Periodic Reporting for period 1 - DIMENSIVE (Data-driven Inference of Models from Embodied Neural Systems In Vertebrate Experiments)
Período documentado: 2020-05-12 hasta 2022-05-11
A crucial challenge in modern neuroscience is the development of methodological tools for incorporating behaviour into the study of brain and mind (Krakauer et al., 2017; Pessoa et al., 2022). Neural recordings usually generate large-scale, non-stationary data, obscuring our understanding of how organisms extract, generate and leverage valuable information from interactions with their environment. Accurately explaining the role of neural activity in information processing involves resolving various conceptual, technical and methodological issues at many levels of organization (from neural biochemistry to behaviour and learning).
The state-of-the-art view is that the activity of neural populations organizes dynamically with behaviour in a complex, ongoing interaction, being steered but not deterministically controlled by sensory input (Harris, 2005). This complex interplay between brain, body and environment is an open problem in academic fields from neuroscience to cognitive robotics. Even simple behaviours show complex interdependencies that involve many levels of organization and non-stationary dynamics, and current efforts are still too limited in providing an integrative view. Thus, there is a pressing demand for mathematical, physical and computational theories to disentangle and explain the complex dynamics of nonequilibrium neural computation (Grün, 2009).
This MSCA aimed to develop and apply mathematical tools and experimental setups to study large neural systems during animal behaviour, i.e. in out-of-equilibrium conditions driven by sensorimotor interaction.
The state-of-the-art view is that the activity of neural populations organizes dynamically with behaviour in a complex, ongoing interaction, being steered but not deterministically controlled by sensory input (Harris, 2005). This complex interplay between brain, body and environment is an open problem in academic fields from neuroscience to cognitive robotics. Even simple behaviours show complex interdependencies that involve many levels of organization and non-stationary dynamics, and current efforts are still too limited in providing an integrative view. Thus, there is a pressing demand for mathematical, physical and computational theories to disentangle and explain the complex dynamics of nonequilibrium neural computation (Grün, 2009).
This MSCA aimed to develop and apply mathematical tools and experimental setups to study large neural systems during animal behaviour, i.e. in out-of-equilibrium conditions driven by sensorimotor interaction.
This MSCA pursued contributions in three categories:
1. The development of mathematical methods for inferring large-scale, non-equilibrium models of neural activity
2. The use of information-theoretical tools to characterize information integration in large networks
3. The study of predictive mechanisms in close-loop sensorimotor loops
1. Mean-field methods for studying non-equilibrium neural populations. Advances in high-throughput neural recording technologies provide unique opportunities to study information processing in the brain during behaviour (Lin et al., 2022). Nevertheless, the data generated by such open neural systems generally exhibits out-of-equilibrium dynamics and large fluctuations that are difficult to capture with current tools, thus demanding new theoretical and computational methods for its study. One of the main outputs of this MSCA, published in Nature Communications, was the development of a novel framework unifying advanced mean-field methods (based on ideas from information geometry applied to nonequilibrium, stochastic neural networks, Aguilera et al., 2021). Such methods can be used both to approximate the dynamics of neural spike trains and to infer the parameters of a model that best characterize recorded data, resulting in better performance than other state-of-the-art mean-field methods even in regions with large fluctuations or critical dynamics. These methods have already been explored in preliminary results in somatosensory cortical populations with promising results (Poc-López & Aguilera, 2021). Future work will test in depth the possibilities of these new, nonequilibrium mean-field methods for inferring models of neural activity during animal behaviour.
2. Integrated information in neural systems. The question of how the brain integrates different information sources into a coherent whole is regarded as a crucial aspect of brain function (Tononi et al., 1994; Wang et al., 2021). Integration has been extensively addressed by methods from information theory and information geometry (Oizumi et al., 2016), providing valuable predictions and explanations in neuroscience (e.g. the role of different brain areas in conscious activity, Tononi et al., 2016). Nevertheless, recent work suggests that integrated information measures might be too simplistic to capture the multiscale nature of integration in complex systems (Rosas et al., 2020). Additionally, integrated information has usually been explored under open-loop assumptions, not taking into account the continuous sensorimotor coupling between a system and its environment. In this MSCA, we have explored how integrated information behaves in very large networks coupled to an external environment (Aguilera & Di Paolo, 2021). In this work, we detail the conditions of system-environment interaction which integrated information can grow with the neural network size, guaranteeing scaling properties of integration. We find that this happens in a specific type of critical phase transition, which we refer to as “critical integration”. In addition, we can use integrated information measures to distinguish the case in which a neural network alone displays high integration of information, in comparison with a case in which a neural network is integrated with its environment and display large integrated information emergent from the interaction of the two.
