Searching for Activation Functions
Abstract
The choice of activation functions in deep networks has a significant effect on the training dynamics and task performance. Currently, the most successful and widelyused activation function is the Rectified Linear Unit (ReLU). Although various handdesigned alternatives to ReLU have been proposed, none have managed to replace it due to inconsistent gains. In this work, we propose to leverage automatic search techniques to discover new activation functions. Using a combination of exhaustive and reinforcement learningbased search, we discover multiple novel activation functions. We verify the effectiveness of the searches by conducting an empirical evaluation with the best discovered activation function. Our experiments show that the best discovered activation function, , which we name Swish, tends to work better than ReLU on deeper models across a number of challenging datasets. For example, simply replacing ReLUs with Swish units improves top1 classification accuracy on ImageNet by 0.9% for Mobile NASNetA and 0.6% for InceptionResNetv2. The simplicity of Swish and its similarity to ReLU make it easy for practitioners to replace ReLUs with Swish units in any neural network.
Searching for Activation Functions
Prajit Ramachandran^{†}^{†}thanks: Work done as a member of the Google Brain Residency program (g.co/brainresidency)., Barret Zoph, Quoc V. Le 

Google Brain 
{prajit,barretzoph,qvl}@google.com 
1 Introduction
At the heart of every deep network lies a linear transformation followed by an activation function . The activation function plays a major role in the success of training deep neural networks. Currently, the most successful and widelyused activation function is the Rectified Linear Unit (ReLU) (Hahnloser et al., 2000; Jarrett et al., 2009; Nair & Hinton, 2010), defined as . The use of ReLUs was a breakthrough that enabled the fully supervised training of stateoftheart deep networks (Krizhevsky et al., 2012). Deep networks with ReLUs are more easily optimized than networks with sigmoid or tanh units, because gradients are able to flow when the input to the ReLU function is positive. Thanks to its simplicity and effectiveness, ReLU has become the default activation function used across the deep learning community.
While numerous activation functions have been proposed to replace ReLU (Maas et al., 2013; He et al., 2015; Clevert et al., 2015; Klambauer et al., 2017), none have managed to gain the widespread adoption that ReLU enjoys. Many practitioners have favored the simplicity and reliability of ReLU because the performance improvements of the other activation functions tend to be inconsistent across different models and datasets.
The activation functions proposed to replace ReLU were handdesigned to fit properties deemed to be important. However, the use of search techniques to automate the discovery of traditionally humandesigned components has recently shown to be extremely effective (Zoph & Le, 2016; Bello et al., 2017; Zoph et al., 2017). For example, Zoph et al. (2017) used reinforcement learningbased search to find a replicable convolutional cell that outperforms humandesigned architectures on ImageNet.
In this work, we use automated search techniques to discover novel activation functions. We focus on finding new scalar activation functions, which take in as input a scalar and output a scalar, because scalar activation functions can be used to replace the ReLU function without changing the network architecture. Using a combination of exhaustive and reinforcement learningbased search, we find a number of novel activation functions that show promising performance. To further validate the effectiveness of using searches to discover scalar activation functions, we empirically evaluate the best discovered activation function. The best discovered activation function, which we call Swish, is , where is a constant or trainable parameter. Our extensive experiments show that Swish consistently matches or outperforms ReLU on deep networks applied to a variety of challenging domains such as image classification and machine translation. On ImageNet, replacing ReLUs with Swish units improves top1 classification accuracy by 0.9% on Mobile NASNetA (Zoph et al., 2017) and 0.6% on InceptionResNetv2 (Szegedy et al., 2017). These accuracy gains are significant given that one year of architectural tuning and enlarging yielded 1.3% accuracy improvement going from Inception V3 (Szegedy et al., 2016) to InceptionResNetv2 (Szegedy et al., 2017).
2 Methods
In order to utilize search techniques, a search space that contains promising candidate activation functions must be designed. An important challenge in designing search spaces is balancing the size and expressivity of the search space. An overly constrained search space will not contain novel activation functions, whereas a search space that is too large will be difficult to effectively search. To balance the two criteria, we design a simple search space inspired by the optimizer search space of Bello et al. (2017) that composes unary and binary functions to construct the activation function.
As shown in Figure 1, the activation function is constructed by repeatedly composing the the “core unit”, which is defined as . The core unit takes in two scalar inputs, passes each input independently through an unary function, and combines the two unary outputs with a binary function that outputs a scalar. Since our aim is to find scalar activation functions which transform a single scalar input into a single scalar output, the inputs of the unary functions are restricted to the layer preactivation and the binary function outputs.
Given the search space, the goal of the search algorithm is to find effective choices for the unary and binary functions. The choice of the search algorithm depends on the size of the search space. If the search space is small, such as when using a single core unit, it is possible to exhaustively enumerate the entire search space. If the core unit is repeated multiple times, the search space will be extremely large (i.e., on the order of possibilities), making exhaustive search infeasible.
For large search spaces, we use an RNN controller (Zoph & Le, 2016), which is visualized in Figure 2. At each timestep, the controller predicts a single component of the activation function. The prediction is fed back to the controller in the next timestep, and this process is repeated until every component of the activation function is predicted. The predicted string is then used to construct the activation function.
Once a candidate activation function has been generated by the search algorithm, a “child network” with the candidate activation function is trained on some task, such as image classification on CIFAR10. After training, the validation accuracy of the child network is recorded and used to update the search algorithm. In the case of exhaustive search, a list of the top performing activation functions ordered by validation accuracy is maintained. In the case of the RNN controller, the controller is trained with reinforcement learning to maximize the validation accuracy, where the validation accuracy serves as the reward. This training pushes the controller to generate activation functions that have high validation accuracies.
Since evaluating a single activation function requires training a child network, the search is computationally expensive. To decrease the wall clock time required to conduct the search, a distributed training scheme is used to parallelize the training of each child network. In this scheme, the search algorithm proposes a batch of candidate activation functions which are added to a queue. Worker machines pull activation functions off the queue, train a child network, and report back the final validation accuracy of the corresponding activation function. The validation accuracies are aggregated and used to update the search algorithm.
3 Search Findings
We conduct all our searches with the ResNet20 (He et al., 2016a) as the child network architecture, and train on CIFAR10 (Krizhevsky & Hinton, 2009) for 10K steps. This constrained environment could potentially skew the results because the top performing activation functions might only perform well for small networks. However, we show in the experiments section that many of the discovered functions generalize to larger models. Exhaustive search is used for small search spaces, while an RNN controller is used for larger search spaces. The RNN controller is trained with Policy Proximal Optimization (Schulman et al., 2017), using the exponential moving average of rewards as a baseline to reduce variance. The full list unary and binary functions considered are as follows:

