Leonardo Palma Batista
The challenge is twofold. On one hand, we want to leverage AI to make processes and systems more optimized, like nyris and its visual search solution for spare parts and products. AI helps our users identify what they are seeing with just one click. But on the other hand, we need to ensure that the AI models we develop are themselves efficient and environmentally friendly, because two thirds of the energy used to run applications in your phone, doesn’t come from your phone. The energy comes from servers which are running our models on the cloud. The problem is, there is no cloud, it's just another computer, and that computer uses the same carbon based energy like everything else.
Our approach to this challenge is to treat it as a search problem. We sift through our database of information to find the right solutions. This involves extending our search capabilities and upgrading information to provide our customers with the most environmentally friendly models.
Models play a crucial role in our business. We use neural networks to compress image data into code, which we then feed into our search engine. This dynamic component of our pipeline serves diverse customers and requires constant updates to ensure satisfaction.
Updating and replacing models is a cyclical process. It involves gathering requirements, collecting the right data, training and validating models, and finally deploying them. However, this cycle is expensive in terms of time, resources, and environmental impact.
To simplify this process, we've developed a method of cherry-picking models. This involves two levels of selection. The first level involves selecting models that are compatible with our machine learning framework. The second level involves a more specific selection process based on our problem statement.
We employ a clustering analysis to further refine our model selection. This involves working with image embeddings and evaluating how well these embeddings are clustered. The result is a shortlist of models that perform well in our specific use case.
Once we've shortlisted our models, we put them through a fine-tuning process. The results have been promising, with a significant reduction in the number of models required for experimentation.
Our initial experiments have been encouraging, and we're excited to apply this approach to more practical problems. Not only does this method simplify the model selection process, but it also saves time and resources, making the process more sustainable.
We're keen to quantify the impact of this approach on our mission and extend this study to other domains. We're also interested in exploring how this method can be applied to other types of data, such as 3D data.
In conclusion, our approach to green AI is a testament to the power of sustainable practices in the field of machine learning. By simplifying our processes and making them more efficient, we're contributing to a greener future for AI.