Transfer Learning — part 2

Oftentimes, it is not the case and one can have just a few examples to learn from..This is few-shot learning problem..Imagine having just one example (one-shot learning) or even no labeled chihuahua at all (zero-shot learning)..It is not yet zero-shot learning, but this scheme can work for few-shot learning..After observing a few examples of the new class, you can hope to learn to recognize the new class with kNN..We can construct a vector-space embedding for both labels and training examples so that a training example and its label are mapped as close to each other as possible in such a common space.This approach is actively used in image classification: a common space embedding is learned for images and for words and words serve as labels..(2013) used pretrained embeddings trained on Wikipedia texts and they learned neural network based mapping of images to word-embedding vector space.Taken from Socher et al. (2013)Norouzi et al..Having trained a multi-class image classifier, they used predicted probabilities of classes to perform probability-weighted average of the word-embedding vectors corresponding to labels of the classes.Romera-Paredes and Torr (2015) developed a linear transformation based approach to zero-shot learning that however requires a characterization of labels and training examples in terms of attributes..(2018).Few-shot learningFew-shot learning is related to the field of Meta-Learning (learning how to learn) where a model is required to quickly learn a new task from a small amount of new data.Lake et al..A new training example is memorized, and then, when a similar new testing example arrives, kNN searches its memory for similar examples and finds the memorized training example and its label..Alternatively, one can try to augment Deep Neural Networks with memory in a more direct way that allows for end-to-end training.Koch, Zemel, and Salakhutdinov (2015) developed few-shot learning method based on nearest neighbour classification with similarity metric learned by a Siamese neural network..(2016) developed a few-shot learning method using Memory-Augmented Neural Network (MANN)..The idea of their model was similar to Neural Turing Machines (Graves, Wayne, and Danihelka 2014): a neural network extended with an external memory module so that the model is differentiable and can be trained end-to-end.. More details

Leave a Reply