upside . We didn't quite nail that product proposition . The third one , then , was for the market research industry of giving them something that they could embed . That met the heavy skepticism , as you talked about , even those that would take it . It didn't create a recurring revenue model for us . They hated recurring revenue models because they don't want to carry a cost base that they can't guarantee that they'll use . A lot of market researchers are ad hoc sales , their sales cycle is ad hoc , so they would use us in their sales decks to win the deals , but they're not use us , ultimately , in delivering the goods . Now that one . Well , at each one of those the change came from staring at insolvency papers so going okay , we're going to have to put this business under , scale it back and redo it . Then . The last kind of this current product has been a very interesting journey because it was actually going the right way . It had a very strong product market fit . It was serving the marketers inside the media industry , particularly from the media owner standpoint , actually , surprisingly , whereas a lot of the attention providers were serving the media agencies . For MediaBind and I would say , on that one though , covid hit .

Speaker 2 43:59

After the first tranche of three funding hit , obviously everybody got skittish . Some tough conversations had to happen . I had to let go of a really good team and we scaled back . A couple of people stayed through . Then , after COVID hit , the recession threats hit and we were trying to go for series A funding and that didn't go through .

Speaker 2 44:26

So we had hit another reset button and at that time we were just like I'd say , the biggest thing is cut deep and cut once . If you ever have to make cuts , having been through a few rounds of cuts , it sucks . It's a very , very visceral human experience . But also surround yourself with the people who've got your back , even if you do fail , and I think that one's real hard . And so , like Diego , my co-founder through and through , has had my back and I'd say , if you get lucky enough to find a good partner through business , hold on to that . I think yeah , I hate giving advice I'd say that the biggest things that I would probably tell myself are always , always , always , no matter what you do , even if the business has to move in a certain direction , be people for it .

Speaker 2 45:20

Just be human about it . I've listened to a lot of lawyers and followed the legal process . Process does win the day , but , god , you can avoid a lot of pain if you're real human about it , and I definitely didn't nail it on a few of those conversations . I messed it up in a number of ways and that was a big one . Lawyers love them to death and make sure you always keep working on being the right fit .

Speaker 2 45:45

It's a marriage , because you're going to need the energy when you don't have it . They put rocket fuel in the tank when you need it and then also really understand your business stage . I didn't understand your business stage in the market . This whole idea of growth at all costs just never really gelled with me . But I tried to live it and I failed at every turn , only to when I went back to my core . This is what I believe in and the moment I did that , everything started to work . It's just really weird . It just started to work and again , I think that's believing yourself as a founder and all that type of stuff .

Speaker 2 46:24

I mean you're told no , you're an idiot and your ideas are dumb and you're selling snake oil and all that type of stuff and you're like well , I still think it's . I still think empathy and body language are . Empathy is really important in a world of the internet and I still think the body language is probably our key identifier of how to empathize .

Speaker 1 46:44

Well , thank you , matt . I know it's been quite a journey . I also know it's on an upward swing and I think will continue to be so . Now talk about an upward swing and we should jump onto a quick far round . But

Quick fire round

Speaker 1 46:57

I'm gonna invert for these questions , ladi , because I happen to know that you got married fairly recently . So I don't know I don't guess you know what you think your best and worst characteristics are . I wanna know what your wife thinks your best and worst characteristics are .

Speaker 2 47:13

So , in anticipation of this , I went . I said I know , henry , he typically asked some of these questions as a listen to your podcast , and so I actually had a conversation with her . So the best characteristic , she said positivity . She's like , you're sickingly positive . And she said , like , even in the middle of the night when you can't sleep , you'll turn around and go yeah , but it's okay , cause there's so many interesting things going on , and so you're just like . She's like almost to the point . That's annoying , but it really helps . She said your positivity is just infectious Resilience . She said boy , do you know how to take a knock and just get back up again and keep going . So , yeah , she said that .

Speaker 2 47:59

And then she said the third . She said it's kind of a two-parter . It's the best and worst is your kindness . You really want to help other people , but you do it at the fourth , which is the worst , at the cost of your own time and therefore time management . You are rubbish at time and I am rubbish at time management , absolutely rubbish at time management . Thankfully , I've got a great board and I've got great operators around me who are not rubbish at time management , so it's always something that's constant work . And yeah , I would say . So there you go positivity , resilience and , I would say , kindness on the two fronts . And then I put other people's problems and I like to solve those before I solve my own , and then time management is horrendous .

Speaker 1 48:44

Yeah , I think it sounds like a pretty fair summation . It'd also be interesting to ask that question like in five years ?

Speaker 2 48:51

Yeah , I will . I've written down the answer , so we could do that .

Speaker 1 48:56

A couple of final questions . What do you wish you ? What do you know now that you wish you'd known , say , 10 years ago ?