3. Closed loop prediction and virtual reality setups. Information theory and statistical inference are increasingly underpinning modern algorithms in machine learning methods for intelligent agents (e.g. intrinsic measures, maximum entropy reinforcement learning). In particular, there is a growing interest in Bayesian perspectives on action-oriented prediction and control inspired by nonequilibrium statistical mechanics (Clark, 2013; Friston, 2010) and its relation with theoretical neurosciences (Friston, 2018). Recent experiments in neuroscience, using innovative visual-flow feedback manipulations in virtual reality environments (VR), have suggested that the visual cortex encodes predictive visual feedback information during locomotion (Keller et al., 2012; Torigoe et al., 2021). While the potential contribution of VR setups for studying predictive neural mechanisms integrating behaviour is ground-breaking, research to date is mostly based on heuristic measures of prediction (e.g. correlation with mismatch signals) due to the lack of a solid theoretical framework. This MSCA action has provided a detailed mathematical review pointing out methodological problems in these theories’ underlying assumptions about nonequilibrium brain-environment interaction (Aguilera et al., 2022).
1. The development of mathematical methods for inferring large-scale, non-equilibrium models of neural activity
2. The use of information-theoretical tools to characterize information integration in large networks
3. The study of predictive mechanisms in close-loop sensorimotor loops
1. Mean-field methods for studying non-equilibrium neural populations. Advances in high-throughput neural recording technologies provide unique opportunities to study information processing in the brain during behaviour (Lin et al., 2022). Nevertheless, the data generated by such open neural systems generally exhibits out-of-equilibrium dynamics and large fluctuations that are difficult to capture with current tools, thus demanding new theoretical and computational methods for its study. One of the main outputs of this MSCA, published in Nature Communications, was the development of a novel framework unifying advanced mean-field methods (based on ideas from information geometry applied to nonequilibrium, stochastic neural networks, Aguilera et al., 2021). Such methods can be used both to approximate the dynamics of neural spike trains and to infer the parameters of a model that best characterize recorded data, resulting in better performance than other state-of-the-art mean-field methods even in regions with large fluctuations or critical dynamics. These methods have already been explored in preliminary results in somatosensory cortical populations with promising results (Poc-López & Aguilera, 2021). Future work will test in depth the possibilities of these new, nonequilibrium mean-field methods for inferring models of neural activity during animal behaviour.
2. Integrated information in neural systems. The question of how the brain integrates different information sources into a coherent whole is regarded as a crucial aspect of brain function (Tononi et al., 1994; Wang et al., 2021). Integration has been extensively addressed by methods from information theory and information geometry (Oizumi et al., 2016), providing valuable predictions and explanations in neuroscience (e.g. the role of different brain areas in conscious activity, Tononi et al., 2016). Nevertheless, recent work suggests that integrated information measures might be too simplistic to capture the multiscale nature of integration in complex systems (Rosas et al., 2020). Additionally, integrated information has usually been explored under open-loop assumptions, not taking into account the continuous sensorimotor coupling between a system and its environment. In this MSCA, we have explored how integrated information behaves in very large networks coupled to an external environment (Aguilera & Di Paolo, 2021). In this work, we detail the conditions of system-environment interaction which integrated information can grow with the neural network size, guaranteeing scaling properties of integration. We find that this happens in a specific type of critical phase transition, which we refer to as “critical integration”. In addition, we can use integrated information measures to distinguish the case in which a neural network alone displays high integration of information, in comparison with a case in which a neural network is integrated with its environment and display large integrated information emergent from the interaction of the two.
3. Closed loop prediction and virtual reality setups. Information theory and statistical inference are increasingly underpinning modern algorithms in machine learning methods for intelligent agents (e.g. intrinsic measures, maximum entropy reinforcement learning). In particular, there is a growing interest in Bayesian perspectives on action-oriented prediction and control inspired by nonequilibrium statistical mechanics (Clark, 2013; Friston, 2010) and its relation with theoretical neurosciences (Friston, 2018). Recent experiments in neuroscience, using innovative visual-flow feedback manipulations in virtual reality environments (VR), have suggested that the visual cortex encodes predictive visual feedback information during locomotion (Keller et al., 2012; Torigoe et al., 2021). While the potential contribution of VR setups for studying predictive neural mechanisms integrating behaviour is ground-breaking, research to date is mostly based on heuristic measures of prediction (e.g. correlation with mismatch signals) due to the lack of a solid theoretical framework. This MSCA action has provided a detailed mathematical review pointing out methodological problems in these theories’ underlying assumptions about nonequilibrium brain-environment interaction (Aguilera et al., 2022).
The development of a unified framework for non-equilibrium mean-field approximations (Aguilera et al., 2021) offers an exciting opportunity to systematically advance in the study of large-scale, non-equilibrium biological and social dynamics operating near critical regimes. In addition, since the methods are directly applicable to machine learning, the general introduced here is relevant as well in machine learning applications. In parallel, the advance in defining an information theoretical framework for studying nonequilibrium closed-loop interaction in neural systems (Aguilera & Di Paolo, 2021; Aguilera et al., 2022) prescribes a route towards the study of neural activity in behaving animals. The synergy between these contributions could eventually provide a novel approach applicable to experimental animal neuroimaging research.