AI is proving its worth by underpinning operations in a range of sectors. Fast-moving consumer goods (FMCG) for example, is already being revolutionised by the innovative possibilities that AI can bring.
Whether it’s utilising data to provide accurate forecasts of consumer demand – ensuring inventory is efficiently stocked – or implementing chatbots to improve and personalise customer engagement, AI is transforming the FMCG sector.
If wielded properly, the future possibilities that could be achieved by AI are boundless. In order to get an AI that works for your business – an AI that is not only strong but also affordable – following at least one of the following three key practices is essential.
1 – Transfer learning from one AI to another
Transfer learning is a process that takes a pre-existing AI that is already strong in one general area and tailors it to fit your specific requirements.
For example, there are AIs that can take different images and identify what is in them. They are trained by learning distinguishing patterns, called features, from a vast set of pre-annotated images. Then they rely on those learned features to classify unseen images.
These AIs may be used for applications like automatically captioning photos. But this pre-existing learning can also be modified and transferred to other uses, such as classifying products.
What if a fruit packaging business wanted AI that could identify the difference between Braeburn apples and Gala apples, or the difference between ripe and unripe peaches? Instead of starting from scratch, they could exploit the previous training an AI has with image recognition and refine it to fit their particular needs or requirements.
Because the AI doesn’t have to learn a plethora of brand-new information, the business can get their application off the ground faster — and more affordably too.
2- Include domain-specific expert knowledge
Often AI is an advanced form of pattern-finding. So, by being specific to the AI about the kinds of patterns you need to be matched, it will be much stronger at finding them.
Suppose you want an AI for optical character recognition (OCR) of serial numbers on scanned or photographed invoices. Suppose you also know that the serial numbers will follow a certain pattern of letters and numbers, such as those within a national insurance number or on a registration plate. An AI built to specifically recognise that kind of pattern will perform more reliably and accurately than just using a general-purpose OCR AI.
By taking care to incorporate what you already know about the data, you can achieve the best possible results. After all, strong, affordable AI does not replace your expertise but retains and builds upon it.
The same is true when teaching AI to extrapolate patterns into extreme situations it hasn’t encountered before.
This time, let’s say a supermarket is forecasting the number of deliveries it will make in a week and is using an intricate machine learning model to provide this forecast. Many such models will perform well when given inputs that are similar or within similar ranges to those already seen. But they may become unpredictable or unsuitable when you go beyond the kinds of scenarios the model was trained on.
For instance, the waters become muddied once the AI is put into a scenario for which you don’t have any data yet, such as being in a new town, or delivering to twice as many customers as normal. An AI not built with this kind of use in mind may perform poorly in these scenarios and may provide unreliable estimations.
If you’re going to want an AI to perform in these uncharted territories, this will fundamentally affect how you build it. You need to consider what kinds of patterns it’s assuming, preparing it to extrapolate beyond where it has any previous observations or data to work from. In other words, it’s a design-time decision to build an AI that extrapolates in a reasonable and useful fashion.
3- Feed your AI high-quality data
AI arises from teasing out given kinds of patterns from a given cache of data. The more general the patterns are, the wider the possibilities of the AI become. But not only does this mean you need more data, it may also become more of a black box, an AI with opaque inner workings.
AI only knows what you tell it, meaning it only knows the data you give it and the assumptions it makes about how that data behaves. The more relevant the data is to the task at hand, the more reliable the conclusions.
Say a beverage company is launching a new drink. They won’t have any data, related to purchasing behaviour, on that particular drink pre-launch. But they will have sales pattern data from other drinks they’ve sold. It is these historical patterns that the AI can work with.
The same can even be applied to a young company that doesn’t have any historical data at all. It’s about finding ways to incorporate what you do know from sources like market research and giving as much relevant data as you can to the AI.
It’s not as simple as just drowning the AI in any old data, though. It needs to be useful data. The best place to start is to consider what data you already have and what it is you’re looking to achieve.
On one level, building strong and affordable AI is about working with what you’ve got. The decisions you make should be influenced by how you will put that AI to use.
So, depending on your context and resources, you can train a generalist AI on lots of high-quality data, or design the AI to retain and build on your domain expertise, or use transfer learning to make the most of general AI learning. The design-level decisions you make here can, in time, lead to stronger AI-informed decisions for your business.
Get in touch to see how our bespoke solutions can help you accelerate innovation with AI.