Experimentation at Scale: a Q&A with Serg Masís from SyngentaJuha Kiili
Our product lead, Juha Kiili, talked to Serg Masís, an agronomic data scientist at Syngenta, on why and how Syngenta is using Valohai for experiment tracking. Here is a summary of the conversation.
Syngenta is a leading provider of agricultural science and technology, and it is particularly focused on seed and crop protection products. The goal is to improve global food security by enabling millions of farmers to make better use of available resources.
Syngenta set out to predict plant growth at different stages. That is, the firm would like to know how long it takes for a plant to move from one stage to another. This is critical for farmers to understand when to plant, the type of crop to plant, and when to apply crop protection products. For this, Syngenta's data scientist, Serg compared two initial approaches; time-series model and aggregation method. The data features used for the prediction of plant growth in one stage can be different from the features used in another stage. Some of the data features include sunlight, precipitation, soil properties, etc. The two approaches were compared based on a single use case i.e. single crop or a single region. In the end, the aggregation method yielded the best result. However, there was a challenge, Serg noted:
"Once we started using the same approach (aggregation) for half a dozen countries and several crops, I started to doubt myself. So, that's when I thought I have to make an experiment at scale; I can't be waiting days for these results… that's when I used Valohai to build a pipeline".
Figure. Syngenta Experimentation Pipeline in Valohai.
Serg built an experimentation pipeline with Valohai to launch massive searches and summarize results. He ran two approaches, LightGBM and LSTM, in parallel. In the end, he figured out that LightGBM was better than LSTM. In addition, the experimentation pipeline can be reused for future projects by colleagues at Syngenta.