Speaker 2 49:05

Validate , validate , validate the problem . What does that mean ? I just got really excited about . I mean I knew in my gut that there was a problem , but what I didn't do is I didn't spend enough time speaking to people in the market and I spoke to , like I went and calculated the other day it was about 1,600 , 1,700 people in the first five years and so that still wasn't enough . We didn't isolate and I got caught up in like trying to find a way in to make this boost . I was more , it turned out I was more interested in trying to bootstrap the business than I was actually solving the problem and recognizing what was required to solve the problem , and I hadn't didn't really clock that until the last four years , so I could have saved myself probably about six years worth of time and effort .

Speaker 2 49:54

Now we had moving markets . There were things that were just out of our control and things changed . You know the Mike Tyson phrase . You know everyone has a plan till you get punched in the face . But spend more time validating . I think just take your time . Actually , the rat race isn't there . There's very few products that need that . Where there's a true race to market , in a lot of cases if you build it right before you rush it , you're gonna actually end up with more markets here on the longterm . So I think A think longterm and B just validate . Don't feel pressured yeah , don't feel pressured to try and raise money or do whatever . Just validate and actually take that signal of failure is actually a good thing and learn from it and keep going .

Speaker 1 50:41

So , matt , it sounds like you weren't necessarily rushed , though , in creating the product and , from even here , you're incredibly knowledgeable around this space and that you know the intricacies of the products very , very well . It almost sounds like you were assuming there was a problem that needed to be solved , which maybe how can we ?

Speaker 2 51:04

I was assuming the problem was recognized . I was assuming that it was more obvious than it was . And it just wasn't . And it was actually the second order , third order level of the problem . That was more of the problem we should have honed in on at the time . So the root cause remains the same of the problem . And the root cause is not going away , it's got .

Speaker 2 51:30

I think it was actually interviewed a product manager today and I think he summarized it very well . He said you got all the ingredients for the recipe but the bread isn't rising . And why is that ? And that was very , very real up until about a year and a half ago , two years ago , where we isolated a problem where we could be of true value , differentiated value , not just like nudge the efficiency a little bit , but really really say here is where behavioral metrics can really fundamentally make a visceral difference that the user can feel in their day-to-day lives . And that's where that's what we had had not really clocked . Initially we thought accessibility was the issue in the first one . With the API it wasn't . It was that the real product problem was out there . But I mean we weren't the only ones . Our competitors I mean Aftik Diva raised what ? 65 million in the first five years of our company and Realize , raised 45 million and frankly they were dealing with the same problems . The market just wasn't there yet .

Speaker 1 52:39

So I hear you Listen to the market .

Speaker 2 52:42

I validate you the market .

Speaker 1 52:43

Yeah , I hear you . So we're at time , I'm conscious . So final question what's your favorite almost impactful book or recent book Could be a piece of media , it doesn't have to be a book .

Speaker 2 52:54

Yeah , favorite one Bill Browder's red notice . So hang on , what's that ? I don't . I've never even heard of that . Oh my God . It's like it reads like a spy novel . Bill Browder was a Jewish American entrepreneur who went to Russia during the nationalization and then privatization so the Yeltsin era and set up , did very well , very well I like super smart numbers guy but then got caught up in the Putin era and I'm not going to give away anymore . It is worth . It literally reads like an international spy thriller . It is so good , but that's all real and it's yeah . Have mad respect for that dude . He's a talk about a real entrepreneur , so yeah .

Speaker 1 53:46

Matt , that sounds like a very good recommendation . I get a few of them . Most of them , have to confess , I don't read , but I will put that one on the list .

Speaker 2 53:53

This one is like a holiday read , like you can actually put it onto the fiction level , but it's just nonfiction . It's amazing .

Speaker 1 54:00

Amazing , amazing . Matt , thanks so much . Great as always to talk to you and thank you for putting up with my annoying , panicky questions , but I learned a lot .

Speaker 2 54:12

No , I appreciate it , henry , and it's . I love listening to your podcast , so thanks for having me on .

Speaker 1 54:19

So did you get all that ? Matt's a super smart guy . He's very lucid and always thought provoking . Even though I'm pretty familiar with the business , I actually went back and listened a couple of times , in particular to some of the explanations around behavioral and sensor data and how it may affect the industry . As I mentioned , I'm on Matt's board , so I may be a little biased , but I think he's a pioneer and a visionary in this space , and Element Human is a company that's going places . On that note , in my next interview I'll have on Catherine Topp , the CEO of Yabbel , one of the leaders in the synthetic and augmented data space . Also , there's a skip close on . Thanks , as always , to Insight Platforms for their support and see you next time . Rewarded by wee-wee .