Hence, the mixtures need to be estimated iteratively using Expectation-Maximization (E-M). The optimization problem for obtaining the separating hyperplane is given by: A massive cluster in the superior temporal cortex reflected perfect parallel processing, firmly constraining the extension of the cerebral locus of the bottleneck. The occipitotemporal (ventral, or “what”) pathway begins in the striate cortex (V1) and projects to the angular gyrus for language processing, to the inferior temporal lobe for object identification, and to the limbic structures. First, a random, unbiased six-faced die was considered. It is a significantly under-constrained problem to arrive at the boundaries without any explicit retinotopic information. In parallel, we assessed paired-pulse suppression in primary somatosensory and visual cortex. PCA is a general technique for reducing dimensionality. A) Boundaries derived using different visual stimuli in one mouse. This dataset has individual neuronal responses recorded from mice of different transgenic Cre-lines. Secondly, we considered a six-faced die biased by the proportion of different area sizes (or the number of pixels/neurons used during training). The plots on the left (A, C) are obtained from natural movie responses and that on the right (B, D) are obtained from resting state responses. This step has been demonstrated in. Felleman, in Encyclopedia of Neuroscience, 2009. The values within the bracket denote the classification accuracy. Historically, the functions of multiple cortical visual areas have been studied in non-human primates, which have well-defined areal parcellations based on visual field representations [1–4]. This is because a high correlation was observed among the nearby pixels in wide-field data. This work uses the Cre-lines “Emx1-IRES” (the whole cortex), and “Nr5a1” (layer 4 specific), which has recordings from all the six visual areas. In each session, different stimuli were used. Learn vocabulary, terms, and more with flashcards, games, and other study tools. We note that the pixels in the wide-field dataset represent the pooled response of neurons, whereas the two-photon dataset captures individual neuron responses. https://doi.org/10.1371/journal.pcbi.1008548.s008. Similarly, for the two-photon dataset, non-averaged 240 secs of resting state responses were used after excluding the initial and final 30 secs of total 5 mins recording from “Session A” of the dataset. The motivation for using GMM is that it uses a multimodal probability density function to represent each visual area. The corresponding classification accuracies are reported in the second row of Fig 3D. For example, in the case of speaker verification biometric, UBM is a speaker-independent GMM that is trained using speech samples obtained from a large set of speakers. Contributed equally to this work with: Visual cortex of higher mammals can be segmented into different functional visual areas. (11) This random classifier will give a chance level accuracy of 16.67%, irrespective of the dataset. Here, we collect responses of the visual cortex to various types of stimuli and ask if we could discover unique clusters from this … However, our critical observation is that data-driven models can classify these areas accurately based on responses to a range of visual stimuli, which extends as well to responses in the resting state. A single SVM can solve only a binary class problem, a one-against-one approach was used to model the multi-class problem, and the final class label was determined using voting strategy. Yes For classifying the test data, the Bayes classifier given in Eq 9 was used. Recently, functional imaging techniques have been used to define the topographic organization of visual areas, especially in humans. It lies at the rear of the brain (highlighted in the image), above the cerebellum . The duration of these three movies were 30, 30, and 120 secs, in order. Similarly, Fig 8E and 8F shows the same result for data from the Emx1-IRES subset of the two-photon dataset. Yes If the features of pixels within each area are similar, then it should be possible to classify them using the trained models. https://doi.org/10.1371/journal.pcbi.1008548.s009. This approach has been most useful for the identification of area MT in primates, whose neurons are nearly all selective for the direction of object motion. Yes Let oi = f(xi) be a non-linear function modeled by ANN, where xi is the input vector representing the neuronal response and f(.) Visual cortex in the macaque comprises about 60 per cent of neocortex. The details of the optimization algorithm can be found in [35]. Precise visual area boundaries can be identified based on the sign of visual field representations [1, 9, 13–18], based on the principle that adjacent visual areas share a common representation of either the vertical or horizontal meridian and have essentially mirror-imaged maps across the common border. In all experiments done with visual stimuli, the responses were averaged across trials to obtain stimulus-specific responses. Study 6.3.1 Parallel Processing In The Visual Cortex 2 flashcards from Jason McBride's class online, or in Brainscape's iPhone or Android app. Here, we collect responses of the visual cortex to various types of stimuli and ask if we could discover unique clusters from this dataset using machine learning methods. Project administration, This parallel processing of different aspects of sensation