Perfumes and Artificial Intelligence?


Hi, I’ll start my personal blog with this post, and I will be posting more about this perfume project called Epoch, that I have been keeping under wraps for a while over the next few weeks.

One of my hobbies is making perfumes. I love the possibilities of learning about what makes a perfume a perfume, and a scent a scent. You get surprising results from mixing different essential oils and aroma-chemicals. There are thousands of these aroma-chemicals available on the market, and you need to learn about as many of them as possible if you want to be a proficient perfumer.

I also love digging into new technologies and finding new interesting uses for them. I have worked over the last couple of years on deep learning projects for applications in the fields of cosmetics and health. You quickly realise the applications are basically unlimited, as long as you have enough computing power and data. You also realise deep learning can quickly become a lengthy and costly business as you need access to a GPU computing platform to speed things up and process that data quickly.

Consider how difficult and a lengthy process making a reasonably good perfume that most people would like is. It can take weeks or months to get the result you want as aroma-chemicals react with each other over time. Most perfumers will agree that as much a process of patience as it is an art, as there are no real shortcuts you can take. Experience helps shorten this process, but you will rarely get a good result instantly (this documentary will show you how much of an art form making perfume can be).

In short, you mix your samples. You are happy with the end result. You then leave them alone for a couple of weeks.

You then go back to them, and find out they have changed more than you expected…. and not always for the better!

You tweak the formulas. You wait again. Repeat the process until you’re happy with it. You then test it, and realise it doesn’t last longer than a couple of hours. Or it gradually turns sour, and complete changes over the course of a day.

You tweak again. You wait again…

It is definitely a game of patience. You gradually learn what works and what doesn’t and how perfumes evolve over time.

It then hit me. I started considering whether you could teach an AI to mix a perfume, and help shorten this lengthy process? I have a lot of notes on what works and what doesn’t as a starting point. I could ask people to smell randomly mixed samples and get their input.

That’s exactly what I ended up doing.

I trained an AI with this small data set I created and decided to let the AI come up with a new perfume recipe. I mixed it… and the result was awful.

But I then still asked people what their feelings were, with surprising results — some people actually liked it. I proceeded to take the information I collected to further train the AI. I let it come up with a new recipe. And repeated the process.

I have been doing this for a couple of years now, and I'm now getting close to a perfume that I would consider smells acceptable to the rest of the world. There were some interesting iterations of it during this process, for sure. Some of them were unremarkable, but some had an amazing sillage that I didn’t think possible considering the ingredients involved.

I am definitely on the right track and it's time to share this with the rest of the world.

I intend to release the first iteration of this AI-based perfume in the course of 2019. Each iteration will be numbered in the way a trained AI is numbered (using Epochs, giving its the name to the perfume).

As part of the game, users agree to give their feedback to improve the next iteration which will be released a few months later..

I want you to be part of this experiment, and you will get to experience how the perfume evolve over time. There is only one caveat. This experiment will only work if users give feedback…

You will be given a way to send your impressions in exchange for a discount on a bottle of future iterations of the perfume, and I also intend to make the results publicly available and share some of my research.

Sign up to the Epoch newsletter below to receive updates about this project, and to be given the opportunity to get your hands on it very soon.

Epoch #001 is intended to be released around June 2019.

Alex Seewald