Unary functions: , , , , , , , , , , , , , , , , , , , , , , ,

Binary functions: , , , , , , , , ,
where indicates a perchannel trainable parameter and is the sigmoid function. Different search spaces are created by varying the number of core units used to construct the activation function and varying the unary and binary functions available to the search algorithm.
Figure 3 plots the top performing novel activation functions found by the searches. We highlight several noteworthy trends uncovered by the searches:

Complicated activation functions consistently underperform simpler activation functions, potentially due to an increased difficulty in optimization. The best performing activation functions can be represented by or core units.

A common structure shared by the top activation functions is the use of the raw preactivation as input to the final binary function: . The ReLU function also follows this structure, where and .

The searches discovered activation functions that utilize periodic functions, such as and . The most common use of periodic functions is through addition or subtraction with the raw preactivation (or a linearly scaled ). The use of periodic functions in activation functions has only been briefly explored in prior work (Parascandolo et al., 2016), so these discovered functions suggest a fruitful route for further research.
Since the activation functions were found using a relatively small child network, their performance may not generalize when applied to bigger models. To test the robustness of the top performing novel activation functions to different architectures, we run additional experiments using the preactivation ResNet164 (RN) (He et al., 2016b), Wide ResNet 2810 (WRN) (Zagoruyko & Komodakis, 2016), and DenseNet 10012 (DN) (Huang et al., 2017) models. We implement the 3 models in TensorFlow and replace the ReLU function with each of the top novel activation functions discovered by the searches. We use the same hyperparameters described in each work, such as optimizing using SGD with momentum, and follow previous works by reporting the median of 5 different runs.
Function  RN  WRN  DN 

ReLU []  93.8  95.3  94.8 
94.5  95.5  94.9  
94.3  95.3  94.8  
94.1  94.8  94.6  
94.0  95.1  94.4  
93.9  94.7  94.9  
93.9  94.2  94.5  
91.5  92.1  92.0  
85.1  92.1  91.1 
Function  RN  WRN  DN 

