Scene interpretation for sensory substitution and intelligent agents: a neurocomputational approach

我研究的长期目标是开发具有感官替代装置的辅助技术。由于大脑的巨大可塑性和紧密相连的皮质区域，有可能用一种感觉模式来补充和替代另一种感觉模式。然而，目前的感觉替代系统不能分析视觉或听觉场景，因为它们本质上是在没有解释的情况下描述环境。因此，它们使用户的试听或视觉过载，这可能会使用户感到不舒服或不健康。听觉、视觉和多通道场景的解释对于开发与环境和人类交互的智能代理也是非常必要的。但实际上，智能代理的场景解释或多或少是对场景的有限描述和对非常具体应用的限制。*通过这个研究计划，我想在未来5年内尽可能地推动多模式场景解释的研究。*为了最大化影响，在进行研究的同时，我将经常将结果转移到感官替代(FQRNTéquipe)和交互代理的人工智能(CHIST-ERA Iglu)应用中。*为了最大限度地减少环境足迹，算法的设计和实现将有利于具有最低功耗的体系结构。为此，在过去的两年里，我将重点介绍尖峰神经元网络的实现。*我将研究计划限制在视觉来源上(它们是给定形状的移动来源，简单特征的纹理获得，以便在第一次通过场景中粗略地分离多模式对象。*2)感知学习，以构建多模式复杂特征和表示，以保持特征(形状、音色、音调)的可感知组织；*3)对象的语义表征；*4)场景分析和解释。*5)低功率神经形态实现。*随着新大脑假体的开发，感觉替代具有非常强大的潜力来“练习”和重新布线大脑。能够通过场景分析感知环境的智能代理是新的人工智能经济和产业的基石。此外，加拿大迫切需要发展尖峰神经网络和神经形态系统行业。

The long-term goal of my research is the development of assistive technologies with sensory substitution devices. Thanks to the great plasticity of the brain and the densely interconnected cortical areas, it is possible to complement and substitute one sensory modality with another. However current sensory substitution systems are not able to analyze visual or auditory scenes as they are essentially describing the environment without interpretation. Therefore they overload the audition or the vision of the user which might be uncomfortable or unhealthy. Interpretation of auditory, visual and multimodal scenes is also of great need to the development of intelligent agents that interact with their environment and human beings. But actually scene interpretation for intelligent agents is more or less a limited description of scenes and restraints to very specific applications.***Through this research program I want to push as much as possible the research in multimodal scene interpretation for the next 5 coming years.***To maximize the impact and while doing the research, I will frequently transfer results to applications in sensory substitution (FQRNT équipe) and in artificial intelligence for interacting agents (CHIST-ERA IGLU).***To minimize the environmental footprint, the design and realization of the algorithms will favour architectures with the lowest power consumption. To this end, I will focus during the last 2 years on implementations with networks of spiking neurons. ***I am restricting the research program to sources that are visual (they are moving sources of given shapes, texture earning of simple features to coarsely segregate in a first pass the multimodal objects inside the scene.***2) Perceptual learning to build multimodal complex features and representations that preserve the perceptive organization of characteristics (shapes, timbre, pitch);***3) Semantic characterization of objects;***4) Scene analysis and interpretation.***5) Low-power neuromorphic implementations.***Sensory substitution has a very strong potential for "exercises" and rewiring of the brain along with development of new brain prosthesis. Intelligent agents that can perceive their environment through scene analysis is a corner stone to the new AI economy and industry. Also, there is a strong need to develop the Canadian industry of spiking neural networks and neuromorphic systems.