Keynote 1 – Ali Sayed, EPFL

Social Machine Learning

A wide range of learning algorithms are available in the literature, including sophisticated structures that are based on feedforward, recurrent, or convolutional neural networks. The performance of these architectures matches or exceeds human performance in many important applications. However, they are susceptible to adversarial attacks that can drive them to erroneous decisions under minimal perturbations. They are also often trained with data that arise from homogeneous statistical distributions. And, once trained, the internal structure of these systems remains fixed and are expected to deliver reliable decisions thereafter. For all practical purposes, learning is turned off following training. Contrast these situations with learning by humans: they learn from different types of data and even minimal clues are sufficient in many instances. Humans are also more difficult to fool by small perturbations, and they continue to learn and accumulate experiences over time.

Motivated by these considerations, we will discuss one architecture for learning that exploits important characteristics of social interactions. We refer to the new framework as Social Machine Learning, and it consists of two main connected blocks. One block represents the memory component of the learning machine since it will learn the underlying clues, store them, and regularly update them. This ability adds a new level of richness to the learning process and is different from traditional boosting techniques because the processing is fully decentralized.

A second block represents the processing component of the social learning machine, and it consists of a graph structure linking the various clue models. This block performs classification by exploiting repeated social interactions among agents connected by a graph topology. The agents observe heterogeneous data arising from different statistical sources. This ability is different from neural network structures where information flows in a particular direction rather than arbitrarily over the graph edges, and where the feature data feeding into the graph is now highly heterogeneous. Moreover, the interactions among the agents on the graph take advantage of the “wisdom of the crowd” paradigm, which should lead to more robust learning. This is because it is more difficult to deceive a group of agents than an individual agent, especially when different parts of the group are observing different clues, not all of which can be perturbed similarly.  

Analyses based on statistical learning theory indicate that, under reasonable conditions, the social machine learning structure can learn with high confidence. Moreover, the proposed architecture handles heterogeneity in data more gracefully, is able to learn with performance guarantees, is more resilient to attacks by exploiting the power of the group, and enables continuous learning.