Expert Gate: Lifelong Learning with a Network

For all the different alternatives, we use the adaptive gradient in the training of the autoencoder. Table 1 shows the classification accuracy for the task labeling problem achieved by each of the different choices and our Expert Gate autoencoder. It can be noticed that the linear gate (Linear Autoencoder) fails to recognize the examples from the Flowers dataset – the linearly learned subspace might not be different from the subspace learned for Birds (due to the visual similarity). There is a slight relative improvement between the gate with the standardization (ReLU activation function and Sigmoid activation function) and without (No standardization). A small improvement can also be seen when using ReLU over the sigmoid activation function. Lastly, our design choice along with finetuning after the autoencoder learned on ImageNet achieves the highest accuracy in recognizing the three different tasks.