Computer Vision News - January 2017

The move from hand-designed to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. A novel work “ Learning to learn by gradient descent by gradient descent ” by the Google DeepMind team takes the current paradigm of machine learning for image recognition one step forward. Just as the field has transitioned from hand-crafted features for object recognition to machine-learned features, enabling us to deal with new domains without having a specialized knowledge of them, they propose that we move from hand-designed update rules to a learned update rule. The DeepMind team proposes another neuron network, which they call the optimizer , with its own set of parameters. It is given the task of producing the parameter updates for the recognition-feature producing network, which they call the optimizee . The optimizer receives information from the optimizee, and prompts parameter updates based on this information, then following the update it receives information on performance, prompting further parameter updates based on that. The DeepMind team wrote the final optimizee parameters as a function of the optimizer parameters, in the following form: Where : the optimizer parameters; f: the function in question; g t is modeled by the recurrent neural network (RNN), as we shall see next. While the objective function depends only on the last parameter value for training the optimizer, it is convenient to have an objective that depends on the entire trajectory of optimization. To this end, the DeepMind team used a two-layer LTSM network , looking at a single coordinate to define the optimizer and sharing it across different parameters of the optimizee. The inputs of this mini-network are the optimizee gradient and the previous hidden state. The output is the update for the corresponding optimizee parameter. Learning to learn by gradient descent by gradient descent Tool +1 = + (△ ( , Computer Vision News Tool 9