Adversarial attacks and defenses using feature-space stochasticity.

Recent studies in deep neural networks have shown that injecting random noise in the input layer of the networks contributes towards ℓp-norm-bounded adversarial perturbations. However, to defend against unrestricted adversarial examples, most of which are not ℓp-norm-bounded in the input layer, such input-layer random noise may not be sufficient. In the first part of this study, we generated a novel class of unrestricted adversarial examples termed feature-space adversarial examples. These examples are far from the original data in the input space but adjacent to the original data in a hidden-layer feature space and far again in the output layer. In the second part of this study, we empirically showed that while injecting random noise in the input layer was unable to defend these feature-space adversarial examples, they were defended by injecting random noise in the hidden layer. These results highlight the novel benefit of stochasticity in higher layers, in that it is useful for defending against these feature-space adversarial examples, a class of unrestricted adversarial examples.

Ketamine-induced 1-Hz oscillation of spontaneous neural activity is not directly visible in the hemodynamics.

The extent to which resting-state hemodynamics reflects the underlying neural activity is still under debate. Especially in the delta frequency band (0.5-4 Hz), it is unclear whether the hemodynamics can directly track the dynamics of underlying neural activity. Based on a recent report showing that ketamine administration induced a 1-Hz neural activity oscillation in the retrosplenial cortex, we conducted simultaneous recordings of the calcium signal and hemodynamics in mice and examined whether the hemodynamics tracked the oscillatory neural activity. Although we observed that the oscillation induced by ketamine appeared in the calcium signal, no sign of oscillation was detected in the simultaneously recorded hemodynamics. Consistently, there was a notable decrease in the correlation between simultaneously recorded calcium signal and hemodynamics. However, on a much longer time scale (10-60 min), we unexpectedly observed an ultraslow increase of hemodynamic signals specifically in the same cortical region exhibiting the neural activity oscillation. These results indicated that hemodynamics cannot track the 1-Hz oscillation in neural activity, although the presence of neural activity oscillation was detectable on a longer timescale. Such ultraslow hemodynamics may be useful for detecting abnormal neural activity induced by psychotic drugs or mental disorders.

Age-associated reduction of nuclear shape dynamics in excitatory neurons of the visual cortex.

Neurons decline in their functionality over time, and age-related neuronal alterations are associated with phenotypes of neurodegenerative diseases. In nonneural tissues, an infolded nuclear shape has been proposed as a hallmark of aged cells and neurons with infolded nuclei have also been reported to be associated with neuronal activity. Here, we performed time-lapse imaging in the visual cortex of Nex-Cre;SUN1-GFP mice. Nuclear infolding was observed within 10 min of stimulation in young nuclei, while the aged nuclei were already infolded pre-stimulation and showed reduced dynamics of the morphology. In young nuclei, the depletion of the stimuli restored the nucleus to a spherical shape and reduced the dynamic behavior, suggesting that nuclear infolding is a reversible process. We also found the aged nucleus to be stiffer than the young one, further relating to the age-associated loss of nuclear shape dynamics. We reveal temporal changes in the nuclear shape upon external stimulation and observe that these morphological dynamics decrease with age.

Thalamocortical circuits for the formation of hierarchical pathways in the mammalian visual cortex.

External sensory inputs propagate from lower-order to higher-order brain areas, and the hierarchical neural network supporting this information flow is a fundamental structure of the mammalian brain. In the visual system, multiple hierarchical pathways process different features of the visual information in parallel. The brain can form this hierarchical structure during development with few individual differences. A complete understanding of this formation mechanism is one of the major goals of neuroscience. For this purpose, it is necessary to clarify the anatomical formation process of connections between individual brain regions and to elucidate the molecular and activity-dependent mechanisms that instruct these connections in each areal pair. Over the years, researchers have unveiled developmental mechanisms of the lower-order pathway from the retina to the primary visual cortex. The anatomical formation of the entire visual network from the retina to the higher visual cortex has recently been clarified, and higher-order thalamic nuclei are gaining attention as key players in this process. In this review, we summarize the network formation process in the mouse visual system, focusing on projections from the thalamic nuclei to the primary and higher visual cortices, which are formed during the early stages of development. Then, we discuss how spontaneous retinal activity that propagates through thalamocortical pathways is essential for the formation of corticocortical connections. Finally, we discuss the possible role of higher-order thalamocortical projections as template structures in the functional maturation of visual pathways that process different visual features in parallel.

Developmental stage-specific spontaneous activity contributes to callosal axon projections.