ReLU []  74.2  77.8  83.7 
75.1  78.0  83.9  
74.8  78.6  84.2  
75.2  76.6  81.8  
73.4  77.1  74.3  
75.2  76.7  83.1  
74.8  76.0  78.6  
66.1  68.3  67.9  
52.8  70.6  68.1 
The results are shown in Tables 2 and 2. Despite the changes in model architecture, six of the eight activation functions successfully generalize. Of these six activation functions, all match or outperform ReLU on ResNet164. Furthermore, two of the discovered activation functions, and , consistently match or outperform ReLU on all three models.
While these results are promising, it is still unclear whether the discovered activation functions can successfully replace ReLU on challenging real world datasets. In order to validate the effectiveness of the searches, in the rest of this work we focus on empirically evaluating the activation function , which we call Swish. We choose to extensively evaluate Swish instead of because early experimentation showed better generalization for Swish. In the following sections, we analyze the properties of Swish and then conduct a thorough empirical evaluation comparing Swish, ReLU, and other candidate baseline activation functions on number of large models across a variety of tasks.
4 Swish
To recap, Swish is defined as , where is the sigmoid function and is either a constant or a trainable parameter. Figure 5 plots the graph of Swish for different values of . If , Swish is equivalent to the Sigmoidweighted Linear Unit (SiL) of Elfwing et al. (2017) that was proposed for reinforcement learning. If , Swish becomes the scaled linear function . As , the sigmoid component approaches a – function, so Swish becomes like the ReLU function. This suggests that Swish can be loosely viewed as a smooth function which nonlinearly interpolates between the linear function and the ReLU function. The degree of interpolation can be controlled by the model if is set as a trainable parameter.
Like ReLU, Swish is unbounded above and bounded below. Unlike ReLU, Swish is smooth and nonmonotonic. In fact, the nonmonotonicity property of Swish distinguishes itself from most common activation functions. The derivative of Swish is
The first derivative of Swish is shown in Figure 5 for different values of . The scale of controls how fast the first derivative asymptotes to and . When , the derivative has magnitude less than for inputs that are less than around . Thus, the success of Swish with implies that the gradient preserving property of ReLU (i.e., having a derivative of 1 when ) may no longer be a distinct advantage in modern architectures.
The most striking difference between Swish and ReLU is the nonmonotonic “bump” of Swish when . As shown in Figure 7, a large percentage of preactivations fall inside the domain of the bump (, which indicates that the nonmonotonic bump is an important aspect of Swish. The shape of the bump can be controlled by changing the parameter. While fixing is effective in practice, the experiments section shows that training can further improve performance on some models. Figure 7 plots distribution of trained values from a Mobile NASNetA model (Zoph et al., 2017). The trained values are spread out between and and have a peak at , suggesting that the model takes advantage of the additional flexibility of trainable parameters.
Practically, Swish can be implemented with a single line code change in most deep learning libraries, such as TensorFlow (Abadi et al., 2016) (e.g., x * tf.sigmoid(beta * x) or tf.nn.swish(x) if using a version of TensorFlow released after the submission of this work). As a cautionary note, if BatchNorm (Ioffe & Szegedy, 2015) is used, the scale parameter should be set. Some high level libraries turn off the scale parameter by default due to the ReLU function being piecewise linear, but this setting is incorrect for Swish. For training Swish networks, we found that slightly lowering the learning rate used to train ReLU networks works well.
5 Experiments with Swish
We benchmark Swish against ReLU and a number of recently proposed activation functions on challenging datasets, and find that Swish matches or exceeds the baselines on nearly all tasks. The following sections will describe our experimental settings and results in greater detail. As a summary, Table 3 shows Swish in comparison to each baseline activation function we considered (which are defined in the next section). The results in Table 3 are aggregated by comparing the performance of Swish to the performance of different activation functions applied to a variety of models, such as Inception ResNetv2 (Szegedy et al., 2017) and Transformer (Vaswani et al., 2017), across multiple datasets, such as CIFAR, ImageNet, and EnglishGerman translation.^{1}^{1}1To avoid skewing the comparison, each model type is compared just once. A model with multiple results is represented by the median of its results. Specifically, the models with aggregated results are (a) ResNet164, Wide ResNet 2810, and DenseNet 10012 across the CIFAR10 and CIFAR100 results, (b) Mobile NASNetA and InceptionResNetv2 across the 3 runs, and (c) WMT Transformer model across the 4 newstest results. The improvement of Swish over other activation functions is statistically significant under a onesided paired sign test.
Baselines  ReLU  LReLU  PReLU  Softplus  ELU  SELU  GELU 

Swish Baseline  9  7  6  6  8  8  8 
Swish Baseline  0  1  3  2  0  1  1 
Swish Baseline  0  1  0  1  1  0  0 
5.1 Experimental Set Up
We compare Swish against several additional baseline activation functions on a variety of models and datasets. Since many activation functions have been proposed, we choose the most common activation functions to compare against, and follow the guidelines laid out in each work:

