towards_few_shot_learning | Jupyter notebook with some steps towards few shot

towards_few_shot_learning is a Jupyter Notebook library. towards_few_shot_learning has no bugs, it has no vulnerabilities, it has a Strong Copyleft License and it has low support. You can download it from GitHub.

We use the EMNIST dataset of handwritten digits to test a simple approach for few shot learning. Choosing a fully connected net with inputs and layer outputs between 0 and 1 and no bias parameters we first trained the network with a subset of the digits. The pre-trained net is used for few shot learning with the untrained digits. Two basic idea were necessary: first the training of the first layer was disabled (or very slow) during few shot learning, and second using a shot consists of one untrained digit together with four previously trained digits and perform a training up to a predefined threshold. This way we reach a 90% accuracy for all handwritten digits after 10 shots. This jupyter notebook contains the generation of the tabels and images in the pdf: few_shot_paper.pdf.