Cooperative Trust Among Neural Networks Drives Deeper Learning

Feedback is the essence of training. When you receive relevant feedback from any source (e.g,  teacher, friend, data-driven algorithm), you’re better able to adjust and thereby improve your performance. And when the feedback comes continuously from a wide range of authoritative sources, you can use it to develop an intelligent repertoire of responses to a wide range of potential scenarios.

Training an artificial neural network is the process of incorporating feedback that drives the adjustment of weights associated with neurons’ respective contributions to some learning task. But you can’t always be sure that your neural nets—such as those within your convolutional, recurrent, and other deep-learning models—are receiving trustworthy feedback that contributes to their success at the relevant learning task.

Ensuring the effectiveness of your neural nets usually relies on cooperative trust with one or more of the following partner entities that provide statistical feedback:

• Your training-data partners: Supervised learning is the process of feeding forward to a neural network some “ground truth” training data. followed by the feedback (also known as “backpropagation”) of error values that drive adjustment of the weights associated with neural nodes’ accuracy in processing that data. With successive runs of feedforward, backpropagation, and weight updates, neural networks apply optimization methods (e.g., gradient descent) to minimize a statistical loss function and thereby identify the desired pattern with high confidence. This supervised-learning process breaks down if the training data is inaccurately labeled or otherwise misleading. This might happen if, for example, you’ve outsourced labeling of your training data to people who lack sufficient expertise in identifying the attributes of interest in the data (such as when one asks non-oncologists to identify various cancers).
• Your discriminator-algorithm partners: Adversarial learning involves the feeding forward of an authentic-seeming faux pattern or other digital object (eg., photo) that was generated by one neural network (called the “generative network”) for ingest by another neural network (called the “discriminative network”). The former network relies on supervised learning in order to adjust its weights in an effort to “trick” the discriminator into believing that a fabricated pattern is an authentic object that might have come from actual training data. Training the generative network involves feeding it with back-propagated error values sourced from the discriminative network. This feedback indicates the degree to which the discriminative network believes that a particular pattern is either authentic or fabricated. Contrary to its name, this “adversarial” process is in fact collaborative. In other words, as noted here, it falls apart if the discriminative network isn’t feeding back its actual assessment of the authenticity of the pattern being fed forward from the generative network. As “generative adversarial networks” take root in cybersecurity applications, cross-domain anti-spoofing machine-learning applications will fail if hackers’ generative applications are able to gain unauthorized access to feedback from their targets’ discriminative neural nets.
• Your transfer-learning partners: Transfer learning involves feeding forward of relevant statistical learning that was gained algorithmically in other learning tasks in order to reuse and test it in new tasks where it may be relevant. As discussed here, transfer learning may involve applying data, algorithms, weights, and other statistical knowledge that were gained in neural-net apps involving human users to identical or similar apps in which all or some of the users are automated bots. Or it may attempt to transfer learning across AI-driven games, websites, and other applications that were built for diverse purposes and user interaction scenarios (e.g., streaming, collaboration, e-commerce, mobility, industrial, and Internet of Things). Transfer learning falls apart if the extraneous, unfamiliar learning tasks being fed forward to a neural net have little or no relevance to that net’s ability to master the new task. Likewise, transfer learning amounts to nought if the backpropagated feedback (ie., training data or discriminative network error values) being transferred from an extraneous learning-task environment doesn’t help the client neural net to converge its loss function or otherwise learn the unfamiliar task. For example, it may be risky to trust all the statistical learning transferred from human responses in AI-driven software-simulation environments (e.g, flight simulators) to actual operational environments (e.g., auto-piloted autonomous aircraft). That’s because humans may have longer latencies of awareness, decision, and action in real-time scenarios than sensor-equipped intelligent machines operating in those same situations.

Going forward, machine learning developers will need to model a growing range of multi-partner scenarios in which myriad types of neural nets, training data sources, and transfer learning interact under varying degrees of partner trustworthiness. For many sensitive applications, it might be useful to assess how the classification performance of a particular deep learning model is impacted by declines in the accuracy, completeness, or recency of specific training-data partner sources. This can help to data scientists to assess the sensitivity of generative models to faked training data, as well as collateral ability of discriminative nets to distinguish generated images (generated from faux training images) vis-a-vis authentic images associated with the source domains.

If you’re a machine learning developer who wants to take the next step into continuous intelligence, please start your journey at this informative IBM site.