Gaussian Error Linear Unit (GELU) (Hendrycks & Gimpel, 2016): , where is the cumulative distribution function of the standard normal distribution. GELU is a nonmonotonic function that has a shape similar to Swish with .
We evaluate both Swish with a trainable and Swish with a fixed (which for simplicity we call Swish1, but it is equivalent to the Sigmoidweighted Linear Unit of Elfwing et al. (2017)). Note that our results may not be directly comparable to the results in the corresponding works due to differences in our training setup.
5.2 Cifar
We first compare Swish to all the baseline activation functions on the CIFAR10 and CIFAR100 datasets (Krizhevsky & Hinton, 2009). We follow the same set up used when comparing the activation functions discovered by the search techniques, and compare the median of 5 runs with the preactivation ResNet164 (He et al., 2016b), Wide ResNet 2810 (WRN) (Zagoruyko & Komodakis, 2016), and DenseNet 10012 (Huang et al., 2017) models.
Model  ResNet  WRN  DenseNet 

LReLU  94.2  95.6  94.7 
PReLU  94.1  95.1  94.5 
Softplus  94.6  94.9  94.7 
ELU  94.1  94.1  94.4 
SELU  93.0  93.2  93.9 
GELU  94.3  95.5  94.8 
ReLU  93.8  95.3  94.8 
Swish1  94.7  95.5  94.8 
Swish  94.5  95.5  94.8 
Model  ResNet  WRN  DenseNet 

LReLU  74.2  78.0  83.3 
PReLU  74.5  77.3  81.5 
Softplus  76.0  78.4  83.7 
ELU  75.0  76.0  80.6 
SELU  73.2  74.3  80.8 
GELU  74.7  78.0  83.8 
ReLU  74.2  77.8  83.7 
Swish1  75.1  78.5  83.8 
Swish  75.1  78.0  83.9 
The results in Tables 5 and 5 show how Swish and Swish1 consistently matches or outperforms ReLU on every model for both CIFAR10 and CIFAR100. Swish also matches or exceeds the best baseline performance on almost every model. Importantly, the “best baseline” changes between different models, which demonstrates the stability of Swish to match these varying baselines. Softplus, which is smooth and approaches zero on one side, similar to Swish, also has strong performance.
5.3 ImageNet
Next, we benchmark Swish against the baseline activation functions on the ImageNet 2012 classification dataset (Russakovsky et al., 2015). ImageNet is widely considered one of most important image classification datasets, consisting of a 1,000 classes and 1.28 million training images. We evaluate on the validation dataset, which has 50,000 images.
We compare all the activation functions on a variety of architectures designed for ImageNet: InceptionResNetv2, Inceptionv4, Inceptionv3 (Szegedy et al., 2017), MobileNet (Howard et al., 2017), and Mobile NASNetA (Zoph et al., 2017). All these architectures were designed with ReLUs. We again replace the ReLU activation function with different activation functions and train for a fixed number of steps, determined by the convergence of the ReLU baseline. For each activation function, we try 3 different learning rates with RMSProp (Tieleman & Hinton, 2012) and pick the best.^{2}^{2}2For some of the models with ELU, SELU, and PReLU, we train with an additional 3 learning rates (so a total of 6 learning rates) because the original 3 learning rates did not converge. All networks are initialized with He initialization (He et al., 2015).^{3}^{3}3For SELU, we tried both He initialization and the initialization recommended in Klambauer et al. (2017), and choose the best result for each model separately. To verify that the performance differences are reproducible, we run the InceptionResNetv2 and Mobile NASNetA experiments 3 times with the best learning rate from the first experiment. We plot the learning curves for Mobile NASNetA in Figure 8.
Model  Top1 Acc. (%)  Top5 Acc. (%)  

LReLU  79.5  79.5  79.6  94.7  94.7  94.7 
PReLU  79.7  79.8  80.1  94.8  94.9  94.9 
Softplus  80.1  80.2  80.4  95.2  95.2  95.3 
ELU  75.8  79.9  80.0  92.6  95.0  95.1 
SELU  79.0  79.2  79.2  94.5  94.4  94.5 
GELU  79.6  79.6  79.9  94.8  94.8  94.9 
ReLU  79.5  79.6  79.8  94.8  94.8  94.8 
Swish1  80.2  80.3  80.4  95.1  95.2  95.2 
Swish  80.2  80.2  80.3  95.0  95.2  95.0 
Model  Top1 Acc. (%)  Top5 Acc. (%) 

