If Human Can Learn from Few Samples, Why Can’t AI? An Attempt On Similar Object Recognition With Few Training Data Using Meta-Learning
Subhayu Dutta, Saptiva Goswami, Sonali Debnath, Subhrangshu Adhikary, Anandaprova Majumder · 2023
An age-old problem that persists with deep learning is that it requires thousands of training samples to learn. Transfer learning is an advancement to deep learning through which learned models can be adapted to recognize new data. But, learning new data with transfer learning depends on the properties of the previous data the model was trained on. In cases where the properties of the new data are different and all the classes of the new data have very few contrasting properties, transfer learning does not produce reliable results. Therefore, we require a method to learn properties from the dataset having low variation among classes using very few samples. To achieve this, we have proposed the usage of a Few-Shot Meta-Learning algorithm and tested this on two open-source datasets having similar-looking classes. The results from the experiment show that the proposed method can reach up to 100% classification accuracy with only 30 samples. Using the same number of samples, transfer learning obtained up to 41.9% accuracy, and general convolution neural network obtained up to 32.0% accuracy. The model can be implemented to reduce the requirement of a large number of annotated training samples to train deep learning models.