Periodic Reporting for period 1 - GROW-AI (Growing Machines Capable of Rapid Learning in Unknown Environments)
Periodo di rendicontazione: 2023-01-01 al 2025-06-30
Despite major advances in the field of artificial intelligence, especially in the field of neural networks, these systems still pale in comparison to even simple biological intelligence. Current machine learning systems take many trials to learn, lack common-sense, and often fail even if the environment only changes slightly. The enormous potential of autonomous machines remains unfulfilled and we still lack robots to fill our dishwashers or go on autonomous search-and-rescue missions. The grand goal of GROW-AI is to create machines with a more general intelligence, allowing rapid adaption in unknown situations. In stark contrast to current neural networks, whose architectures are designed by human experts and whose large number of parameters are optimized directly, evolution does not operate directly on the parameters of biological nervous systems. Instead, these nervous systems are grown and self-organize through a much smaller genetic program that produces rich behavioral capabilities right from birth and the ability to rapidly learn. Neuroscience suggests this "genomic bottleneck" is an important regularizing constraint, allowing animals to generalize to new situations. However, currently there does not exist a solution to creating a similar system artificially. Taking inspiration from the fields of artificial life, neurobiology, and machine learning, in the GROW-AI we aim to learn genomic bottleneck algorithms instead of manually designing them. If successful, this project will greatly improve the autonomy of machines and significantly increase the range of real-world tasks they can solve.
In the first half of the project, we were able to successfully develop the proposed genomic bottleneck approach, we call Neural Developmental Programs (NDPs). This approach is able to grow a fully functioning neural network from a single seed "neuron". Inspired by biological development, the NDP allows networks to self-assemble and evolve solely based on local communication between its components. Unlike conventional AI, which requires extensive manual design, these networks can adapt their shape and structure dynamically, making them more robust.
During the development of the NDP, we identified a key challenge: ensuring phenotypic complexity requires maintaining neuronal diversity, but this diversity comes at the cost of optimization stability. To address this, we introduce two mechanisms: (a) equipping neurons with an intrinsic state inherited upon neurogenesis; (b) lateral inhibition, a mechanism inspired by biological growth, which controls the pace of growth, helping diversity persist. We show that both mechanisms contribute to neuronal diversity and that, equipped with them, NDPs achieve comparable results to existing direct and developmental encodings in complex locomotion tasks.
A significant limitation of most current AI models is that they are trained once and then remain fixed, limiting their ability to handle new challenges. This is in stark contrast to biological neural networks that are characterized by their high degree of plasticity, a core property that enables the remarkable adaptability of natural organisms. Building on the NDP, we were able to develop a class of self-organizing neural networks capable of synaptic and structural plasticity in an activity and reward-dependent manner which we call Lifelong Neural Developmental Program (LNDP). Our results demonstrate the ability of the model to learn from experiences in different control tasks starting from randomly connected or empty networks. We further show that structural plasticity is advantageous in environments necessitating fast adaptation or with non-stationary rewards.
During the development of the NDP, we identified a key challenge: ensuring phenotypic complexity requires maintaining neuronal diversity, but this diversity comes at the cost of optimization stability. To address this, we introduce two mechanisms: (a) equipping neurons with an intrinsic state inherited upon neurogenesis; (b) lateral inhibition, a mechanism inspired by biological growth, which controls the pace of growth, helping diversity persist. We show that both mechanisms contribute to neuronal diversity and that, equipped with them, NDPs achieve comparable results to existing direct and developmental encodings in complex locomotion tasks.
A significant limitation of most current AI models is that they are trained once and then remain fixed, limiting their ability to handle new challenges. This is in stark contrast to biological neural networks that are characterized by their high degree of plasticity, a core property that enables the remarkable adaptability of natural organisms. Building on the NDP, we were able to develop a class of self-organizing neural networks capable of synaptic and structural plasticity in an activity and reward-dependent manner which we call Lifelong Neural Developmental Program (LNDP). Our results demonstrate the ability of the model to learn from experiences in different control tasks starting from randomly connected or empty networks. We further show that structural plasticity is advantageous in environments necessitating fast adaptation or with non-stationary rewards.
The GROW-AI project has made significant advances beyond the state-of-the-art by developing self-growing, lifelong-learning AI systems that evolve and adapt to their environments. Unlike most traditional AI models, which are manually designed and fixed, the approaches developed in this project enable networks to continuously adapt to their environment.
Additionally, typical neural networks are rigid and tied to the specifics of a particular input and output space, which prevents them from being optimized across domains with differing dimensions. Going beyond the current state-of-the-art, our approach allows more structurally flexible neural networks (SFNNs), which enable a single AI model to adjust its architecture dynamically to tackle different problems with different input/output dimensions.
The potential impact of these developments is substantial. In robotics, self-growing AI could create adaptive machines that modify their behavior in real time, making them more effective in unpredictable environments such as disaster response and industrial automation. However, for these innovations to be fully realized, further research is needed to scale these models for real-world applications and further improve their computational efficiency.
Additionally, typical neural networks are rigid and tied to the specifics of a particular input and output space, which prevents them from being optimized across domains with differing dimensions. Going beyond the current state-of-the-art, our approach allows more structurally flexible neural networks (SFNNs), which enable a single AI model to adjust its architecture dynamically to tackle different problems with different input/output dimensions.
The potential impact of these developments is substantial. In robotics, self-growing AI could create adaptive machines that modify their behavior in real time, making them more effective in unpredictable environments such as disaster response and industrial automation. However, for these innovations to be fully realized, further research is needed to scale these models for real-world applications and further improve their computational efficiency.