may be fundamental to the operation of the visual as well as the somatosensory systems, and may prove to be a feature o fall sensory pathways. Conceptualization, Previous work has segmented the mouse visual cortex into different areas based on the organization of retinotopic maps. The proposed supervised approach is able to predict the area labels accurately using neuronal responses to various visual stimuli. D) Pixels selected for training the supervised model are limited to center x% of the radius of the visual area. Traditionally, cortical areas have been defined using multiple, convergent criteria, including architectonics, cortical connections, topographic organization, functional properties, and behavioral contributions. This dataset is collected using two-photon microscopy and consists of individual neuronal responses recorded from the six visual areas. Similar to PCA, LDA is also used to reduce the dimension of population neuronal responses. One third of the cortical area of the human brain is dedicated to visual information processing. Information processing in the cerebral cortex involves interactions among distributed areas. In the semi-supervised approach, the center-most pixels were the only supervised information given to the algorithm. (9) In addition to this, the results in Tables 3 and 4 and Fig 4, show that the classification accuracy and confusion matrix are poor for the two-photon dataset when compared to the wide-field dataset. Sight questions. Ming Hu, Roles This result provides additional support for our conjecture that there are intrinsic response characteristics of each visual area that generalize across stimuli and enable classification. Following processing in this region, the visual neuronal impulses are directed to secondary visual cortex or V2. The experiments for collecting the wide-field dataset (Section 1.2.1) were carried out under protocols approved by MIT’s Animal Care and Use Committee (Protocol Approval Number: 1020-099-23) and conform to NIH guidelines. Step 4: The center cluster/grid of each visual area is labeled using the retinotopic map. The region that receives information directly from the LGN is called the primary visual cortex , (also called V1 and striate cortex). https://doi.org/10.1371/journal.pcbi.1008548.t005. (7) (8) Here we show the confusion matrices for all mice and Cre-lines from the wide-field and two-photon datasets, respectively. Unlike the wide-field imaging dataset, this dataset has individual neuronal responses recorded from different areas and mice. Since the training data for the supervised model are sampled randomly from each visual area, the classification accuracy observed in Fig 3A and 3B can be an artifact of correlated responses. The details of the semi-supervised clustering approach used in this paper are given below: The result of clustering the visual areas using the semi-supervised approach is given in Table 5. After recovery from surgery, mice were acclimatized to head fixation and then imaged while awake. Evidence from anatomical tracer studies as well as lesions of the primary auditory cortex (AI) indicate that the principal relay nucleus of the auditory thalamus, the ventral part of the medial geniculate (MGv), projects in parallel to AI and the rostral area on the supratemporal plane of the macaqu … In this section, we compare resting state responses with single-trial and trial-averaged stimulus-induced responses of various durations. The LDA space significantly improved the ratio of average intra-area and inter-area correlations to 8.8 and 9.7 for the wide-field and two-photon datasets, respectively. There is higher variability in the neuronal responses, which leads to poorer performance. Visualization, Single cells interconnect according to precise interlaminar patterns to form columns, whereas columns interact not only locally, within the same hypercolumn, but also with distant columns representing other parts of the visual field. Adaptive thresholding for every iteration eliminates the need for the penalty term given in [42]. This section briefly describes two datasets collected using wide-field and two-photon imaging, respectively. In Sections 2.1 and 2.2, the natural movies and other stimuli were presented multiple times and averaged to obtain the stimulus-induced response. These raw signals were preprocessed using PCA and LDA before they were given to the classifiers. The results obtained in Tables 3 and 4 are not based on a single classifier. Parallel processing of visual space by neighboring neurons in mouse visual cortex. Utilizing this ground truth information, data-driven models are built using supervised and semi-supervised approaches to identify the visual area boundaries. A simple single hidden-layer ANN with 30 nodes was chosen to classify the neuronal response. Copyright: © 2021 Kumar et al. The visual cortex of the brain is the area of the cerebral cortex that processes visual information.It is located in the occipital lobe.Sensory input originating from the eyes travels through the lateral geniculate nucleus in the thalamus and then reaches the visual cortex. The wide-field dataset was collected by the authors on awake, head-fixed mice, which transgenically express GCaMP6f or GCaMP6s. Yes The entries denote “% accuracy (± standard deviation)”. The problem of fitting a GMM is an incomplete data problem. (5). Imaging and visual stimulation. https://doi.org/10.1371/journal.pcbi.1008548.g001. D) An