Publications

Building autonomous --- i.e., choosing goals based on one's needs -- and adaptive -- i.e., surviving in ever-changing environments -- agents has been a holy grail of artificial intelligence (AI). A living organism is a prime example of such an agent, offering important lessons about adaptive autonomy. Here, we focus on interoception, a process of monitoring one's internal environment to keep it within certain bounds, which underwrites the survival of an organism. To develop AI with interoception, we need to factorize the state variables representing internal environments from external environments and adopt life-inspired mathematical properties of internal environment states. This paper offers a new perspective on how interoception can help build autonomous and adaptive agents by integrating the legacy of cybernetics with recent advances in theories of life, reinforcement learning, and neuroscience.

Abstract The presence of interictal epileptiform discharges on EEG may indicate increased epileptic seizure risk. In highly epileptogenic lesions, such as cortical tubers in tuberous sclerosis, these discharges can be recorded with intracranial stereotactic EEG as part of the evaluation for epilepsy surgery. Yet the network mechanisms that underwrite the generation and spread of these discharges remain poorly understood. Here, we investigate the dynamics of interictal epileptiform discharges using a combination of quantitative analysis of invasive EEG recordings and mesoscale neural mass modelling of cortical dynamics. We first characterise spatially organised local dynamics of discharges recorded from 36 separate tubers in eight patients with tuberous sclerosis. We characterise these dynamics with a set of competing explanatory network models using dynamic causal modelling. Bayesian model comparison of plausible network architectures suggests that the recurrent coupling between neuronal populations within, as well as adjacent to, the tuber core explains the travelling wave dynamics observed in these patient recordings. Our results indicate that tuber cores are the spatial sources of interictal discharges that behave like travelling waves with dynamics most probably explained by locally recurrent tuber–perituberal networks. This view integrates competing theories regarding the pathological organisation of epileptic foci and surrounding cortex in patients with tuberous sclerosis by through coupled oscillator dynamics. This recurrent coupling can explain the spread of ictal dynamics and also provide an explanation interictal discharge spread. In the future, we will explore the possible implications of our findings for epilepsy surgery approaches in tuberous sclerosis. image Key points Interictal epileptiform discharges (IEDs) are abnormal electrical patterns observed in the brains of people with epilepsy and may indicate seizure risk. In tuberous sclerosis, a condition causing epileptic lesions called cortical tubers, IEDs spread from the tuber core to surrounding brain tissue, forming travelling waves. This study used invasive EEG recordings and mathematical models to identify that recurrent connections between the tuber core and its surroundings explain this wave‐like spread. Further in silico simulations demonstrate that this recurrent network architecture supports both interictal discharges and seizure‐like dynamics under different levels of local inhibition

Hierarchies feature prominently in anatomical accounts of cortical organisation. An open question is which computational (algorithmic) processes are implemented by these hierarchies. One renowned hypothesis is that cortical hierarchies implement a model of the world's causal structure and serve to infer environmental states from sensory inputs. This view, which casts perception as hierarchical Bayesian inference, has become a highly influential concept in both basic and clinical neuroscience. So far, however, a direct correspondence between the predicted order of hierarchical Bayesian computations and the sequence of evoked neuronal activity has not been demonstrated. Here, we present evidence for this correspondence from neuroimaging and electrophysiological data in healthy volunteers. Trial-wise sequences of hierarchical computations were inferred from participants' behaviour during a social learning task that required multi-level inference about intentions. We found that the temporal sequence of neuronal activity matched the order of computations as predicted by the theory. These findings provide strong evidence for the operation of hierarchical Bayesian inference in human cortex. Furthermore, our approach offers a novel strategy for the combined computational-physiological phenotyping of patients with disorders of perception, such as schizophrenia or autism.

Abstract The social environment presents the human brain with the most complex information processing demands. The computations that the brain must perform occur in parallel, combine social and nonsocial cues, produce verbal and nonverbal signals and involve multiple cognitive systems, including memory, attention, emotion and learning. This occurs dynamically and at timescales ranging from milliseconds to years. Here, we propose that during social interactions, seven core operations interact to underwrite coherent social functioning; these operations accumulate evidence efficiently—from multiple modalities—when inferring what to do next. We deconstruct the social brain and outline the key components entailed for successful human–social interaction. These include (i) social perception; (ii) social inferences, such as mentalizing; (iii) social learning; (iv) social signaling through verbal and nonverbal cues; (v) social drives (e.g. how to increase one’s status); (vi) determining the social identity of agents, including oneself and (vii) minimizing uncertainty within the current social context by integrating sensory signals and inferences. We argue that while it is important to examine these distinct aspects of social inference, to understand the true nature of the human social brain, we must also explain how the brain integrates information from the social world.

Inferences about brain function, using functional neuroimaging data, require models of how the data were derived. A variety of models are used in practice that range from conceptual models of functional anatomy to nonlinear mathematical models of hemodynamic responses (e.g. as measured by functional magnetic resonance imaging, fMRI) and neuronal responses. In this review, we discuss the most important models used to analyse functional imaging data and demonstrate how they are interrelated. Initially, we briefly review the anatomical foundations of current theories of brain function on which all mathematical models rest. We then introduce some basic statistical models (e.g. the general linear model) used for making classical (i.e. frequentist) and Bayesian inferences about where neuronal responses are expressed. The more challenging question, how these responses are caused, is addressed by models that incorporate biophysical constraints (e.g. forward models from the neural to the hemodynamic level) and/or consider causal interactions between several regions, i.e. models of effective connectivity. Some of the most refined models to date are neuronal mass models of electroencephalographic (EEG) responses. These models enable mechanistic inferences about how evoked responses are caused, at the level of neuronal subpopulations and the coupling among them.