LReLU  72.5  91.0 
PReLU  74.2  91.9 
Softplus  73.6  91.6 
ELU  73.9  91.3 
SELU  73.2  91.0 
GELU  73.5  91.4 
ReLU  72.0  90.8 
Swish1  74.2  91.6 
Swish  74.2  91.7 
Model  Top1 Acc. (%)  Top5 Acc. (%) 

LReLU  78.4  94.1 
PReLU  77.7  93.5 
Softplus  78.7  94.4 
ELU  77.9  93.7 
SELU  76.7  92.8 
GELU  77.7  93.9 
ReLU  78.4  94.2 
Swish1  78.7  94.2 
Swish  78.7  94.0 
Model  Top1 Acc. (%)  Top5 Acc. (%) 

LReLU  79.3  94.7 
PReLU  79.3  94.4 
Softplus  79.6  94.8 
ELU  79.5  94.5 
SELU  78.3  94.5 
GELU  79.0  94.6 
ReLU  79.2  94.6 
Swish1  79.3  94.7 
Swish  79.3  94.6 
The results in Tables 6–10 show strong performance for Swish. On InceptionResNetv2, Swish outperforms ReLU by a nontrivial . Swish performs especially well on mobile sized models, with a boost on Mobile NASNetA and a boost on MobileNet over ReLU. Swish also matches or exceeds the best performing baseline on most models, where again, the best performing baseline differs depending on the model. Softplus achieves accuracies comparable to Swish on the larger models, but performs worse on both mobile sized models. For Inceptionv4, the gains from switching between activation functions is more limited, and Swish slightly underperforms Softplus and ELU. In general, the results suggest that switching to Swish improves performance with little additional tuning.
5.4 Machine Translation
We additionally benchmark Swish on the domain of machine translation. We train machine translation models on the standard WMT 2014 EnglishGerman dataset, which has 4.5 million training sentences, and evaluate on 4 different newstest sets using the standard BLEU metric. We use the attention based Transformer (Vaswani et al., 2017) model, which utilizes ReLUs in a 2layered feedforward network between each attention layer. We train a 12 layer “Base Transformer” model with 2 different learning rates^{4}^{4}4We tried an additional learning rate for Softplus, but found it did not work well across all learning rates. for 300K steps, but otherwise use the same hyperparameters as in the original work, such as using Adam (Kingma & Ba, 2015) to optimize.
Model  newstest2013  newstest2014  newstest2015  newstest2016 

LReLU  26.2  27.9  29.8  33.4 
PReLU  26.3  27.7  29.7  33.1 
Softplus  23.4  23.6  25.8  29.2 
ELU  24.6  25.1  27.7  32.5 
SELU  23.7  23.5  25.9  30.5 
GELU  25.9  27.3  29.5  33.1 
ReLU  26.1  27.8  29.8  33.3 
Swish1  26.2  28.0  30.1  34.0 
Swish  26.5  27.6  30.0  33.1 
Table 11 shows that Swish outperforms or matches the other baselines on machine translation. Swish1 does especially well on newstest2016, exceeding the next best performing baseline by BLEU points. The worst performing baseline function is Softplus, demonstrating inconsistency in performance across differing domains. In contrast, Swish consistently performs well across multiple domains.
6 Related Work
Swish was found using a variety of automated search techniques. Search techniques have been utilized in other works to discover convolutional and recurrent architectures (Zoph & Le, 2016; Zoph et al., 2017; Real et al., 2017; Cai et al., 2017; Zhong et al., 2017) and optimizers (Bello et al., 2017). The use of search techniques to discover traditionally handdesigned components is an instance of the recently revived subfield of metalearning (Schmidhuber, 1987; Naik & Mammone, 1992; Thrun & Pratt, 2012). Metalearning has been used to find initializations for oneshot learning (Finn et al., 2017; Ravi & Larochelle, 2016), adaptable reinforcement learning (Wang et al., 2016; Duan et al., 2016), and generating model parameters (Ha et al., 2016). Metalearning is powerful because the flexibility derived from the minimal assumptions encoded leads to empirically effective solutions. We take advantage of this property in order to find scalar activation functions, such as Swish, that have strong empirical performance.
While this work focuses on scalar activation functions, which transform one scalar to another scalar, there are many types of activation functions used in deep networks. Manytoone functions, like max pooling, maxout (Goodfellow et al., 2013), and gating (Hochreiter & Schmidhuber, 1997; Srivastava et al., 2015; van den Oord et al., 2016; Dauphin et al., 2016; Wu et al., 2016; Miech et al., 2017), derive their power from combining multiple sources in a nonlinear way. Onetomany functions, like Concatenated ReLU (Shang et al., 2016), improve performance by applying multiple nonlinear functions to a single input. Finally, manytomany functions, such as BatchNorm (Ioffe & Szegedy, 2015) and LayerNorm (Ba et al., 2016), induce powerful nonlinear relationships between their inputs.
Most prior work has focused on proposing new activation functions (Maas et al., 2013; Agostinelli et al., 2014; He et al., 2015; Clevert et al., 2015; Hendrycks & Gimpel, 2016; Klambauer et al., 2017; Qiu & Cai, 2017; Zhou et al., 2017; Elfwing et al., 2017), but few studies, such as Xu et al. (2015), have systematically compared different activation functions. To the best of our knowledge, this is the first study to compare scalar activation functions across multiple challenging datasets.
Our study shows that Swish consistently outperforms ReLU on deep models. The strong performance of Swish challenges conventional wisdom about ReLU. Hypotheses about the importance of the gradient preserving property of ReLU seem unnecessary when residual connections (He et al., 2016a) enable the optimization of very deep networks. A similar insight can be found in the fully attentional Transformer (Vaswani et al., 2017), where the intricately constructed LSTM cell (Hochreiter & Schmidhuber, 1997) is no longer necessary when constantlength attentional connections are used. Architectural improvements lessen the need for individual components to preserve gradients.
7 Conclusion
In this work, we utilized automatic search techniques to discover novel activation functions that have strong empirical performance. We then empirically validated the best discovered activation function, which we call Swish and is defined as . Our experiments used models and hyperparameters that were designed for ReLU and just replaced the ReLU activation function with Swish; even this simple, suboptimal procedure resulted in Swish consistently outperforming ReLU and other activation functions. We expect additional gains to be made when these models and hyperparameters are specifically designed with Swish in mind. The simplicity of Swish and its similarity to ReLU means that replacing ReLUs in any network is just a simple one line code change.
Acknowledgements
We thank Esteban Real, Geoffrey Hinton, Irwan Bello, Jascha SohlDickstein, Jon Shlens, Kathryn Rough, Mohammad Norouzi, Navdeep Jaitly, Niki Parmar, Sam Smith, Simon Kornblith, Vijay Vasudevan, and the Google Brain team for help with this project.
References