The developing neocortex exhibits spontaneous network activity with various synchrony levels, which has been implicated in the formation of cortical circuits. We previously reported that the development of callosal axon projections, one of the major long-range axonal projections in the brain, is activity dependent. However, what sort of activity and when activity is indispensable are not known. Here, using a genetic method to manipulate network activity in a stage-specific manner, we demonstrated that network activity contributes to callosal axon projections in the mouse visual cortex during a 'critical period': restoring neuronal activity during that period resumed the projections, whereas restoration after the period failed. Furthermore, in vivo Ca2+ imaging revealed that the projections could be established even without fully restoring highly synchronous activity. Overall, our findings suggest that spontaneous network activity is selectively required during a critical developmental time window for the formation of long-range axonal projections in the cortex.

Modular strategy for development of the hierarchical visual network in mice.

Hierarchical and parallel networks are fundamental structures of the mammalian brain1-8. During development, lower- and higher-order thalamic nuclei and many cortical areas in the visual system form interareal connections and build hierarchical dorsal and ventral streams9-13. One hypothesis for the development of visual network wiring involves a sequential strategy wherein neural connections are sequentially formed alongside hierarchical structures from lower to higher areas14-17. However, this sequential strategy would be inefficient for building the entire visual network comprising numerous interareal connections. We show that neural pathways from the mouse retina to primary visual cortex (V1) or dorsal/ventral higher visual areas (HVAs) through lower- or higher-order thalamic nuclei form as parallel modules before corticocortical connections. Subsequently, corticocortical connections among V1 and HVAs emerge to combine these modules. Retina-derived activity propagating the initial parallel modules is necessary to establish retinotopic inter-module connections. Thus, the visual network develops in a modular manner involving initial establishment of parallel modules and their subsequent concatenation. Findings in this study raise the possibility that parallel modules from higher-order thalamic nuclei to HVAs act as templates for cortical ventral and dorsal streams and suggest that the brain has an efficient strategy for the development of a hierarchical network comprising numerous areas.

Response Selectivity of the Lateral Posterior Nucleus Axons Projecting to the Mouse Primary Visual Cortex.

Neurons in the mouse primary visual cortex (V1) exhibit characteristic response selectivity to visual stimuli, such as orientation, direction and spatial frequency selectivity. Since V1 receives thalamic visual inputs from the lateral geniculate nucleus (LGN) and lateral posterior nucleus (LPN), the response selectivity of the V1 neurons could be influenced mostly by these inputs. However, it remains unclear how these two thalamic inputs contribute to the response selectivity of the V1 neurons. In this study, we examined the orientation, direction and spatial frequency selectivity of the LPN axons projecting to V1 and compared their response selectivity with our previous results of the LGN axons in mice. For this purpose, the genetically encoded calcium indicator, GCaMP6s, was locally expressed in the LPN using the adeno-associated virus (AAV) infection method. Visual stimulations were presented, and axonal imaging was conducted in V1 by two-photon calcium imaging in vivo. We found that LPN axons primarily terminate in layers 1 and 5 and, to a lesser extent, in layers 2/3 and 4 of V1, while LGN axons mainly terminate in layer 4 and, to a lesser extent, in layers 1 and 2/3 of V1. LPN axons send highly orientation- and direction-selective inputs to all the examined layers in V1, whereas LGN axons send highly orientation- and direction-selective inputs to layers 1 and 2/3 but low orientation and direction selective inputs to layer 4 in V1. The distribution of preferred orientation and direction was strongly biased toward specific orientations and directions in LPN axons, while weakly biased to cardinal orientations and directions in LGN axons. In spatial frequency tuning, both the LPN and LGN axons send selective inputs to V1. The distribution of preferred spatial frequency was more diverse in the LPN axons than in the LGN axons. In conclusion, LPN inputs to V1 are functionally different from LGN inputs and may have different roles in the orientation, direction and spatial frequency tuning of the V1 neurons.

Chromatic micromaps in primary visual cortex.

The clustering of neurons with similar response properties is a conspicuous feature of neocortex. In primary visual cortex (V1), maps of several properties like orientation preference are well described, but the functional architecture of color, central to visual perception in trichromatic primates, is not. Here we used two-photon calcium imaging in macaques to examine the fine structure of chromatic representation and found that neurons responsive to spatially uniform, chromatic stimuli form unambiguous clusters that coincide with blobs. Further, these responsive groups have marked substructure, segregating into smaller ensembles or micromaps with distinct chromatic signatures that appear columnar in upper layer 2/3. Spatially structured chromatic stimuli revealed maps built on the same micromap framework but with larger subdomains that go well beyond blobs. We conclude that V1 has an architecture for color representation that switches between blobs and a combined blob/interblob system based on the spatial content of the visual scene.

