Use the future to build the present
Neural Network Architectures
Comment
Stakeholder Type
,
1Quantum Revolution& Advanced AI2HumanAugmentation3Eco-Regeneration& Geo-Engineering4Science& Diplomacy1.11.21.31.42.12.22.32.43.13.23.33.43.54.14.24.34.44.5HIGHEST ANTICIPATIONPOTENTIALAdvancedArtificial IntelligenceQuantumTechnologiesBrain-inspiredComputingBiologicalComputingCognitiveEnhancementHuman Applications of Genetic EngineeringRadical HealthExtensionConsciousnessAugmentation DecarbonisationWorldSimulationFuture FoodSystemsSpaceResourcesOceanStewardshipComplex Systems forSocial EnhancementScience-basedDiplomacyInnovationsin EducationSustainableEconomicsCollaborativeScience Diplomacy
1Quantum Revolution& Advanced AI2HumanAugmentation3Eco-Regeneration& Geo-Engineering4Science& Diplomacy1.11.21.31.42.12.22.32.43.13.23.33.43.54.14.24.34.44.5HIGHEST ANTICIPATIONPOTENTIALAdvancedArtificial IntelligenceQuantumTechnologiesBrain-inspiredComputingBiologicalComputingCognitiveEnhancementHuman Applications of Genetic EngineeringRadical HealthExtensionConsciousnessAugmentation DecarbonisationWorldSimulationFuture FoodSystemsSpaceResourcesOceanStewardshipComplex Systems forSocial EnhancementScience-basedDiplomacyInnovationsin EducationSustainableEconomicsCollaborativeScience Diplomacy

Sub-Field:

1.3.1Neural Network Architectures

Efforts to build brain-like processors take a number of different forms. All of them, however, take inspiration from what neuroscientists have discovered about the structure and operation of the brain. That means building networks of nodes that mimic the action of the brain’s neurons, and having the nodes emit signals in the same way that the neuron soma’s spike to allow neurons to communicate.5 The topology, size and exact nature of the experimental networks vary immensely, because it is not yet clear how large a network has to be, and how interconnected it must be, for it to demonstrate neural-type properties.6 Continued progress is likely to require better theoretical underpinnings, conceptual refinements in computational neuroscience and better models of the brain’s mechanisms.

Understanding the sub-mechanisms of the brain’s component parts will also be important. The cortical structures, the thalamus, the cerebellum, the hippocampus and the basal ganglia all play roles within the brain that could have technological significance if we can learn to replicate their operation. Additionally, it is not enough to map the connections and topology of our networks of neurons. We need to understand the dynamics, synaptic and structural plasticity of the brain, and genetically-defined developmental programs that are responsible for a large portion of the brain’s wiring.

Alternative approaches to brain-inspired computing include those that consider proposals and hypotheses about how things happen at a cognitive level, elucidating rules and descriptions of behaviour and planning rather than seeking to generate these by emulating neuronal activity. There is still debate over whether analogue or digital processing offers the best route to mimicking the brain, and it is possible that a hybrid system, combining the energy efficiency of analog and the precision of digital, might provide competitive performance.

Future Horizons:

×××

5-yearhorizon

Brain structures and sub-structures are mapped and simulated

We have practically useful neuronal architectures for navigation and map formation based on rodents (rat, bat) and insects (bee) brains, plus hippocampus and entorhinal cortex, navigation complex for insects. These circuits help us understand, model, and program practically usable building blocks, or algorithms. They are deployed in small autonomous domestic robots. Emerging neuromorphic technologies include olfaction-inspired chemical sensors; retina- and eye-inspired vision sensors (with fovea, targeted microsaccades, adaptive thresholds and active sensing); active spinal-cord inspired controllers and central pattern generators for flexible snake- or salamander-like robots.

10-yearhorizon

Neuromorphic computing provides useful technology

Based on understanding of human and animal reaching and grasping, soft robots can direct actions at objects in our everyday environments: a robot arm that can grasp and manipulate things, helping with maintenance work at home and in factory floors, support elderly and sick, supporting construction workers etc. Simultaneous localisation and mapping (SLAM) provides robots with mobility in complex indoor and outdoor environments, allowing them to perform inspection tasks in complex and hard-to-reach environments. 3D vision enables artificial systems to perform movement around and seamless interaction — including gesture and gaze-based interaction — with physical objects including humans. Brain-machine interfaces allow processing of biological systems on a low-power chip for control of prosthetic devices, monitoring of heart and brain activity, seizure prediction, and for fast control of computer games with, for example, EEG.

25-yearhorizon

Brain maps begin to show useful, granular detail

We finally understand the processes and causal relationships between different levels of the brain, going from molecular to cellular to circuit to population to cortical micro-area to larger cortical areas and subcortical structures. We know the connectivity and structure as well as the genetic programs and rules of plasticity by which the structure is formed. We know how to cure different malfunctions, and have mathematical models of brain processes on different levels that can be used to implement similar functions to solve technical problems.

Neural Network Architectures - Anticipation Scores

How the experts see this field in terms of the expected time to maturity, transformational effect across science and industries, current state of awareness among stakeholders and its possible impact on people, society and the planet. See methodology for more information.

GESDA Best Reads and Key Resources