Paper Reviewed: DeCAF - A Deep Convolutional Activation Feature for Generic Visual Recognition

I just read a very interesting and exciting paper on Computer Vision based on Deep Learning. The title of the paper is "DeCAF: A Deep Convolutional Activation Feature for Generic Visual Recognition". It demonstrates methods to visualize features learned inside a Deep Neural Network more specifically the neuron activations of a specific layer inside the network. It also uses the parameters learned in the convolutional layers to demonstrate that the features learned are generic enough to be used for other classification tasks. The author holds the parameters learned in the convolutional layers and only trained the network in the final layer for classification. This demonstrates that the parameters are be "reused" across different classification problems based on images and achieve good generalization.

The big deal here is both theoretical and practical. It is theoretical because the learning is "transferrable" meaning once the machine has done learning a visual task (e.g. differentiate a face or non-face images), it is relatively easy to train it to learn another visual task. It is practical because you can now train your machine on a big dataset that is similar to your task and then use the learned parameters to fine-tune your machine for the task of interest. Note that for deep learning, the most difficult layers to train is the first few layers of the network. Using human vision as an analogy, the first few layers are modelling the eye (i.e. iris, cornea, retina) the last few layers are the visual cortex of the brain. It is very important to get the right information to the brain before making any classification! However, once you have a developed eye (i.e. a set of trained parameters), the brain just needs to discriminate the inputs to make a decision. It is common that the training dataset for the task of interest is small which is easy to be overfitted but now we can try to train the machine with a similar and big dataset and then fine-tune with the actual dataset of the task, this will make it very useful!