intermediate step in the clustering process. For more information about PLOS Subject Areas, click https://doi.org/10.1371/journal.pcbi.1008548.g006. wk, denote the respective component weight. The results suggest that responses from visual cortical areas can be classified effectively using data-driven models. (10) ¶‡These authors are joint senior authors on this work. Funding: Supported by National Institutes of Health (NIH) grants EY007023 and EY028219 (MS), and the Center for Computational Brain Research (CCBR), IIT Madras, N.R. Here we present the inter and intra-area correlations for various individual mice from the dataset, which had responses recorded from three or more sessions. In between two consecutive stimuli, a grey and blank screen was shown for 2 secs to capture the baseline responses. Bayes classifier in this setup reduces to a maximum likelihood classifier (Eq 9). In Fig 8, average intra-area and inter-area correlations computed from input responses were shown for an example mouse and a Cre-line from the wide-field and two-photon datasets, respectively. An animal running must go left or right around a tree; it cannot do both. The procedure for training these parameters using E-M is detailed in [27]. The findings suggest that the activity patterns of different visual areas can be used to reliably and accurately classify their borders. The two-photon dataset captures responses of only few individual neurons from each area (Table 2). The visual cortex was initially divided into small clusters of equal size such that they each had adequate data for MAP adaptation. Gaussian mixture models (GMMs) are used to model visual areas in this method. Following up on these results using visual stimuli, we hypothesized that each area of the mouse brain has unique responses that can be used to classify the area independently of stimuli. Each area has a distinct representation of the visual field and presumably a unique contribution to visual information processing. The synaptic and circuit mechanisms underlying such plasticity are being progressively understood. D.J. where x is a D-dimensional random vector that describes the neuronal response, , are the unimodal component densities with mean and covariance Σk. 5.1A) and orientation selectivity (Hubel and Wiesel, 1969). For both the datasets, this led to accuracy of ≈ 14% to 18% (similar to the random chance level of 16% for 6 classes). We have shown the proposed supervised methods to work with generative (GMM, Unimodal Bayes) and discriminative (SVM, ANN) classifiers. For every iteration, a threshold of the score is determined, and the clusters are merged. Anatomical connectivity suggests that certain areas form local hierarchical relations such as within the visual system. The boundaries obtained for different mice and stimuli are shown in Fig 6. With the wide-field dataset, clustering neuronal responses using a constrained semi-supervised classifier showed graceful degradation of accuracy. The results of supervised classifiers suggest that the activity of each area has a specific statistical characteristic which can be used identify the area. The most prominent of these architectonic studies were those performed by Brodmann (1909) and von Economo and Koskinas (1925) who subdivided the occipital cortex into three distinct zones. Azimuth 0° and 90° correspond to the midline and the far periphery of the contralateral visual field, respectively. For the wide-field dataset, the non-averaged 800 secs of resting state responses excluding initial and final 50 secs of 15 mins recording sessions (to exclude non-stationary transients) were used in this analysis. The advantage of using a feed-forward neural network over a conventional GMM is that the ANN is trained in a discriminative way. However, multiple reentrant pathways, both within and between areas, allow interaction between these different modular systems. This shows that the obtained results are mainly because of the proposed dimension reduction using PCA and LDA rather than the classifier (Fig 9). The process can take a mere 13 milliseconds, according to a 2017 study at MIT in the United States. Learn faster with spaced repetition. This experiment shows that the observed results are not an artifact of the supervised approach. Mriganka Sur, ... Hiroki Sugihara, in Progress in Brain Research, 2013. 6. https://doi.org/10.1371/journal.pcbi.1008548.s003, https://doi.org/10.1371/journal.pcbi.1008548.s004. These cluster-specific GMMs are probability density functions that represent the responses of each cluster with UBM as the reference. Although the results are inferior compared to the supervised classification, the boundaries were observed to be close to the ground truth retinotopic boundaries in Fig 6. In Fig 4, we show examples of confusion matrices, obtained using mouse M1 from wide-field dataset and Emx1-IRES Cre-line from two-photon dataset, respectively. https://doi.org/10.1371/journal.pcbi.1008548.g004. C) Results on resting state responses for two mice. The hypothesis, given its initial characterisation in a paper by David Milner and Melvyn A. Goodale in 1992, argues that humans possess two distinct visual systems. Further examples of correlations analysis are shown in S5–S7 Figs. PCA is used extensively in diverse fields from neuroscience to physics because it is a simple, non-parametric method of obtaining relevant information from complex datasets. The clusters formed using the resting state