M.I.N.D.S – Minimal Instance Neural Data System

In this Study, we compare the unseen data recognition power provided by a set of few shot learning techniques, applied over a large set of pre-trained models including ResNet, InceptionV3, ShuffleNet, etc. Few-shot learning or low sample learning is a type of learning which tries to classify data having very limited labeled examples, and it has been of interest due to its promise in situations with rare or expensive data. Agriculture or CSI sensing, where every signal fluctuation might result in a totally new unpredictable signal variation. And yet, few shot learning performance can vary considerably by model architecture. Since these pre-trained models have proven to be effective in extracting features of novel data, we can further study their effectiveness in classifying new data across different datasets by fine-tuning them with multiple quantitatively and qualitatively different few shot learning methods. We compare the merits and demerits of different models used and interpret the accuracy, precision and generalization capability of each model. The results gained in this study clearly showed that ShuffleNet is the best model for the use of few-shot learning applications with an accuracy of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 1. 0 5 \%}$</tex> in a <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1 0}$</tex>-shots setting.