Long-Range Interhemispheric Projection Neurons Show Biased Response Properties and Fine-Scale Local Subnetworks in Mouse Visual Cortex.

Integration of information processed separately in distributed brain regions is essential for brain functions. This integration is enabled by long-range projection neurons, and further, concerted interactions between long-range projections and local microcircuits are crucial. It is not well known, however, how this interaction is implemented in cortical circuits. Here, to decipher this logic, using callosal projection neurons (CPNs) in layer 2/3 of the mouse visual cortex as a model of long-range projections, we found that CPNs exhibited distinct response properties and fine-scale local connectivity patterns. In vivo 2-photon calcium imaging revealed that CPNs showed a higher ipsilateral (to their somata) eye preference, and that CPN pairs showed stronger signal/noise correlation than random pairs. Slice recordings showed CPNs were preferentially connected to CPNs, demonstrating the existence of projection target-dependent fine-scale subnetworks. Collectively, our results suggest that long-range projection target predicts response properties and local connectivity of cortical projection neurons.

Natural images are reliably represented by sparse and variable populations of neurons in visual cortex.

Natural scenes sparsely activate neurons in the primary visual cortex (V1). However, how sparsely active neurons reliably represent complex natural images and how the information is optimally decoded from these representations have not been revealed. Using two-photon calcium imaging, we recorded visual responses to natural images from several hundred V1 neurons and reconstructed the images from neural activity in anesthetized and awake mice. A single natural image is linearly decodable from a surprisingly small number of highly responsive neurons, and the remaining neurons even degrade the decoding. Furthermore, these neurons reliably represent the image across trials, regardless of trial-to-trial response variability. Based on our results, diverse, partially overlapping receptive fields ensure sparse and reliable representation. We suggest that information is reliably represented while the corresponding neuronal patterns change across trials and collecting only the activity of highly responsive neurons is an optimal decoding strategy for the downstream neurons.

Characterisation of nonlinear receptive fields of visual neurons by convolutional neural network.

A comprehensive understanding of the stimulus-response properties of individual neurons is necessary to crack the neural code of sensory cortices. However, a barrier to achieving this goal is the difficulty of analysing the nonlinearity of neuronal responses. Here, by incorporating convolutional neural network (CNN) for encoding models of neurons in the visual cortex, we developed a new method of nonlinear response characterisation, especially nonlinear estimation of receptive fields (RFs), without assumptions regarding the type of nonlinearity. Briefly, after training CNN to predict the visual responses to natural images, we synthesised the RF image such that the image would predictively evoke a maximum response. We first demonstrated the proof-of-principle using a dataset of simulated cells with various types of nonlinearity. We could visualise RFs with various types of nonlinearity, such as shift-invariant RFs or rotation-invariant RFs, suggesting that the method may be applicable to neurons with complex nonlinearities in higher visual areas. Next, we applied the method to a dataset of neurons in mouse V1. We could visualise simple-cell-like or complex-cell-like (shift-invariant) RFs and quantify the degree of shift-invariance. These results suggest that CNN encoding model is useful in nonlinear response analyses of visual neurons and potentially of any sensory neurons.

Cell-Type-Specific Thalamocortical Inputs Constrain Direction Map Formation in Visual Cortex.

Finding the relationship between individual cognitive functions and cell-type-specific neuronal circuits is a central topic in neuroscience. In cats, the lateral geniculate nucleus (LGN) contains several cell types carrying spatially and temporally precise visual information. Whereas LGN cell types lack selectivity for motion direction, neurons in the primary visual cortex (area 17) exhibit sharp direction selectivity. Whether and how such de novo formation of direction selectivity depends on LGN cell types remains unknown. Here, we addressed this question using in vivo two-photon calcium imaging in cat area 17, which consists of two compartments receiving different combinations of inputs from the LGN cell types. The direction map in area 17 showed unique fragmented organization and was present only in small and distributed cortical domains. Moreover, direction-selective domains preferentially localized in specific compartments receiving Y and W inputs carrying low spatial frequency visual information, indicating that cell-type-specific thalamocortical projections constrain the formation of direction selectivity.

Neuronal Origin of the Temporal Dynamics of Spontaneous BOLD Activity Correlation.

