In a little suburb in Melbourne, a man has an espresso and smoked salmon bagel in the morning, makes a salad for lunch, then heats up a store-bought frozen pizza for dinner. In another part of the world, a woman has nasi lemak (rice cooked in coconut milk) for breakfast, chicken rice for lunch and pork knuckle stew for dinner.
It may look as if they live similar lives — they both eat three meals a day — but when it comes to the trash they produce, the picture suddenly looks very different.
And when a trash-sorting robot equipped with computer vision and artificial intelligence looks at this picture, it may not always be able to figure out what it’s looking at.
Artificial intelligence for sorting trash
More people are becoming aware of the need to recycle. However, this awareness hasn’t caused a reduction in consumption. Waste continues to pile up and it’s not always sorted in a way that’s suitable for the recycling process. Glass and cardboard and recycled differently but are often thrown in together.
The good news is that waste management companies are starting to introduce AI to tackle this massive feat.
ZenRobotics provides AI-powered robots to sort trash. Waste Robotics provides robot-agnostic software and (if required) robots to processing facilities. AMP Robotics introduced a series of intelligent robots that can sort, pick and place material to achieve 99% accuracy. Greyparrot analyzes and identifies trash on moving conveyor belts, providing waste managers with new insights for more efficient waste management.
The introduction of AI for processes like waste management is a good thing for our planet. But these companies have their work cut out for them. As exciting as it seems, waste management is also daunting in other ways besides just the immensity of the task.
It’s already a big challenge for companies to manage trash in their own regions just because of how large the volume is. But if they’re looking to expand beyond their current regions of operation, things get even more challenging.
Product and brand availability, daily living habits and routines — these are little factors that add diversity to the waste that people from around the world produce.
In Malaysia, we’re more likely to see yellow Mamee Monster bags in the trash, instead of the orange Cheetos ones. Durians are probably more common in Malaysian dustbins, compared to those in the United States. In some countries, people use banana leaves as food wrappings. In other countries, they may use newspapers.
It stands to reason then, that when it comes to developing waste management systems, different countries would require different data sets for training the AI algorithm. Diversity in the data labeling process is a common issue for most AI solutions across multiple industries.
The cost of accuracy
Currently a subject that’s being examined closely, accuracy in AI and machine learning is often drawn back to the need for diversity in every part of the AI development process.
Besides the ethical implications, taking diversity into consideration will enable the development of more accurate AI machines.
Imagine a waste sorting machine that has never encountered a Mamee Monster pack before. Sure, it may be able to tell that it’s some kind of packaging. But what if the pack is squashed up? What if it’s been turned inside out? Would the machine still be able to tell?
One limitation that’s often brought up when considering expansion to other regions, or increasing diversity in the AI development process, is cost. It’s a valid concern; collecting and processing the data takes time and requires human resources.
If you want to go far, go together
One way to overcome this limitation is by working with a partner in the relevant regions. For example, an AI company wanting to expand into Southeast Asia might need training datasets that are accurately labeled by a local team. This US-based company could partner up with a company that has a significant experience in this area.
Besides having a team of data labelers who already understand the data labeling process and don’t have to go through an unnecessarily lengthy training process, this partner might also be able to provide additional insights on the local market.
And instead of building multiple local teams or working with multiple data labeling companies, AI companies might want to look into getting a data labeling partner that already has a diverse team that may be spread out across a single region.
For example, although Supahands may be based in Kuala Lumpur, it has a diversified workforce that is spread out around Southeast Asia.
Rather than hiring separate teams for each country, a company looking to expand into waste management can work with Supahands and achieve the required level of diversity.
Although many sci-fi books and movies tell stories about how robots are going to take over the world, we’re still a long, long way from that. We first need to figure out how to train these robots.
Interested to see what we can do for you when it comes to preparing diverse computer vision training data sets? Get in touch!