3 ms·

> Fast forward to today - the tech is available and the team is assembled to make PlantSnap a reality. With a beta version completed, and 250,000 images in the

• ovi256 9y ago
  To help others understand why, the "Deep Learning" book, which summarizes current state-of-the-art DL, advises having at least 5000 samples (images) per class for OK performance (equal to non-DL approaches) and 50k to 100k for state-of-the-art. A class here would be a plant species. So even with 5k samples, the 250k image corpus would only have 50 species, using this rule of thumb. A good engineer could pick up DL and build a system that performs to this standard, because the tricks are all written down in the literature. If they do better, they either exploit unpublished methods, or researched those methods themselves, with their researchers.
  • Houshalter 9y ago
    But there is a lot of transfer learning between classes. For example, the network may learn to detect "broad leaves" or "red stems" for a common plant. It will then be able to use those feature on less common plants, and require much less training data for them. Transfer learning has been shown to work with just a single image of a desired class. It's also nice that life naturally fits into nested hierarchies because of evolution. So if you can recognize what family it belongs to, then that narrows down the possible sub families it can belong to. That in turn narrows down the possible clades it can belong to, etc, which narrows down the exact species. You couldn't find a more perfect use case for hierarchical softmax!