Resting-state functional connectivity (FC) has become a major functional magnetic resonance imaging method to study network organization of human brains. There has been recent interest in the temporal fluctuations of FC calculated using short time windows ("dynamic FC") because this method could provide information inaccessible with conventional "static" FC, which is typically calculated using the entire scan lasting several tens of minutes. Although multiple studies have revealed considerable temporal fluctuations in FC, it is still unclear whether the fluctuations of FC measured in hemodynamics reflect the dynamics of underlying neural activity. We addressed this question using simultaneous imaging of neuronal calcium and hemodynamic signals in mice and found coordinated temporal dynamics of calcium FC and hemodynamic FC measured in the same short time windows. Moreover, we found that variation in transient neuronal coactivation patterns was significantly related to temporal fluctuations of sliding window FC in hemodynamics. Finally, we show that the observed dynamics of FC cannot be fully accounted for by simulated data assuming stationary FC. These results provide evidence for the neuronal origin of dynamic FC and further suggest that information relevant to FC is condensed in temporally sparse events that can be extracted using a small number of time points.

Mouse optical imaging for understanding resting-state functional connectivity in human fMRI.

Resting-state functional connectivity (FC), which measures the temporal correlation of spontaneous hemodynamic activity between distant brain areas, is a widely accepted method in functional magnetic resonance imaging (fMRI) to assess the connectome of healthy and diseased human brains. A common assumption underlying FC is that it reflects the temporal structure of large-scale neuronal activity that is converted into large-scale hemodynamic activity. However, direct observation of such relationship has been difficult. In this commentary, we describe our recent progress regarding this topic. Recently, transgenic mice that express a genetically encoded calcium indicator (GCaMP) in neocortical neurons are enabling the optical recording of neuronal activity in large-scale with high spatiotemporal resolution. Using these mice, we devised a method to simultaneously monitor neuronal and hemodynamic activity and addressed some key issues related to the neuronal basis of FC. We propose that many important questions about human resting-state fMRI can be answered using GCaMP expressing transgenic mice as a model system.

Astrocytes in the mouse visual cortex reliably respond to visual stimulation.

Astrocytes are known to contact with a great number of synapses and may integrate sensory inputs. In the ferret primary visual cortex, astrocytes respond to a visual stimulus with a delay of several seconds with respect to the surrounding neurons. However, in the mouse visual cortex, it remains unclear whether astrocytes respond to visual stimulations. In this study, using dual-color simultaneous in vivo two-photon calcium imaging of neurons and astrocytes in the awake mouse visual cortex, we examined the visual response of astrocytes and their precise response timing relative to the surrounding neurons. Neurons reliably responded to visual stimulations, whereas astrocytes often showed neuromodulator-mediated global activities, which largely masked small visual responses. Administration of the selective α1-adrenergic receptor antagonist prazosin substantially reduced such global astrocytic activities without affecting the neuronal visual responses. In the presence of prazosin, astrocytes showed weak but consistent visual responses mostly at their somata. Cross-correlation analysis estimated that the astrocytic visual responses were delayed by approximately 5 seconds relative to the surrounding neuronal responses. In conclusion, our research demonstrated that astrocytes in the primary visual cortex of awake mice responded to visual stimuli with a delay of several seconds relative to the surrounding neurons, which may indicate the existence of a common mechanism of neuron-astrocyte communication across species.

Cell Type Specific Representation of Vibro-tactile Stimuli in the Mouse Primary Somatosensory Cortex.

Although the processing of whisker deflections in the barrel area of the rodent primary somatosensory cortex (S1) has been studied extensively, how cutaneous vibro-tactile stimuli are processed in the rodent S1 outside the barrel area has not been fully examined. Particularly, the cell-type specific representation of multiple vibration frequencies in genetically identified inhibitory cells in the S1 has not been examined. Using two-photon calcium imaging, we examined the responses to vibration stimuli of excitatory and inhibitory neurons in the S1 hind limb area of male and female mice. The excitatory cells showed relatively sharp selectivity to vibration stimuli, whereas the inhibitory cells exhibited less selectivity. The excitatory and inhibitory cells with different preferred stimuli were intermingled in a "salt and pepper" manner. Furthermore, the noise correlation tended to be especially strong in excitatory-inhibitory and inhibitory-inhibitory cell pairs that have similar stimulus selectivity. These results suggest that excitatory cells tend to represent specific stimulus information and work together with similarly tuned inhibitory cells as a functionally connected network.

Functional Segregation and Development of Mouse Higher Visual Areas.

