Search

This paper highlights the overall endeavors of the NSF AI Institute for Future Edge Networks and Distributed Intelligence (AI‐EDGE) to create a research, education, knowledge transfer, and workforce development environment for developing technological leadership in next‐generation edge networks (6G and beyond) and artificial intelligence (AI). The research objectives of AI‐EDGE are twofold: "AI for Networks" and "Networks for AI." The former develops new foundational AI techniques to revolutionize technologies for next‐generation edge networks, while the latter develops advanced networking techniques to enhance distributed and interconnected AI capabilities at edge devices. These research investigations are conducted across eight symbiotic thrust areas that work together to address the main challenges towards those goals. Such a synergistic approach ensures a virtuous research cycle so that advances in one area will accelerate advances in the other, thereby paving the way for a new generation of networks that are not only intelligent but also efficient, secure, self‐healing, and capable of solving large‐scale distributed AI challenges. This paper also outlines the institute's endeavors in education and workforce development, as well as broadening participation and enforcing collaboration. [ABSTRACT FROM AUTHOR]

We study the implicit bias of gradient descent methods in solving a binary classification problem over a linearly separable data set. The classifier is described by a non‐linear ReLU model and the objective function adopts the exponential loss function. We first characterize the landscape of the loss function and show that there can exist spurious asymptotic local minima besides asymptotic global minima. We then show that gradient descent (GD) can converge to either a global or a local max‐margin direction or may diverge from the desired max‐margin direction in a general context. For stochastic gradient descent (SGD), we show that it converges in expectation to either the global or the local max‐margin direction if SGD converges. We further explore the implicit bias of these algorithms in learning a multineuron network under certain stationary conditions and show that the learned classifier maximizes the margins of each sample pattern partition under the ReLU activation. [ABSTRACT FROM AUTHOR]