Discussion meeting on Neuroscience, Data Science and Dynamics

The hippocampus supports two important functions: spatial navigation and the storage of episodic memories. Yet, how these two seemingly distinct roles converge in a single circuit remains an open question. In this talk, I will present a neural model that leverages the low-dimensional attractor dynamics of spatial "grid cells" to implement associative, spatial, and episodic memory, thus unifying these distinct functions. Our model, Vector-HaSH (Vector Hippocampal Scaffolded Heteroassociative Memory) operates through a factorization of the content of memories from dynamics that generate error-correcting stable states. This leads to a graceful trade-off between number of stored items and recall detail, unlike the abrupt capacity limits found in classical Hopfield-like memory models. We find that the usage of pre-structured low-dimensional representations also enables high-capacity sequence memorization by recasting the chaining problem of high-dimensional states into one of learning low-dimensional transitions. Further, our presented approach reproduces several hippocampal experiments on spatial mapping and context-dependent representations, and provides a circuit model of the 'memory palaces' used by memory athletes. Thus, this work provides a unified framework for understanding how the hippocampus simultaneously supports spatial mapping, associative memory, and episodic memory.

Nonlinear dynamics aims to elucidate the basic mechanisms necessary to reflect the temporal behavior of a natural system. The data analysis and modeling techniques proposed by artificial intelligence (deep networks, computational reservoirs, recurrent networks, as examples), on the other hand, ostensibly resign the mechanistic vision for a data-oriented modeling paradigm.  In these lectures, these apparently antagonistic approaches will be analyzed in parallel, using as examples my work on the physics of birdsong production and vocal learning.

The basal ganglia are known to influence action selection and modulation of movement vigor, but whether and how they contribute to specifying the kinematics of learned motor skills is not understood. Here, we probe this question by recording and manipulating basal ganglia activity in rats trained to generate complex task-specific movement patterns with rich kinematic structure. We find that the sensorimotor arm of the basal ganglia circuit is crucial for generating the detailed movement patterns underlying the acquired motor skills. Furthermore, the neural representations in the striatum, and the control function they subserve, do not depend on inputs from the motor cortex. Taken together, these results extend our understanding of the basal ganglia by showing that they can specify and control the fine-grained details of learned motor skills through their interactions with lower-level motor circuits.

Flying insects must balance the demands of speed and agility with the precision of their movements to swiftly and accurately respond to environmental stimuli. Achieving this balance requires them to integrate sensory information from various modalities. Of particular importance are visual inputs from their compound eyes and mechanosensory inputs from their antennae, which are crucial for maintaining flight stability. This challenge is particularly pronounced in insects like hawkmoths, which navigate under low light conditions. Prior studies on diverse hawkmoths and other insects have highlighted the critical role of antennal mechanosensory feedback in flight control, akin to the function of halteres in flies. How is such multisensory integration achieved? We addressed this question by conducting recordings from descending neurons in the cervical connective nerve in the Oleander hawkmoths, Daphnis nerii. The moths were provided with visual stimuli comprised of moving spots of light and mechanical stimuli to their antennae. While these stimuli were presented singly or concurrently, we recorded intracellularly from axons of descending interneurons to determine if they respond to one or both stimuli. In addition to a number of exclusively visual or mechanosensory descending neurons, we also identified several neurons that multiplex the visual and mechanosensory signals such that a single neuron encodes both visual stimuli from the compound eyes, and mechanosensory stimuli from the antennal Johnston's organs. Additional experiments at the level of behavior in intact moths reveals that integration of visual and antennal mechanosensory feedback plays a key role in gaze stabilization in flying hawkmoths. Together, these experiments underscore the importance of multisensory integration during flight in hawkmoths.

ynchronization of rhythmic units is essential for various biological functions. The synchronization mechanism is often governed by neural networks. For example, the circadian rhythm in mammals functions by synchronizing the gene expression rhythms of individual neurons within a neural tissue called the suprachiasmatic nucleus. Various movement patterns observed in animal locomotion, such as walking and swimming, are generated by neural networks known as central pattern generators. The synchronization dynamics of oscillator groups strongly depend on the heterogeneity of intrinsic frequencies, noise intensity, and the interaction network. If these factors can be estimated from observations, it can aid in understanding, predicting, and controlling the system, and contribute to elucidating the design principles of robust systems. However, estimation involves many challenges. Among them, estimation becomes exponentially difficult as the model's dimensions and the number of parameters increase. Therefore, it is desirable to assume a model with as low dimensions and as few parameters as possible. In the synchronization phenomena of oscillator groups, the phase oscillator model is expected to be useful. This presentation introduces research on estimation using the phase oscillator model [1,2]. [1]A Matsuki, H Kori, R Kobayashi: Network inference from oscillatory signals based on circle map, arXiv:2407.07445 (2024) [2]F Mori, H Kori: Noninvasive inference methods for interaction and noise intensities of coupled oscillators using only spike time data, Proceedings of the National Academy of Sciences 119, e2113620119 (2022)

Nonlinear dynamics aims to elucidate the basic mechanisms necessary to reflect the temporal behavior of a natural system. The data analysis and modeling techniques proposed by artificial intelligence (deep networks, computational reservoirs, recurrent networks, as examples), on the other hand, ostensibly resign the mechanistic vision for a data-oriented modeling paradigm.  In these lectures, these apparently antagonistic approaches will be analyzed in parallel, using as examples my work on the physics of birdsong production and vocal learning.

