The Data Question

If someone were to ask me ‘the data question’ today, I would answer that more relevant training data always seems to help reduce deep neural network error; but I can’t tell you how much data is required to achieve a specific goal metric.From those heuristics I’ve mentioned earlier, we can guess that for a simple binary classification problem we may get adequate performance from as few as 1000 images per class using transfer learning across a similar source and target domain..Even if the performance we get from that data is not adequate, we can say with some certainty that more data will improve a deep neural network..It may be unlikely we will ever definitively answer ‘the data question.’ We believe there is more work to do here and we certainly hope that our work inspires other researchers to continue to clarify this relationship.To hear more about Transfer Learning, check out our presentation from AnacondaCon called “Deep Learning with just a little data.”. More details

Leave a Reply