SCAI Question 5

Models & Architecture Derived From Natural Intelligence

How do we leverage developmental models and architecture derived from natural intelligence to create new paradigms of AI?

Context & Assumptions

Human intelligence and cognition have capabilities and performance characteristics that are currently unmatched by the best AI systems. While some AI systems outperform specific human capabilities, they fall short in generalisation capabilities and learning efficiency. Natural intelligence is far more flexible, adaptive, responsive, and energy efficient. The brain and our understanding of its functional and cognitive architecture offers a possible reference to implement an intelligence with these performance characteristics.

The functional and developmental organisation of natural intelligence has a considerable impact on how intelligent capabilities form and perform. Hardware substrates in the brain (e.g., neurons) differ from the artificial hardware substrates (e.g., processing units). For instance, the hierarchical structure of spatial reasoning in the brain, between cortical grid cells and hippocampal place cells, provides considerable computational efficiency and robustness that robots are only now beginning to match. In addition, the ability of the brain to perform in-memory computation provides significant efficiencies that cannot be matched by the current separation between computation and memory in existing CPUs and GPUs. There is also evidence that the cognitive architecture of the brain is genetically encoded at birth - the core knowledge hypothesis suggests a strong prior over concepts such as places, objects and motor skill which is already present in biology.

The increased understanding of the functional architecture and performance of the brain provided by cognitive neuroscience can give us new paradigms to develop more capable forms of artificial intelligence. The cognitive sciences (neuroscience, psychology, linguistics, philosophy of mind, anthropology and artificial intelligence) study different aspects of natural intelligence that can be used to inform the design of more capable AI systems. Conversely, progress in AI creates opportunities for understanding brain functions, human cognition, psychology, and development.

Question

How do we leverage developmental models and architectures derived from natural intelligence to create new paradigms of AI?

Answering this core question is tightly coupled to the following additional questions:

Evaluation questions

• Which aspects of natural intelligence cannot currently be replicated by existing AI approaches? This is a moving target, but it is crucial to understand precisely how natural intelligence outperforms existing artificial intelligent systems, and at which tasks.

Structural questions

• What is the right functional decomposition of intelligence that enables these levels of performance and capabilities? The functional relations implicitly define a structure that may be reflected in the structure of the brain.

• What are the intermediate hierarchical structures in the brain that organise neurons into functional reasoning and cognition? While cognition and intelligence do not need to be implemented by neurons, there are existing models of artificial intelligence represented using spiking neural models. An additional intermediate hierarchical structure is required to organise artificial spiking neural models into purposeful computation to allow program synthesis.

• The functional decomposition and cognitive architecture of natural intelligence imply specific and powered inductive biases. What are these inductive biases inherent in natural cognitive architectures?

• How can biological models of motor skills be acquired and composed by artificial intelligence? It is clear that natural intelligence can acquire low-level motor skills efficiently, and incorporate these skills as concepts into higher-level reasoning. There is further evidence from evolutionary biology that developing these low-level skills took considerably more evolutionary time than higher-level reasoning, and may be considered the true cognitive substrate of intelligence.

Performance questions

• Can computational architectures inspired by the cognitive architectures of the brain change, adapt and evolve as easily as the brain does? The power consumed by in-silico intelligence dwarfs the power consumed by biological intelligence, but with a fraction of the performance.

• Can computational architectures inspired by the cognitive architectures of the brain match the energy efficiency of the brain? There is considerable evidence that when a biological agent encounters new scenarios, it is quickly able to adapt to the scenario by reusing previous experience.

• Do the cognitive architectures of the brain implicitly encode an inductive bias that is aligned with human values and judgement?

• How can we ensure and maintain alignment between artificially intelligent agents and humanity?

Indicators of Progress

We expect that the best approaches to answering these questions will include:

• Leveraging development models of natural intelligence and insights from neuroscience and cognitive science and implementing these models in architectures inspired by them. We expect that one candidate form of these models and architectures will be hierarchical compositional probabilistic models that can be reused.

• We also expect that one candidate form of these models that allows the structure of the architecture to be learned and to be adapted from, includes neurosymbolic representations.

• We will also require new theories of model integration (which is not the same as interoperability) and techniques for learning to be used for integrating component models.

• Designing neuromorphic AI software and hardware, and examining the benefits that can be gained by neuromorphic and neurosymbolic approaches.

We expect that the best approaches to measuring progress in answering these questions will be:

• Benchmarking brain-inspired architectures and cognitive systems on tasks relative to human/natural performance.

• Benchmarking on task specialisation, generalisation and few-shot learning.

• Benchmarking developmental models, that allow the cognitive architecture to develop and adapt over time. For example, tasks that have an internal hierarchy with different levels of abstraction that are currently hand-specified (e.g., perception systems driving task-and-motion planning systems) can be derived automatically using developmental models derived from natural intelligence.