P134 Sensory Data Observation is not an Instant Mapping
Chang Sub Kim*
Department of Physics, Chonnam National University, Gwangju 61186, Republic of Korea
*Email: cskim@jnu.ac.kr
Introduction
The brain self-supervises its embodied agent's behavior via perception, learning, and planning action. Researchers have lately accommodated computational algorithms such as error backpropagation [1] and graphical models [2] to enhance our understanding of how the brain works. The accommodated approaches suit reverse-engineering problems but may not account for real brains. This study aims to provide a biologically plausible theory describing sensory generation, synaptic efficacy, and neural activity as all dynamical processes within a physics-attended framework. We address that sensory observation is generally continuous; therefore, one must handle them appropriately, not as an instant mapping prevalent in Kalman filters [3].
Methods We formulate a neurophysical theory for the brain's working under the free energy principle (FEP), advocating that the brain minimizes informational free energy (IFE) for autopoietic reasons [4]. We derive the Bayesian mechanics (BM) that actuates IFE minimization and numerically show how the BM performs the minimization. To this end, we must determine the likelihood and prior probabilities in the IFE, which are nonequilibrium physical densities in the biological brain. Using stochastic-thermodynamic methods, we specify them as path probabilities and identify variational IFE as a classical action in analytical mechanics [5]. Subsequently, we apply the principle of least action and obtain the brain's neural equations of functional motion.
Results Our resulting BM governs the co-evolution of the neural state and momentum variables; the momentum variable represents prediction error in the predictive coding framework [6]. Figure 1 depicts a sensory stream observed in continuous time, contrasting with discrete Kalman emission. We have numerically explored static and time-dependent sensory inputs for various cognitive operations such as passive perception, active inference, and learning synaptic weights. Our results reveal optimal trajectories, manifesting the brain's minimization of the IFE in neural phase space. In addition, we will present the neural circuitries implied by the BM, reflecting a network of neural nodes in the generic cortical column.
Discussion We argued that sensory data generation is a dynamical process, which we incorporated into our formulation for IFE minimization. Our minimization procedure does not invoke the gradient descent (GD) methods in conventional neural networks but arises naturally from the Hamilton principle. In contrast to quasistatic GD updating, our approach can handle fast, time-varying sensory inputs and provides continuous trajectories of least action, optimizing noisy neuronal dynamics. Furthermore, our theory resolved the issue of the lack of local error representation by revealing the momentum variable as representing local prediction error; we also uncovered its neural equations of motion.
Figure 1. Schematic of sensory data observation. The sensory stream is generally continuous, as depicted in the blue noise curve; the neural response is drawn as the red trajectory, retrodicting the sensory causes in continuous time. In contrast, the prevailing Bayesian filtering in the literature handles sensory observation as a discrete mapping delineated by vertical dashed arrows. Acknowledgements Not applicable. References ● https://doi.org/10.1016/j.tics.2018.12.005 ● https://doi.org/10.1016/j.jmp.2021.102632 ● http://dx.doi.org/10.1115/1.3662552 ● https://doi.org/10.1038/nrn2787 ● Landau, L. D., & Lifshitz, E. M. (1976). Mechanics: Course of theoretical physics. Volume 1. 3rd edition. Amsterdam: Elsevier. ● https://doi.org/10.1038/4580
