Designing Neural Network Architectures using Reinforcement Learning

Authors: Bowen Baker, Otkrist Gupta, Nikhil Naik, Ramesh Raskar

ICLR 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct experiments with a space of model architectures consisting of only standard convolution, pooling, and fully connected layers using three standard image classification datasets: CIFAR-10, SVHN, and MNIST. The learning agent discovers CNN architectures that beat all existing networks designed only with the same layer types (e.g., Springenberg et al. (2014); Srivastava et al. (2015)). In addition, their performance is competitive against network designs that include complex layer types and training procedures (e.g., Clevert et al. (2015); Lee et al. (2016)). Finally, the Meta QNN selected models comfortably outperform previous automated network design methods (Stanley & Miikkulainen, 2002; Bergstra et al., 2013).
Researcher Affiliation Academia Bowen Baker, Otkrist Gupta, Nikhil Naik & Ramesh Raskar Media Laboratory Massachusetts Institute of Technology Cambridge MA 02139, USA {bowen, otkrist, naik, raskar}@mit.edu
Pseudocode Yes A ALGORITHM; Algorithm 1 Q-learning For CNN Topologies; Algorithm 2 SAMPLE NEW NETWORK(ϵ, Q); Algorithm 3 UPDATE Q VALUES(Q, S, U, accuracy)
Open Source Code Yes For more information, model files, and code, please visit https://bowenbaker.github.io/metaqnn/
Open Datasets Yes We conduct experiments with a space of model architectures consisting of only standard convolution, pooling, and fully connected layers using three standard image classification datasets: CIFAR-10, SVHN, and MNIST.
Dataset Splits Yes For each experiment, we created a validation set by randomly taking 5,000 samples from the training set such that the resulting class distributions were unchanged.
Hardware Specification No Our experiments using Caffe (Jia et al., 2014) took 8-10 days to complete for each dataset with a hardware setup consisting of 10 NVIDIA GPUs.
Software Dependencies No Our experiments using Caffe (Jia et al., 2014) took 8-10 days to complete for each dataset with a hardware setup consisting of 10 NVIDIA GPUs. (No version for Caffe or other libraries is specified.)
Experiment Setup Yes For every network, a dropout layer was added after every two layers. The ith dropout layer, out of a total n dropout layers, had a dropout probability of i 2n. Each model was trained for a total of 20 epochs with the Adam optimizer (Kingma & Ba, 2014) with β1 = 0.9, β2 = 0.999, ε = 10 8. The batch size was set to 128, and the initial learning rate was set to 0.001.