Abadi et al. (2016)
Martín Abadi, Paul Barham, Jianmin Chen, Zhifeng Chen, Andy Davis, Jeffrey
Dean, Matthieu Devin, Sanjay Ghemawat, Geoffrey Irving, Michael Isard, et al.
Tensorflow: A system for largescale machine learning.
In USENIX Symposium on Operating Systems Design and
Implementation, volume 16, pp. 265–283, 2016. 
Agostinelli et al. (2014)
Forest Agostinelli, Matthew Hoffman, Peter Sadowski, and Pierre Baldi.
Learning activation functions to improve deep neural networks.
arXiv preprint arXiv:1412.6830, 2014. 
Ba et al. (2016)
Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton.
Layer normalization.
In Advances in Neural Information Processing Systems, 2016. 
Bello et al. (2017)
Irwan Bello, Barret Zoph, Vijay Vasudevan, and Quoc V Le.
Neural optimizer search with reinforcement learning.
In International Conference on Machine Learning, pp. 459–468, 2017. 
Cai et al. (2017)
Han Cai, Tianyao Chen, Weinan Zhang, Yong Yu, and Jun Wang.
Reinforcement learning for architecture search by network
transformation.
arXiv preprint arXiv:1707.04873, 2017. 
Clevert et al. (2015)
DjorkArné Clevert, Thomas Unterthiner, and Sepp Hochreiter.
Fast and accurate deep network learning by exponential linear units
(elus).
arXiv preprint arXiv:1511.07289, 2015. 
Dauphin et al. (2016)
Yann N Dauphin, Angela Fan, Michael Auli, and David Grangier.
Language modeling with gated convolutional networks.
arXiv preprint arXiv:1612.08083, 2016. 
Duan et al. (2016)
Yan Duan, John Schulman, Xi Chen, Peter L Bartlett, Ilya Sutskever, and Pieter
Abbeel.
Rl2: Fast reinforcement learning via slow reinforcement learning.
arXiv preprint arXiv:1611.02779, 2016. 
Elfwing et al. (2017)
Stefan Elfwing, Eiji Uchibe, and Kenji Doya.
Sigmoidweighted linear units for neural network function
approximation in reinforcement learning.
arXiv preprint arXiv:1702.03118, 2017. 
Finn et al. (2017)
Chelsea Finn, Pieter Abbeel, and Sergey Levine.
Modelagnostic metalearning for fast adaptation of deep networks.
arXiv preprint arXiv:1703.03400, 2017. 
Goodfellow et al. (2013)
Ian J Goodfellow, David WardeFarley, Mehdi Mirza, Aaron Courville, and Yoshua
Bengio.
Maxout networks.
In International Conference on Machine Learning, 2013. 
Ha et al. (2016)
David Ha, Andrew Dai, and Quoc V Le.
Hypernetworks.
arXiv preprint arXiv:1609.09106, 2016. 
Hahnloser et al. (2000)
Richard HR Hahnloser, Rahul Sarpeshkar, Misha A Mahowald, Rodney J Douglas, and
H Sebastian Seung.
Digital selection and analogue amplification coexist in a
cortexinspired silicon circuit.
Nature, 405(6789):947, 2000. 
He et al. (2015)
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun.
Delving deep into rectifiers: Surpassing humanlevel performance on
imagenet classification.
In Proceedings of the IEEE international conference on computer
vision, pp. 1026–1034, 2015. 
He et al. (2016a)
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun.
Deep residual learning for image recognition.
In Proceedings of the IEEE conference on computer vision and
pattern recognition, pp. 770–778, 2016a. 
He et al. (2016b)
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun.
Identity mappings in deep residual networks.
In European Conference on Computer Vision, pp. 630–645.
Springer, 2016b. 
Hendrycks & Gimpel (2016)
Dan Hendrycks and Kevin Gimpel.
Bridging nonlinearities and stochastic regularizers with gaussian
error linear units.
arXiv preprint arXiv:1606.08415, 2016. 
Hochreiter & Schmidhuber (1997)
