Autonomous Linguistic Emergence in neural Networks

Deep neural networks (DNNs) are specialized computational models lacking a standard interface. If a complex task requires different DNNs, an ad-hoc connection must be laboriously designed.

Inspired by human language, ALiEN wants to replace such ad-hoc interfaces with generic communication protocols optimized for ease of learning by DNNs that might have different architectures and functions.

ALiEN “languages” are not hand-crafted: they emerge by training DNNs to share information through communication, offering the scalability and robustness to noise that is an asset of learned systems.A first set of experiments will study, in tightly controlled settings, the impact of input, training community size and communication channel on the expressiveness and ease of acquisition of emergent protocols.

The emergence of general protocols will be encouraged by a training environment characterized by varied inputs and interaction among numerous DNNs.

The best emerged protocols will also be taught in a supervised way to new DNNs, with the final aim of establishing a “universal” DNN language.

Next, I will explore how emergent protocols can help interfacing out-of-the box, state-of-the-art DNNs, only requiring the addition of light input and output layers to existing pre-trained models.

Finally, a simplified home automation use case will demonstrate the usefulness of emergent protocols in a scenario that features some of the complexities to be expected in real-life applications.

The project will also thoroughly analyze the emerging protocols, with the concurrent aims of i) identifying and favoring features that make them more expressive and easier to learn; ii) enhancing interpretability; and iii) gathering scientific insights into communication emergence in a non-human “species”.All in all, ALiEN will take a first bold step towards enabling autonomous DNN interaction, and thus genuinely adaptive AI systems.