Research

A framework for artificial “Bayesian Brain.”

Given the growing ubiquity of things like autonomous vehicles, robots, and arrays of edge computing devices and sensors, the zenith of the AI age may end up being a distributed but interconnected ecosystem where both human and human-designed artifacts collaborate seamlessly. A prerequisite for fluent interaction in such ecosystems is human trust, which demands the artifacts to embody a form of intelligence that, at least appears to us, shares a coherent common sense with us. This desideratum leads me to a fundamental inquiry, which also serves as the driving force behind my research:

How can we build machines that reason in a trustworthy manner—adhering to clearly defined rules that mirror idealized principles of human thought?

As von Neumann indicated, to build machines that “think” like us, we must first know exactly what our “thinking” consists of. Two fundamental questions that persistently emerge in any form of human “thinking” are:

  • “What can I know?” –concerning how we perceive the world; and
  • “How shall I do?” –concerning how we act in the world.

Bayesian brain theory (BBT), which models the human brain as a probabilistic inference engine, furnishes a simple and intuitive decomposition of this perception-action loop, i.e. the set of cognitive processes that allow us to interpret sensory signals and respond appropriately. BBT formalizes the core desiderata of human cognition—identifying what we should do to best achieve our goals, given our available information—with two principles of rational inference and decision. The rational inference principle guides how one’s beliefs are affected by perceptions in terms of Bayesian belief updating, while the rational decision principle depicts how beliefs and desires lead to judicious actions.

BBT entails a formal definition and explicit mechanics of the intelligence for human, which opens the door for us to engineer the intelligence for artifacts. Given any feasible principles of inference and decision, the rules to regulate their rationality can be distinguished into 4 pivotal categories:

  • Rules that regulate perceptions;
  • Rules that regulate actions;
  • Rules that regulate how perceptions influence actions; and
  • Rules that regulate how actions influence perceptions.

As such, my research goal can be summarized as “4R for 4R“:

Leveraging 4 sets of Rules embedded in the logic of Bayesian principles to create Reliable, Responsible, and Resilient Robots.

Rule 1: Many Eyes, One Vision

The first, and maybe the most primitive, set of rules is to regulate the rationality of multi-modal perception for robots, which is rooted in the classic “perceptual binding” problem in human cognition: how does the brain integrate signals from different sensory modalities (e.g., vision, audition, touch,…) to yield a unified and consistent percept?

Much like the human nervous system—which senses, communicates, computes, and actuates using distributed and diverse components—robotic systems often acquire multi-modal information via heterogeneous, multi-robot teams. Although diversity enriches the information collected, it also poses significant challenges for distilling consistent knowledge from a complex mixture of various data streams.

Rule 2: Stay Alert to Risk

Another equivalently basic set of rules is about guiding the action of robots in the existence of risk, which stems from the major “residual motor uncertainty” limitation in the physical world: the execution of an intentional movement is never perfectly reliable, i.e. it is noisy.

Human brain inherently performs a sensorimotor integration that uses sensory feedback signals to reduce movement uncertainty. Such a paradigm inspires the robotic systems to correct the motor command on the basis of online sensory information. However, for any online model-based control strategy, there exists an inherent ambiguity: precise models can only be progressively inferred from incrementally available data. To address the disparity, which can be particularly pronounced in early learning stages with limited data, explicitly accounting for this ambiguity from a risk-averse perspective is demanded.

  • Yingke Li, Yifan Lin, Enlu Zhou and Fumin Zhang, “Risk-Aware Model Predictive Control Enabled by Bayesian Learning,” 2022 American Control Conference (ACC), Atlanta, GA, USA, 2022, pp. 108-113, doi: 10.23919/ACC53348.2022.9867207.
  • Yingke Li, Yifan Lin, Enlu Zhou and Fumin Zhang, “Bayesian Risk-averse Model Predictive Control with Consistency and Stability Guarantees,” IEEE Transactions on Automatic Control (TAC), under review.

Rule 3: Optimize Across Scales

While the first two sets of rules equip robots with separate abilities to perform consistent perceptions and risk-aware actions, they omit a critical element: intention. The missing link between “what we know” and “how we do” is the strategic utilization of information to achieve our desires. Due to limited computational resources, a third set of rules is mandated to optimize our decisions with minimal effort.

When navigating decisions amid the challenges of incomplete or noisy information, human cognitive process naturally engages multiple scales of decision-making. Consider the example of crafting a travel plan: typically, we sketch a long-horizon itinerary (e.g., Boston-Atlanta-LA) and determine the short-horizon transportation to the next destination (e.g., by airplane or train). Elaborate transportation planning between subsequent stops can be deferred due to inherent uncertainties and unpredictabilities (e.g., weather conditions). This inclination to defer non-urgent decisions until additional information surfaces mirrors a broader philosophy that leverages Bayesian principles in multi-layer planning with multiple resolutions and horizons.

Rule 4: Curiosity is Knowledge

The third set of rules optimizes the relevance and priority of information in relation to given desires, regardless of how it was obtained. Nevertheless, as the Chinese philosopher Wang Yangming stated, “Knowledge is the beginning of action; action is the completion of knowledge.” Wisdom arises through action, so the final set of rules encourages proactive decision-making to facilitate knowledge acquisition.

Active inference (AIF), a biologically plausible principle that originated in computational neuroscience, formalizes the idea of gathering information under the constraint that preferred outcomes are achieved. AIF mandates a balance between information-seeking and goal-directed behavior, essentially suggesting that intelligence partly consists of a balanced curiosity—a drive to explore in order to optimize long-term rewards.

Recent Post

Top Category