Sepp Hochreiter and Jürgen Schmidhuber.
Long shortterm memory.
Neural Computation, 9(8):1735–1780, 1997. 
Howard et al. (2017)
Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang,
Tobias Weyand, Marco Andreetto, and Hartwig Adam.
Mobilenets: Efficient convolutional neural networks for mobile vision
applications.
arXiv preprint arXiv:1704.04861, 2017. 
Huang et al. (2017)
Gao Huang, Zhuang Liu, Kilian Q Weinberger, and Laurens van der Maaten.
Densely connected convolutional networks.
In Conference on Computer Vision and Pattern Recognition,
2017. 
Ioffe & Szegedy (2015)
Sergey Ioffe and Christian Szegedy.
Batch normalization: Accelerating deep network training by reducing
internal covariate shift.
In International Conference on Machine Learning, pp. 448–456, 2015. 
Jarrett et al. (2009)
Kevin Jarrett, Koray Kavukcuoglu, Yann LeCun, et al.
What is the best multistage architecture for object recognition?
In 2009 IEEE 12th International Conference on Computer Vision,
2009. 
Kingma & Ba (2015)
Diederik Kingma and Jimmy Ba.
Adam: A method for stochastic optimization.
In International Conference on Learning Representations, 2015. 
Klambauer et al. (2017)
Günter Klambauer, Thomas Unterthiner, Andreas Mayr, and Sepp Hochreiter.
Selfnormalizing neural networks.
arXiv preprint arXiv:1706.02515, 2017. 
Krizhevsky & Hinton (2009)
Alex Krizhevsky and Geoffrey Hinton.
Learning multiple layers of features from tiny images.
Technical report, Technical report, University of Toronto, 2009. 
Krizhevsky et al. (2012)
Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton.
Imagenet classification with deep convolutional neural networks.
In Advances in Neural Information Processing Systems, pp. 1097–1105, 2012. 
Maas et al. (2013)
Andrew L Maas, Awni Y Hannun, and Andrew Y Ng.
Rectifier nonlinearities improve neural network acoustic models.
In International Conference on Machine Learning, volume 30,
2013. 
Miech et al. (2017)
Antoine Miech, Ivan Laptev, and Josef Sivic.
Learnable pooling with context gating for video classification.
arXiv preprint arXiv:1706.06905, 2017. 
Naik & Mammone (1992)
Devang K Naik and RJ Mammone.
Metaneural networks that learn by learning.
In Neural Networks, 1992. IJCNN., International Joint
Conference on, volume 1, pp. 437–442. IEEE, 1992. 
Nair & Hinton (2010)
Vinod Nair and Geoffrey E Hinton.
Rectified linear units improve restricted boltzmann machines.
In International Conference on Machine Learning, 2010. 
Parascandolo et al. (2016)
Giambattista Parascandolo, Heikki Huttunen, and Tuomas Virtanen.
Taming the waves: sine as activation function in deep neural
networks.
2016. 
Qiu & Cai (2017)
Suo Qiu and Bolun Cai.
Flexible rectified linear units for improving convolutional neural
networks.
arXiv preprint arXiv:1706.08098, 2017. 
Ravi & Larochelle (2016)
Sachin Ravi and Hugo Larochelle.
Optimization as a model for fewshot learning.
2016. 
Real et al. (2017)
Esteban Real, Sherry Moore, Andrew Selle, Saurabh Saxena, Yutaka Leon Suematsu,
Quoc Le, and Alex Kurakin.
Largescale evolution of image classifiers.
arXiv preprint arXiv:1703.01041, 2017. 
Russakovsky et al. (2015)
Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma,
Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al.
Imagenet large scale visual recognition challenge.
International Journal of Computer Vision, 115(3):211–252, 2015. 
Schmidhuber (1987)
Jurgen Schmidhuber.
Evolutionary principles in selfreferential learning.
