For AI to succeed, it needs to fail first
There’s a lot of talk about ethical and bias free AI. And it’s only logical, if you compare a machine to a human, learning goes through the same cycles.
Take a child as an example. If you teach a child that a fork is called a spoon, the child will call the fork a spoon. It’s a reality it’s been taught. Once other humans start correcting the child, the spoon will become the fork (and even that can be hard to achieve, depending on who it was that taught the child that the fork is the spoon in the first place, parental influence goes a long way). Given that I’m maniacally focused on etiquette at the dining table - I eat pizza and burgers with knife and fork (not spoon) much to my partners chagrin, let’s leave that example of the table and focus on another example.
Research has shown that for facial recognition, it’s harder for the machine to differentiate between black women’s faces compared to differentiating between white women’s (or men’s faces). See this great Wired article from last year - https://www.wired.com/story/best-algorithms-struggle-recognize-black-faces-equally. An extract from that article: “At sensitivity settings where Idemia’s algorithms falsely matched different white women’s faces at a rate of one in 10,000, it falsely matched black women’s faces about once in 1,000—10 times more frequently.” To add to that, the more recent controversy about Twitter’s algorithm on black and white - https://www.theverge.com/2020/9/20/21447998/twitter-photo-preview-white-black-faces only strengthens the case that facial recognition is definitely not ‘there yet’.
Research on facial recognition and the skewed results are longer term goals, and several reasons, none conclusive, have been named. Photos and quality have been optimized from the starting days of photography for white people, there are more photos of white people on the web to learn from than there are of black people and a more controversial statement ‘that black faces are statistically more similar to one another than white faces are’.
What the examples on facial recognition show is that first of all we’re treading a very thin line between teaching the machine while being discriminatory, yet not to discriminate in the racial sense of the word. However, specifically for this example, the machine needs to discriminate in the individualization of results. And that’s where learning comes in, constant learning. Because it’s men that teach the machine and the machine will accept that the fork is a spoon.
Moving on to my own domain, that of Marketing Automation & Engagement, for AI to succeed to convince there are tons of solutions available today. Some work better than others and the sum of all doesn’t necessarily equate to better results. A few examples of AI techniques used today in Marketing Automation;
- Lookalike
- Next best offer / Next best action
- Lead Scoring
- Subject Line Analysis for communications
- Channel Propensity
- Sentiment Analysis
- ROI and Cost modelling
- LTV (Life Time Value)
There are many many more, and no doubt over time new approaches and techniques will find their way into Marketing Automation platforms, to replicate what your local butcher knows about your taste and his way of getting you to buy that Wagyu steak you hadn’t tried before (as he knows you found it too expensive) - by adding a dozen eggs from his local supplier at no cost. Suppose the butcher learns that you come less often and your buying pattern has changed, and he takes into account climate change and the impact agriculture has on our environment, isn’t it time for the butcher to start to offer plant based products? It’s fascinating to see that real life branding and marketing for plant based meat still uses the word ‘meat’ in all its marketing. Examples from https://www.beyondmeat.com/products/the-beyond-burger/
- Beyond Meatballs
- Beyond Breakfast Sausage
- Beyond Beef
- Beyond Sausage
- Beyond Beef Crumbles
My point of view is that this approach again deals with learning. Consumers need to transition (be transitioned) from meat to plant based. A hard liner approach will not reach the general population, vegan’s and vegetarians already had their mind made up and know what to buy. But it’s the regular consumer, the frequent meat eater that needs to be convinced to change their behavior.
Machine learning and AI are expensive approaches, which over time will definitely lead to cost savings, ROI improvement, better advise and better engagement, and thereby better customer experience. But good AI takes time and testing, extensive testing. In order for AI to succeed, it needs to fail. Data scientists, marketers, researchers and consumers all play a crucial role in this approach. The human teaches the machine, the machine tries, the machine fails, the machine learns, the machine tries again. It’s the last three steps of trying, failing and learning from that failure that makes for great AI. By setting the premise that a fork is a spoon, without corrective action, AI will never learn.
All photos are my own. Please share your point of view! I love to learn.