This report explores the evolution and current state of neuro- symbolic artificial intelligence, an approach that integrates neural network capabilities with symbolic reasoning. We trace the historical context from early AI aspirations to modern implementations and successes, highlighting key paradigms, and other logical and semantical considerations. We argue against the “scaling is all you need” hypothesis, and point to persistent challenges in reliable symbolic reasoning with deep and large models. We conclude by suggesting that despite numerous implementation choices and the ”broad church” nature of neuro-symbolic AI, these approaches offer the most promising path towards AI systems that combine pattern recognition with robust reasoning, particularly for applications requiring structured knowledge, explainability, and trustworthiness.

The Winograd Schema Challenge—a set of twin sentences involving pronoun reference disambiguation that seem to require the use of commonsense knowledge—was proposed by Hector Levesque in 2011. By 2019, a number of AI systems, based on large pre-trained transformer-based language models and fine-tuned on these kinds of problems, achieved better than 90% accuracy. In this paper, we review the history of the Winograd Schema Challenge and discuss the lasting contributions of the flurry of research that has taken place on the WSC in the last decade. We discuss the significance of various datasets developed for WSC, and the research community's deeper understanding of the role of surrogate tasks in assessing the intelligence of an AI system.

Both activity-dependent (AD) and activity-independent (AI) processes play important roles in neural development. For example, in the development of the vertebrate visual system, molecular guidance cues that are largely activity-independent provide a rough topography of early projections, while activity-dependent refinement of termination zones occurs later on through correlated retinal activity. A key question concerns the nature of the interaction between these processes. Knockout experiments involving the beta2 subunit of nicotinic acetylcholine receptors and bone morphogenic protein (BMP) suggest that these two processes make genuinely separate contributions - but leave open the precise nature of their interaction. In this article we show how a novel, computational framework (dubbed INTEGRATE) can illuminate the scope and limits of the AI-AD interaction, including facts about critical periods and timing.

A challenge to understanding the mind and brain is to integrate biochemical, genetic, developmental and neuroscientific information. We present a system for integrated simulation of biochemistry, neurodevelopment and neural activity within a unifying framework of genetic control. Using this system, we have developed a novel model for the formation of topographic projections. We also simulate activity-dependent developmental processes which underlie receptive field refinement and ocular dominance column formation. As an illustration of the overall utility of integrated neurogenetic simulation, we show how axon guidance and learning together may explain the results of a critical study of topographic map development (Brown et al., 2000, Cell).