An imitation classifier may trained to imitate an arbitrarily large single network or ensemble. In imitation learning there may an unlimited amount of training data because the task of the imitator machine learning system is only to match the output of the imitated system, regardless of whether the imitated system is correct and even regardless of whether the correct answer is known. The imitation training data may even include data generated at random or data produced by a synthesizer or a generator network.
In imitation learning, any problem of over fitting may be addressed simply by generating more data. In addition, if the imitated system is deterministic, there would be no problem of overlapping probability distributons or a non-zero minimum Bayes error rate. Thus, in imitation learning, the imitator system may be grown to a size and number of learned parameters to achieve arbtrarily accurate performance on the imitation task. Under the control of a human + AI learning management system, the imitator system may be trained to convergence to an optimum cost/performance trade-off.
One way to train a network L to imitate a network K is to have node-to-node is-equal-to knowledge sharing links for network K to network L. The architectures of the two networks do not need to be the same. It is only necessary for some of the nodes in K to be linked to nodes in L. For example, it is sufficient to have each output node in network K linked to an output node in L. However, links to and from nodes other than output nodes may enhance the learning process and have other beneficial effects. For some of these benefits, it is not necessary to link the output nodes. If either system K or system L is an ensemble rather than a single network, additional benefits may be obtained by having knowledge sharing links among the members of the ensemble.
If system K is a very large network or an ensemble, it may be possible to achieve better price/performance by training a smaller network or a smaller ensemble L to imitate the larger system. For this purpose, the outputs of K should be linked to the outputs of L. In addition, selecting some inner nodes of K to link to inner nodes of L may help the learning process.
There are various uses for letting network K have one or more companion networks L, for example:
When a larger, higher performance network is build by adding new structure to an existing network, there is a natural association between any node in the original network and the corresponding node in the new, larger network. Imitation learning or node-to-node knowledge sharing may help to transfer knowledge from the old network to the new one. Either imitation learning or knowledge sharing may be applied even for data for which the base network was not trained. In fact, the data does not even need to be labeled. However, if the data is labeled, the knowledge sharing or imitation learning may be selective, only being applied when the base network gets the correct answer. This process helps the new network retain the knowledge of the base network when it is useful.
by James K Baker and Bradley J Baker