How insects find and feed on flowers is crucial for plant pollination. Plants provide several chemical and visual cues to attract insects. However, even as insects find their host flowers, they have the additional challenge of targeting a tiny nectary opening on floral surfaces to reach the sugary reward. This becomes especially challenging for moths and butterflies because they use a long and flexible mouthpart called proboscis to draw the nectary. Additionally, nocturnal hawkmoths feed while hovering in front of flowers, often at very low lights levels during nighttime. How do moths target the tiny nectary hole in flowers, despite the low visual resolution of the nectary opening? We tracked the motion of hawkmoth proboscis as they fed from artificial, 3D-printed flowers and show that hawkmoths systematically explore floral surfaces to detect tactile features such as curvatures to target the nectary. We found that over repeated visits, they preferentially explore in ways that increases the efficiency of finding the nectary. Systematic exploration and targeting objects in the environment require expert control over appendage movements. How do hawkmoths sense and control the proboscis motion to achieve such precise movements? Pilifers are paired bristled organs at the proboscis base and are well placed to provide proprioceptive feedback about relative movements of the proboscis. To study the role of pilifers as proprioceptive organs, we drove lateral motions of the proboscis in anaesthetized head-fixed moths while simultaneously recording neural responses from the pilifer nerve. Our recordings reveal that pilifer mechanosensory neurons are sensitive to lateral motions, either to the left or to the right. Like other mechanosensory organs, they respond extremely rapidly, often within a few milliseconds. We build neural models which reveal that the neural filtering properties such as the stimulus feature and selectivity function of the pilifer mechosensors are strikingly like other insect mechanosensors, including the strain sensors on wings and halteres, abdominal mechanosensors etc. suggesting an important role for sensor mechanics and motion in encoding relevant information.

Grid cells in the medial entorhinal cortex (MEC) fire when an animal is located at the vertices of a hexagonal grid that extends across the environment. The population activity of grid cells serves as an allocentric representation of the current location of the animal. Recent studies have identified a class of grid cells that represent locations ahead of the animal. How do these predictive representations emerge from the wetware of the MEC? To address this question, we developed a detailed conductance-based model of the MEC network, constrained by existing data on the biophysical properties of stellate cells and the topology of the MEC network. Our model revealed two mechanisms underlying the emergence of a predictive code in the MEC. The first relied on a time scale associated with the HCN conductance. The other depended on the degree of asymmetry in the topology of the MEC network. The former mechanism was sufficient to explain predictive coding in layer II grid cells that represented locations shifted ahead of the current location. The shift was equivalent to ~5% of the diameter of a grid field. The latter mechanism was required to model predictive representations in layer III grid cells that were shifted forward by a distance of ~25% of the diameter of a grid field. A corollary of our model, that the extent of the predictive code changes monotonically along the dorsoventral axis of the MEC following observed changes in the properties of the HCN conductance, is borne out by recent experiments.

Oscillatory phenomena are ubiquitous in the brain. Although there are oscillator-based models of brain dynamics, they do not seem to enjoy the universal computational properties of rate-coded and spiking neuron network models. Use of oscillator-based models is often limited to special phenomena like locomotor rhythms and oscillatory attractor-based memories. If neuronal ensembles are taken to be the basic functional units of brain dynamics, it is desirable to develop oscillator-based models that can explain a wide variety of neural phenomena. To this end, we aim to develop a generalized network of oscillatory neurons. Specifically we propose a novel neural network architecture consisting of Hopf oscillators described in the complex domain. The oscillators can adapt their intrinsic frequencies by tracking the frequency components of the input signals. The oscillators are also laterally connected with each other through a special form of coupling we labeled as “power coupling”. Power coupling allows two oscillators with arbitrarily different intrinsic frequencies to interact at a constant normalized phase difference. The network can be operated in two phases. In the encoding phase the oscillators comprising the network perform a Fourier-like decomposition of the input signal(s). In the reconstruction phase, outputs the trained oscillators are combined to reconstruct the training signals. As a salient example, the network can be trained to reconstruct Electroencephalogram (EEG) signals, paving the way to an exciting class of large scale brain models.

Birdsong is a complex motor activity that arises from the interaction between the central and peripheral nervous systems, the body, and the environment. Its striking similarities to human speech, both in production and learning, make songbirds powerful animal models for studying learned motor skills. In this talk, I will present an interdisciplinary approach to understanding the emergence of behavior. Specifically, I will show neuronal recordings from a telencephalic region involved in sensorimotor integration, revealing well-defined oscillations in local field potentials synchronized with the rhythmic structure of canary (Serinus canaria) song. I will introduce a low-dimensional mathematical model of a neural network that replicates the neural dynamics observed in the experiments. This work highlights the value of low-dimensional models as tools for exploring the neuroscience of perception and generation of complex motor behaviors.

Nonlinear dynamics aims to elucidate the basic mechanisms necessary to reflect the temporal behavior of a natural system. The data analysis and modeling techniques proposed by artificial intelligence (deep networks, computational reservoirs, recurrent networks, as examples), on the other hand, ostensibly resign the mechanistic vision for a data-oriented modeling paradigm.  In these lectures, these apparently antagonistic approaches will be analyzed in parallel, using as examples my work on the physics of birdsong production and vocal learning.