Episode Summary: How can machine learning help us advertise through social media? In this episode, Thomas Jelonek, CEO of Envision.ai, talks to us about how in the next five years, machine learning might automate the laborious guess-and-check process of finding visual content with which users can engage. Right now, finding images and videos that will best generate engagement is a task reserved for a human. He or she shifts through images and video clips that may work for an audience based on anecdotal evidence and perception of past post success. Learn how, according to Thomas, machine learning could help you save time and money, generate you a better ROI, and build you a larger list with more accurate targeting on social media.
Expertise: Computer vision
Brief Recognition: Thomas Jelonek is the CEO of Envision.ai, a company specializing in the use of AI for curating visual content for social media and other platforms in an effort to maximize engagement. He is also the founder of Foreideas, Inc., a company which seeks to provide users with related content from images with which they engage. Thomas did his doctoral studies in computer vision at McGill University.
Big Ideas:
When you need images and video clips to post to social media, you probably hire a freelancer, or maybe you have a person in-house, who scrolls through troves of different images or video clips you’ve acquired or that are in the public domain. How exactly are images and video clips ultimately selected for posting? Chances are you have to interpret the data you may have on what makes certain visual posts engaging for your audience (if you have any data at all), but you’re never really sure if your post is going to take off. Thomas believes machine learning may help with this process in the near future.
Consider that the curation process, the physical sifting through images and videos, is not a task aptly-suited for humans. There’s no real way for someone to sift through stores of thousands of images and videos you may have acquired in a timely manner in order to then determine the best one to post on social media. Thomas suggests AI could prove useful after building data sets detailing engagement statistics. The software could scour millions of images in a short amount of time, cross-referencing those images with the dataset to find patterns that would ultimately provide an answer to the question “What do I post to garner the most engagement?”
According to Thomas, machine learning may even be able to help with brand-building. Thomas believes that in the near future, advertisements will further become less obvious. We’re already seeing some of this, with companies integrating lifestyle messaging and entertaining visuals that seek to engage the user over directly asking for an email or pitching a product. Thomas claims that, in the near future, AI software could scan the Internet and caches of images and videos to build content around a brand story. You may be able to find an image or video related to a specific niche quickly based on data sets that indicate which images and videos performed well in the past given the niche in terms of engagement. In addition, Thomas suggests AI software might help in targeting audiences better, so that, as he predicts, users may even discover interests they hadn’t even considered based on their engagement behavior.
Turning Insight Into Action:
Think about the goal of your social media posts. Do you want the post to generate leads, garner page likes, or sell a product? You can do some preliminary work on trying to find images and videos that you think would help you meet your goal, but consider this preliminary. It’s no good to work on anecdotal markers of success when you choose which images and videos to post. Collect data on how users engage with your posts. This data will serve as the basis for any AI you implement to help you later. It may be tedious, but investment in a solid spreadsheet of engagement data could save you time and money well into the future.
Interview Highlights – Machine Vision in Marketing
(3:06) Dan: What’s possible today at the intersection of computer vision and marketing?
TJ: Video is increasing by massive amounts in terms of the type of content people are using for advertising, and it’s clearly some of the most engaging content people can see. When people see advertising, they’re often flicking through a social media feed fairly quickly, and so we ask the questions: how can visual content make an impact when someone is only going to see it for a couple seconds?
How can you get people to pause their scrolling and click on a video? What is the best way to maximize engagement? What are you goals? Do you want to maximize click through rate? Gain followers? You have two problems: how are you going to manage all of the content you have, and how are you going to maximize that process depending on what your goals are? How do you get people to follow a brand or watch a video? With deep learning and AI, the classification and ranking of this type of content has become quite accurate, so things that weren’t possible a few years ago now are within the realms of possibility.
(5:56) Dan: Picking which clip to rotate on Facebook or Instagram is not something a human curator can do well. If I’m guessing correctly, a human would need to start with a whole bunch of different split-tested clips based on some algorithm trained to coax out what action looks like in a video and rotate that across different sub segments to coax out who should be matched with what video clip in order to get them to watch it. Is that correct?
TJ: The reality is it’s not very sophisticated in terms of what’s going on right now. Companies go to someone and say, “give me some images.” There’s a list of a hundred million images they have to go through. How do they choose which photos to go together with your content to achieve the goal. That’s a really hard problem.
(9:00) Dan: So, in the future we may be looking at a world where we don’t have a human curating two or three images we want to split test, we have a machine curating three hundred and rotating them across two thousand different subsets of our audience and coaxing out as optimal an interaction as we can across all these various segments.
TJ: With a lot of these techniques you can do constant monitoring and see every couple seconds how people are engaging with this, what is working, what is not working. It starts becoming very multidimensional because you start thinking not just of the content but of the meta data you need to marry it with.
(13:31) Dan: Are there further applications for machine vision in advertising that are conceivable today?
TJ: I think what you can do is monitor what’s going on in the market on different social platforms and take the advertiser’s content with what they’re goals are and marry them together. Then you can learn how behaviors are going on over time, build some data set you can learn from, and achieve lift right away. You aren’t guessing from ground zero, whether it’s getting sales or followers for your brand.
(17:05) Dan: Is it possible the future of social will involve massively high competition when computers are trained to make sure that the most targeted images and videos are in front of everyone? What is the future of social media and advertising?
TJ: Certainly you’re going to see sampling almost continuously. You’re going to see what works in real time and be able to make modifications. You’re constantly in real time building the data sets you’re creating to learn from. I think what’s going to happen is there will be advertising more intermeshed with all kinds of other content. The message will be embedded with all kinds of other content to bring you there, and it might not even be clear in the beginning that you’re looking at an advertisement or product placement.
(19:46) Dan: What is the future of social going to look like as these technologies proliferate and as optimization becomes more and more possible?
TJ: Imagine you go to a brand who has some things they want to sell, and then they start taking bits of content from all around the web and build their own story around that. If they make it interesting enough, people know they’re going to a brand or some advertiser, and they stay there because there’s an interesting story to tell. It can be compelling. A brand can build an interesting story around their products, and people stay there and become really engaged. You’re going to see AI systems automatically building a lot of this content around this brand for the story.
(26:43) Dan: Is there anything else that I as a user would notice is different about the way I’m engaging with advertisers?
If they have some very specific interests they’re going to start finding that much more quickly. As people understand how to use some of the meta data associated with that image content, people are going to find things they’re really interested in. A lot of things which you don’t discover now, you will discover then.
Header image credit: LinkedIn
