Heuritech launches the Lab – an Artificial Intelligence dedicated to Ecology
At Heuritech, we are deeply aware that climate change is caused by human action. We, as individuals and as a company specialized in artificial intelligence, believe that not taking ecology into account, would be an act of self-delusion.
As a fashion tech start-up, we understand the environmental footprint that fashion has on the planet – the production volume of apparel has doubled in the past 15 years and it is often stated that fashion is the second most polluting industry. However, we want to be part of the solution, not the problem. Our product aims to forecast demand and trends more accurately for brands to produce more sustainably, with less waste.
A sustainable lab at Heuritech
In parallel to our main mission in fashion, we decided to create a Lab dedicated to modifying our behaviors and pushing forward our machine learning technology towards climate change initiatives, not necessarily in fashion. We believe that our technology can be beneficial to ecological initiatives in many fields.
Today, machine learning makes a lot of sense in many large-scale commercial applications. However, scientists and engineers have spent much less time trying to design systems dedicated to fighting climate change – or at least taking it into account. We are convinced that Machine Learning, like many other technologies, can be used to shape a more sustainable world, if we pick the right battles.
This article presents several initiatives that we are launching at the Lab. We have decided to focus on actions with actual impact, being aware that no one solution will answer to the scale or complexity of the problem. Take a step in the right direction with us to initiate awareness, and to make meaningful changes – yet, very humble ones.
Heuritech’s sole existence is energy-intensive. Servers consume a lot of power to train Artificial Intelligence models and analyze images. Before starting any environmental project, we decided to audit our everyday work and its production of Co2, to understand which part is the most gas-guzzling.
Source of CO2 emission at Heuritech
Our first (approximate) audit on main CO2 emissions generated by Heuritech1.
What concrete actions can we take?
Firstly, we aim to limit flights by grouping meetings, using trains, or using videoconference whenever possible. It seems obvious, but it’s far from easy when clients expect you to come see them in person worldwide. When impossible, we will compensate with carbon offsetting by planting trees (not a real solution, but better than nothing).
Secondly, we will control the amount of energy used by our computers, as GPUs (Graphics Processing Unit) famously need significant power to run. We have decided to primarily have our machines in countries where power sources are known to be greener: for instance, France has low Co2 emissions per kWh2. In addition, we aim to use smaller, cheaper and lighter networks3. We think that, energy-efficiency should be a de-facto metric for comparing different deep learning models (in addition to accuracy and model complexity).
Lastly, we aim to foster an ecological spirit within our team, through discussions and debates as well as supporting all ecological initiatives. This already translates through our commuting practices (we encourage the use of public transportation or biking) and specific recycling projects (using coffee grounds to grow mushrooms).
Applying Heuritech’s artificial intelligence technology to ecological projects
We announced in September our collaboration with NGO Surfrider Foundation Europe. Since then, several Heuritech members have been dedicating a day per week to the project. We take advantage of our image recognition experience and technology to detect plastic pollution in pictures of rivers taken by Surfrider members. By automatically detecting and geo-localizing plastic on river banks, the project aims to map out the pollution in French rivers (and more). Measuring plastic pollution on a large scale is nearly impossible today, and has a colossal impact on local measures and larger political actions undertaken by Surfrider. All our contributions are open source, here (the content will be updated regularly).
This is not restricted to Computer Vision, as most of the technologies we work with at Heuritech could be linked to climate change problems: our Times Series Prediction methodology could be applied to electricity or thermal energy demand forecasting, our worldwide trend detection and comparison methodology can be employed to monitor ecosystem populations etc. More to come soon!
Putting ecology at the core of Academia and Education
As scientists, sharing knowledge is one essential aspect of our daily work. We publish scientific papers, technical blog posts, organize conferences and teach in French universities. We could take this as an opportunity to shed light on ecological problems.
Indeed, Machine Learning is driven by data: to fuel our research and train our models, scientists use datasets. Today, they have the choice to use datasets that seek to have a positive impact. One example is the iNaturalist dataset for instance, which aims at monitoring ecosystems and biodiversity from geolocalized pictures: it could replace standard Image Classification datasets such as Imagenet.
There are so many important and interesting challenges within ecology that we believe research and education should focus on these topics first and foremost. Public and private entities such as ours should commit to releasing fully open data and code that may be relevant to climate change mitigation, so that these become more accessible to students, researchers and other entities.
Finally, we will dedicate our next “Deep learning Meetup” in Paris to applications within ecology, date to be announced by the end of 2019, so please join us!
We decided to open the Lab because we feel a responsibility to have a more positive impact on the environment. We acknowledge that it is a humble step, and does not make us a “green” company. We’d be pleased to be challenged, criticized, or to start new partnerships. Get in touch at email@example.com!
- This chart was made using several assumptions (Blog post to come; Feel free to ask for source / methodology). Examples: Paris-NY for 1 person is around ~1.4T Co2eq ; For servers Estimation of kWh usage per month times average european carbon Co2 emission for electricity 0.3kg Co2eq / kWh.
- HuggingFace for instance recently released smaller footprints networks for NLP https://medium.com/huggingface/distilbert-8cf3380435b5