EBRAIN Ensembles
Neuroscience provides a rich source of inspiration for new types of algorithms and architectures, independent of and/or complementary to the mathematical and logic-based methods and ideas that have largely dominated traditional approaches to AI. Naturally, a better understanding of biological brains could play a vital role in AI applications, including building intelligent industrial machines, tools, appliances, shop floors and manufacturing processes. However, most neurological models, so far, focus on a specific function or a singular task, which may not be suitable for an industrial setup of machine ensembles in shop floors.
Digitally transformed industrial manufacturing systems appear as collective adaptive systems that consist of diverse heterogeneous entities acting autonomously but having to cooperate with each other to accomplish collective tasks. As each entity is autonomous, it is characterized by its own behavior. To enable collaboration, it can also expose functionalities to the outside world. As such, while entities preserve individuality, they still can form collectives for collaboration, much like individual brains form social structures. For the next generation of digital industrial manufacturing, it is ultimately important to address the operational principles of machine ensembles as socially adaptive brain ensembles
In the EBRAIN-Ensembles project, the question is at what level of human cognition we should look for an industrial manufacturing environment, in which we have an evolving workflow composed of spatially as well as temporally related individual tasks, coherently moving forward. A lot of successful research so far, specifically addressed how a generic neural network (NNW) can be trained to perform tasks that require explicit structural knowledge. Historically, we have NNWs with the functionality of back-propagation. More recently, generative adversarial networks, in which two NNWs are contesting each other in a zero-sum game, bring the advantage of mimicking the data, which is useful in situations where a dataset is not available. Ultimately reinforcement learning, which is based on the concepts of rewards and punishments and learning by trial and error, is combined with a convolutional neural network (termed as deep Q-learning network), able to generate intelligence at a superhuman level. All this suggests that deep neural networks can be trained to emulate the rule-like and structured behaviors that characterize behavioral cognition. In contrast to traditional approaches, EBRAIN-Ensembles will be built on top of the brain architecture references and EBRAINS Neuromorphic Compute Platform developed within HBP.
Goals
Fueled by initiatives like Industry 4.0, the industrial manufacturing paradigm has already evolved from mass production to mass customization and is rapidly moving towards personalized production. A major contributor to this quest is AI. However, Industrial AI Eco-Systems prevalent these days are task-specifical, emphasizing just machine learning, leaving behind recognition, multi-modal perception, scene understanding and comprehension. This narrow focus limits the true potential of AI in complex industrial environments where dynamic situations demand more human-like cognitive abilities. Current systems often struggle with adaptability, real-time decision-making in unforeseen circumstances, and the integration of diverse data streams from various sensors. In contrast, EBRAIN-Ensembles will leverage recent understanding in neuroscience and its subsequent developments in computational neuroscience, to empower novel computing paradigms for industrial use-cases. Thus, seeking to replicate the brain's remarkable capacity for learning, adaptation, and generalization.
To this end, the project aims to establish foundational knowledge around human perception, formalize and model principles, derive sensing and compute paradigms, and develop a flexible reasoning engine.
Approach
The approach used in the project focused on (i) formalizing and modeling the core ingredients of human intelligence by identifying the aspects of core knowledge necessary for general-purpose capabilities, which encompass the workflows of an industrial process, (ii) creating efficient hierarchical data representations emulating human perception with digital (sensor) inputs devices based on multi-sensor computational perception and (iii) developing a collective adaptive reasoning engine as a brain ensemble.
Expected and Achieved Results
The project successfully achieved the following key results. (i) We successfully identify and formalize the core knowledge aspects crucial for general-purpose capabilities within industrial processes. By comprehensive modeling of various industrial workflows and deducing a deep understanding of their underlying logic and dependencies and by developing a robust framework tailored to the operational realities and requirements of industrial environments. (ii) Designed and implemented efficient hierarchical data representations that effectively emulate human perceptual processes that are inspired by multi-sensor computational perception of humans, and thus allowing for the processing and interpretation of diverse digital sensor inputs and demonstrated the ability to extract meaningful insights and patterns from sensor data, mirroring the efficiency and selectivity of human perception. (iii) We developed a novel distributed and collaborative reasoning that will consequently empower systems to adapt and learn from dynamic conditions. The implemented ensemble mechanisms are able to collectively analyze data, infer conclusions, and make decisions and thus being more than an individual component.
