https://en.wikipedia.org/wiki/Neural_architecture_search
Neural architecture search (NAS)[1] denotes the process of automatically designing artificial neural networks (ANN), a widely used model in the field of machine learning. Various approaches to NAS have designed networks that are on par or even outperform hand-designed architectures.[2][3]Methods for NAS can be categorized according to the search space, search strategy and performance estimation strategy used:[1]
- The search space defines which type of ANN can be designed and optimized in principle.
- The search strategy defines which strategy is used to find optimal ANN's within the search space.
- Obtaining the performance of an ANN is costly as this requires training the ANN first. Therefore, performance estimation strategies are used obtain less costly estimates of a model's performance.
Detailed explanation and examples can be found in recent survey article.[1]
NAS is closely related to Hyperparameter optimization and is a subfield of Automated machine learning (AutoML).
Contents
- 1NAS with Reinforcement Learning
- 2NAS with Evolution
- 3NAS with Hill-climbing
- 4Multi-objective Neural architecture search
- 5References
NAS with Reinforcement Learning[edit]
Reinforcement learning (RL) can be utilized as a search strategy for NAS. Zoph et al.[2] applied NAS with RL the CIFAR-10 dataset yielded a network architecture that rivals the best manually-designed architecture for accuracy, with a test error rate of 3.65, which is 0.09 percent better and 1.05x faster than the previous model that used a similar design. On the Penn Treebank dataset, that model composed a recurrent cell that outperforms LSTM, reaching a test set perplexity of 62.4, or 3.6 perplexity better than the prior leading system. On the PTB character language modeling task it achieved bits per character of 1.214.[2]
Learning a model architecture directly on a large dataset is a lengthy process. NASNet[3][4] addressed this issue by transferring a building block designed for a small dataset to a larger dataset. The design was constrained to use two types of convolutional cells to return feature maps that serve two main functions when convoluting an input feature map: "Normal Cells" that return maps of the same extent (height and width) and "Reduction Cells" in which the returned feature map height and width is reduced by a factor of two. For the Reduction Cell, the initial operation applied to the cell’s inputs uses a stride of two (to reduce the height and width).[3] The learned aspect of the design included elements such as which lower layer(s) each higher layer took as input, the transformations applied at that layer and to merge multiple outputs at each layer. In the studied example, the best convolutional layer (or "cell") was designed for the CIFAR-10 dataset and then applied to the ImageNet dataset by stacking copies of this cell, each with its own parameters. The approach yielded accuracy of 82.7% top-1 and 96.2% top-5. This exceeded the best human-invented architectures at a cost of 9 billion fewer FLOPS—a reduction of 28%. The system continued to exceed the manually-designed alternative at varying computation levels. The image features learned from image classification can be transferred to other computer vision problems. E.g., for object detection, the learned cells integrated with the Faster-RCNN framework improved performance by 4.0% on the COCO dataset.[3]
In the so-called Efficient Neural Architecture Search (ENAS), a controller discovers neural network architectures by learning to search for an optimal subgraph within a large graph. The controller is trained with policy gradient to select a subgraph that maximizes the validation set's expected reward. The model corresponding to the subgraph is trained to minimize a canonical cross entropy loss. Multiple child models share parameters, ENAS requires fewer GPU-hours than other approaches and 1000-fold less than "standard" NAS. On CIFAR-10, the ENAS design achieved a test error of 2.89%, comparable to NASNet.On Penn Treebank, the ENAS design reached test perplexity of 55.8.[5]
NAS with Evolution[edit]
Several groups employed evolutionary algorithms for NAS.[6][7][8] Mutations in the context of evolving ANNs are operations such as adding a layer, removing a layer or changing the type of a layer (e.g., from convolution to pooling). On CIFAR-10, it was shown that evolution and RL perform comparable, while both outperform Random search.[7]
NAS with Hill-climbing[edit]
Another group used a hill climbing procedure that applies network morphisms, followed by short cosine-annealing optimization runs. Surprisingly, The approach yielded competitive results, requiring resources on the same order of magnitude as training a single network. E.g., on CIFAR-10, the method designed and trained a network with an error rate below 5% in 12 hours on a single GPU.[9]
Multi-objective Neural architecture search[edit]
While most approaches solely focus on finding architecture with maximal predictive performance, for most practical applications there are several other relevant objectives, such as memory consumption, model size or inference time (i.e., the time required to obtain a prediction). Because of that, researchers worked Multi-objective neural architecture search.[10][11]
LEMONADE[10] denotes an evolutionary algorithm with Lamarckism to efficiently optimize multiple objectives. In every generation, children networks are generated to improve the Pareto frontier with respect to the current population of ANNs.
Neural Architect[11] is claimed to be a resource-aware multi-objective reinforcement learning based NAS with network embedding and performance prediction. Network embedding encodes an existing network to a trainable embedding vector. Based on the embedding, a controller network generates transformations of the target network. A multi-objective reward function considers network accuracy, computational resource and training time. The reward is predicted by multiple performance simulation networks that are pre-trained or co-trained with the controller network. The controller network is trained via policy gradient. Following a modification, the resulting candidate network is submitted to both an accuracy network and a training time network. The results of each are combined by a reward engine that passes its output back to the controller network.