Recent studies suggest that higher visual areas (HVAs) in the mouse visual cortex are segregated anatomically into two visual streams, likely analogous to the ventral and dorsal streams in primates. However, HVAs in mice have yet to be characterized functionally. Moreover, it is unknown when the functional segregation of HVAs occurs during development. Here, we investigated spatiotemporal selectivity of HVAs and their development using wide-field calcium imaging. We found that lateral HVAs in the anatomical ventral stream shared similar spatiotemporal selectivity, whereas the spatiotemporal selectivity of anterior and medial HVAs in the anatomical dorsal stream was not uniform and these areas were segregated functionally into multiple groups. This functional segregation of HVAs developed and reached an adult-like pattern ∼10 d after eye opening (EO). These results suggest, not only the functional segregation of ventral and dorsal streams, but also the presence of multiple substreams in the dorsal stream, and indicate that the functional segregation of visual streams occurs gradually after EO.SIGNIFICANCE STATEMENT Investigation of the spatiotemporal selectivity of nine higher visual areas (HVAs) in adult and developing mice revealed that lateral HVAs belonging to the putative ventral stream shared similar spatiotemporal selectivity, whereas the spatiotemporal selectivity of anterior and medial HVAs belonging to the putative dorsal stream was not uniform and these areas were segregated functionally into multiple groups. These results suggest the presence of multiple substreams within the putative dorsal stream for visuospatial processing. Furthermore, we found that initially immature functional segregation among HVAs developed to an adult-like pattern ∼10 d after eye opening. These results provide a foundation for using mouse HVAs as a model to understand parallel processing and its developmental mechanism.

Color Representation Is Retinotopically Biased but Locally Intermingled in Mouse V1.

Dichromatic vision is common in many mammals. However, color processing in the primary visual cortex (V1) of dichromatic mammals is relatively unknown compared to the trichromatic primates. In this study, we investigated the functional organization of color processing in mouse V1. The mouse retina has a graded expression pattern of two opsins along its dorsoventral axis. However, it is not clear whether and how this expression pattern is reflected in the cortical representation at local (several hundred microns) and areal (V1) level. Using in vivo two-photon calcium (Ca2+) imaging and wide-field Ca2+ imaging, we revealed that V1 neurons responded to S (UV)- and M (green)-opsin isolating stimuli with slightly biased color preference depending on retinotopic position in V1. This was consistent with the distribution of retinal opsins. At the cellular level, preferences for S- and M-opsin isolating stimuli were intermingled in a local region encompassing several hundred microns. These results suggest that functional organizations of color information are locally intermingled, but slightly biased depending on the retinotopic position in mouse V1.

Mixed functional microarchitectures for orientation selectivity in the mouse primary visual cortex.

A minicolumn is the smallest anatomical module in the cortical architecture, but it is still in debate whether it serves as functional units for cortical processing. In the rodent primary visual cortex (V1), neurons with different preferred orientations are mixed horizontally in a salt and pepper manner, but vertical functional organization was not examined. In this study, we found that neurons with similar orientation preference are weakly but significantly clustered vertically in a short length and horizontally in the scale of a minicolumn. Interestingly, the vertical clustering is found only in a part of minicolumns, and others are composed of neurons with a variety of orientation preferences. Thus, the mouse V1 is a mixture of vertical clusters of neurons with various degrees of orientation similarity, which may be the compromise between the brain size and keeping the vertical clusters of similarly tuned neurons at least in a subset of clusters.

Transient neuronal coactivations embedded in globally propagating waves underlie resting-state functional connectivity.

Resting-state functional connectivity (FC), which measures the correlation of spontaneous hemodynamic signals (HemoS) between brain areas, is widely used to study brain networks noninvasively. It is commonly assumed that spatial patterns of HemoS-based FC (Hemo-FC) reflect large-scale dynamics of underlying neuronal activity. To date, studies of spontaneous neuronal activity cataloged heterogeneous types of events ranging from waves of activity spanning the entire neocortex to flash-like activations of a set of anatomically connected cortical areas. However, it remains unclear how these various types of large-scale dynamics are interrelated. More importantly, whether each type of large-scale dynamics contributes to Hemo-FC has not been explored. Here, we addressed these questions by simultaneously monitoring neuronal calcium signals (CaS) and HemoS in the entire neocortex of mice at high spatiotemporal resolution. We found a significant relationship between two seemingly different types of large-scale spontaneous neuronal activity-namely, global waves propagating across the neocortex and transient coactivations among cortical areas sharing high FC. Different sets of cortical areas, sharing high FC within each set, were coactivated at different timings of the propagating global waves, suggesting that spatial information of cortical network characterized by FC was embedded in the phase of the global waves. Furthermore, we confirmed that such transient coactivations in CaS were indeed converted into spatially similar coactivations in HemoS and were necessary to sustain the spatial structure of Hemo-FC. These results explain how global waves of spontaneous neuronal activity propagating across large-scale cortical network contribute to Hemo-FC in the resting state.