This is a Pytorch implementation of the Online Deep Learning: Learning Deep Neural Networks on the Fly paper. This algorithm contains a new backpropagation approach called Hedge Backpropagation and it is useful for online learning. In this algorithm you model a overnetwork architeture and the algorithm will try to turn on or turn off some of the hidden layers automatically. This algorithm uses the first hidden layer to train/predict but if it is going bad it starts to use another layers automatically. For more informations read the paper in the 'References' section.

pip install onn

#Importing Library
import numpy as np
from onn.OnlineNeuralNetwork import ONN

#Starting a neural network with feature size of 2, hidden layers expansible until 5, number of neuron per hidden layer = 10 #and two classes.
onn_network = ONN(features_size=2, max_num_hidden_layers=5, qtd_neuron_per_hidden_layer=10, n_classes=2)

#Do a partial training
onn_network.partial_fit(np.asarray([[0.1, 0.2]]), np.asarray([0]))
onn_network.partial_fit(np.asarray([[0.8, 0.5]]), np.asarray([1]))

#Predict classes
predictions = onn_network.predict(np.asarray([[0.1, 0.2], [0.8, 0.5]]))

Predictions -- array([1, 0])

• The algortihm works with batch now. (It is not recommended because this is an online approach. It is useful for experimentation.)
• The algorithm can use CUDA if available. (If the network is very small, it is not recommended. The CPU will process more fast.)

The ONN_THS acts like a non-linear contextual bandit (a reinforcement learning algorithm). This algorithm works with the non-linear exploitation factor (ONN) plus an exploration factor provided by Thompson Sampling algorithm. The ONN_THS works with 'select' and 'reward' actions. For more detailed examples, please look at the jupyter notebook file in this repository. This algorithm was created by me to solve a problem in the company I work.

The great thing about this algorithm is that it can be used in an online manner and it has a non-linear exploitation. This algorithm can learn different kind of data in a reinforcement learning way.

#Importing Library
import numpy as np
from onn.OnlineNeuralNetwork import ONN_THS

#Starting a neural network with feature size of 2, hidden layers expansible until 5, number of neuron per hidden layer = 10 #and two classes.
onn_network = ONN_THS(features_size=2, max_num_hidden_layers=5, qtd_neuron_per_hidden_layer=10, n_classes=2)

#Select an action
arm_selected, exploration_factor = onn_network.predict(np.asarray([[0.1, 0.2]]))

#Reward an action
onn_network.partial_fit(np.asarray([[0.1, 0.2]]), np.asarray([arm_selected]), exploration_factor)

• Alison de Andrade Carrera
• Fábio Silva Vilas Boas
• Daniel Araújo

• Online Deep Learning: Learning Deep Neural Networks on the Fly
• A Survey of Online Experiment Design with the Stochastic Multi-Armed Bandit