One-shot learning through generalized representations with neural networks

Sammanfattning: Despite the rapid progress in the field of machine learning and artificial neural networks, many hurdles yet remain before machines can match human capabilities. One such hurdle is the copious amount of data required for these learning machines to reach adequate performance. There have been many methods to improve learning with limited data, going so far as to only use a single example, known as one-shot learning. A common strategy to take on one-shot learning involves turning the problem into a statistical comparison of two examples. In this paper, we propose a model that attempts to combine the information from a few given examples in order to learn more about the underlying distribution. We hypothesize that the performance of a model that is not reliant on a one-to-one comparison will scale better with increasing data, as it could combine information from different examples. Our proposed model involves training a convolutional neural network (CNN) to represent images in such a way that makes it easier for a smaller artificial neural network to infer whether a specified object is present in the image. To let the system learn about a new object category, the smaller network can simply be trained from scratch on that category. Testing this method reveals that training the CNN does not result in better performance compared to an untrained CNN with random initialization. Despite this, the smaller network learns surprisingly well even when dealing with limited data. As it stands, the proposed model offers no discernible bene- fits compared to previous work that uses statistical comparisons, but there may be room for further testing if the training procedure for the CNN is revised and improved on.