What is Transfer Learning?

Transfer Learning in Computer Vision

Training “deep convolutional neural network” models could take days or months. Using previously trained models from computer vision benchmark datasets such as ImageNet is a shortcut for this process and is known as transfer learning. Transfer learning often involves using models trained on one problem as a starting point for solving related problems. Pre-trained models can be used directly for feature extraction in preprocessing to train new neural networks quickly, and reduce generalization errors. The weights in reused layers may be used in the training process and adapted to new problems, in a way transfer learning can be considered a weight initialization scheme. This may be useful when the first related problem has much more labeled data than the current problem, and the similarity of problem structures is useful in both contexts.

ImageNet Large Scale Image Recognition Challenge (ILSVRC) is an annual competition in which teams enter their models to compete for higher visual recognition accuracy on a given dataset. ILSVRC has inspired novel architectures and training methods for convolutional neural networks. The models produced for ILSVRC are trained on more than a million images to recognize 1000 categories of general features. The competition has also pushed them to the limits of performance in this specific task. These models are available for download or for direct use through APIs. All these advantages make them ideal candidates for the initial model in transfer learning.