On learning how to learn: The metameta… hook.) Diploma
thesis, Institut f. Informatik, Tech. Univ. Munich, 1987. 
Schulman et al. (2017)
John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov.
Proximal policy optimization algorithms.
arXiv preprint arXiv:1707.06347, 2017. 
Shang et al. (2016)
Wenling Shang, Kihyuk Sohn, Diogo Almeida, and Honglak Lee.
Understanding and improving convolutional neural networks via
concatenated rectified linear units.
In International Conference on Machine Learning, pp. 2217–2225, 2016. 
Srivastava et al. (2015)
Rupesh Kumar Srivastava, Klaus Greff, and Jürgen Schmidhuber.
Highway networks.
arXiv preprint arXiv:1505.00387, 2015. 
Szegedy et al. (2016)
Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew
Wojna.
Rethinking the inception architecture for computer vision.
In The IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), June 2016. 
Szegedy et al. (2017)
Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander A Alemi.
Inceptionv4, inceptionresnet and the impact of residual connections
on learning.
In AAAI, pp. 4278–4284, 2017. 
Thrun & Pratt (2012)
Sebastian Thrun and Lorien Pratt.
Learning to learn.
Springer Science & Business Media, 2012. 
Tieleman & Hinton (2012)
Tijmen Tieleman and Geoffrey Hinton.
Lecture 6.5rmsprop: Divide the gradient by a running average of its
recent magnitude.
COURSERA: Neural networks for machine learning, 4(2):26–31, 2012. 
van den Oord et al. (2016)
Aaron van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex
Graves, et al.
Conditional image generation with pixelcnn decoders.
In Advances in Neural Information Processing Systems, pp. 4790–4798, 2016. 
Vaswani et al. (2017)
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones,
Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin.
Attention is all you need.
In Advances in Neural Information Processing Systems, 2017. 
Wang et al. (2016)
Jane X Wang, Zeb KurthNelson, Dhruva Tirumala, Hubert Soyer, Joel Z Leibo,
Remi Munos, Charles Blundell, Dharshan Kumaran, and Matt Botvinick.
Learning to reinforcement learn.
arXiv preprint arXiv:1611.05763, 2016. 
Wu et al. (2016)
Yuhuai Wu, Saizheng Zhang, Ying Zhang, Yoshua Bengio, and Ruslan R
Salakhutdinov.
On multiplicative integration with recurrent neural networks.
In Advances in Neural Information Processing Systems, pp. 2856–2864, 2016. 
Xu et al. (2015)
Bing Xu, Naiyan Wang, Tianqi Chen, and Mu Li.
Empirical evaluation of rectified activations in convolutional
network.
arXiv preprint arXiv:1505.00853, 2015. 
Zagoruyko & Komodakis (2016)
Sergey Zagoruyko and Nikos Komodakis.
Wide residual networks.
In British Machine Vision Conference, 2016. 
Zhong et al. (2017)
Zhao Zhong, Junjie Yan, and ChengLin Liu.
Practical network blocks design with qlearning.
arXiv preprint arXiv:1708.05552, 2017. 
Zhou et al. (2017)
Guorui Zhou, Chengru Song, Xiaoqiang Zhu, Xiao Ma, Yanghui Yan, Xingya Dai, Han
Zhu, Junqi Jin, Han Li, and Kun Gai.
Deep interest network for clickthrough rate prediction.
arXiv preprint arXiv:1706.06978, 2017. 
Zoph & Le (2016)
Barret Zoph and Quoc V Le.
Neural architecture search with reinforcement learning.
In International Conference on Learning Representations, 2016. 
Zoph et al. (2017)
Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V Le.
Learning transferable architectures for scalable image recognition.
arXiv preprint arXiv:1707.07012, 2017.
https://www.groundai.com/project/searchingforactivationfunctions/