responses were also consistent with the retinotopic maps (Fig 6D). Digital image processing, as a computer-based technology, carries out automatic processing, where w is a normal vector to the separating hyperplane, and b is the bias of the same. In this method, visual areas are modeled using a parametric unimodal distribution. A) Diagrammatic representation of a mouse prepared for wide-field imaging. In A, the results are compared across different visual stimuli. (15) Likewise, the somatosensory thalamus of Spalax almost reached the dorsolateral surface, without evidence of lateral geniculate body, the first station in the visual pathway. Felleman, in The Senses: A Comprehensive Reference, 2008. of concurrent processing streams in the visual cortex. The population responses are first projected to a lower dimension space using techniques such as principal component analysis (PCA) and linear discriminant analysis (LDA). A) Confusion matrix obtained using responses of Mouse M1 and Natural Movie stimuli. https://doi.org/10.1371/journal.pcbi.1008548.g008. Even with this small amount of supervised information, the algorithm was able to find boundaries that were consistent with that of retinotopically obtained boundaries. In the current view of visual cortical organization in macaque monkeys, 31 subdivisions have been described consistently by two or more of these criteria and thus are recognized as cortical areas, whereas 20 subdivisions are less consistently described and retain the designation of cortical zone. Each of these levels of organization performs certain neuronal operations under the influence of many other neural networks. D) Boundaries derived from resting state responses. Since this dataset was collected using single-photon wide-field imaging, the spatial resolution is limited to the pixels of the microscopic image. The visual cortex is the largest system in the human brain and is responsible for processing the visual image. In this paper, the responses of six visual areas of mouse cortex to various stimuli were studied using data-driven methods. We tested this hypothesis by using resting state recordings from both datasets. Software, They suggested that these streams are associated with different capabilities: (1) the parietal stream is involved in visual assessment of spatial relationships; (2) the temporal stream is concerned with visual recognition of objects. Visual processing: Parallel-er and Parallel-er Richard T. Born The mammalian visual system processes many different aspects of the visual scene in separate, parallel channels. This indicates that the detection of a target defined by audio-visual conjunction is achieved via the same mechanism as within a single perceptual modality, through separate, parallel processing of the auditory and visual features and serial processing of the feature conjunction elements, rather than by evaluation of a fused multimodal percept. Single cells interconnect according to precise interlaminar patterns to form columns, whereas columns interact not only locally, within the same hypercolumn, but also with distant columns representing other parts of the visual field. The principal component corresponding to the largest eigenvalue captures the maximum variance present in the dataset. Structure can be represented at multiple levels, including transitional probabilities, ordinal position, and identity of units. In Fig 8C, 8D, 8G, and 8H, the correlations computed in the LDA domain are shown. In both datasets, the mouse visual cortex was first partitioned into different visual areas using a retinotopic map [9]. The area borders obtained by the classification are close to the retinotopic boundaries for all visual stimuli and all mice used. With this dataset, we get an accuracy significantly better than random chance. (first row of Fig 3D). along the processing chain show more complex selectivity. These results were verified to be consistent by training and testing responses of different mice to natural movie stimuli (Fig 3B). Our findings show that resting state responses are an important complement to visual responses in classifying areas. Moreover, each area can contribute to different types of analysis; for example, the information about direction of motion analyzed in MT can be used to detect the movement of an object, a person's self-motion, or the shape of a camouflaged vehicle moving against a background. McGill University, CANADA, Received: January 31, 2020; Accepted: November 17, 2020; Published: February 4, 2021. This result suggests that stimulus-driven responses contain better-discriminating responses compared to responses without an overt stimulus. E. Zavitz, ... N.S.C. These techniques rely on retinotopic maps within representations of the contralateral visual field to derive visual areas. Since this dataset consists of individual neuronal responses, the numbers of neurons from each area varied from session to session. For each visual area, the parameters are determined using maximum likelihood estimation (MLE). In the secondary visual cortex (BA 18) neuronal representations of color, form, and movement remain segregated, but in a way different from the one in the primary visual cortex12. The scalability of the proposed approach on the dataset from the Allen Institute and the experiments on the resting state analysis (with no overt stimuli) adds to this hypothesis. The LDA projection matrix is computed using scatter matrix Σb and covariance matrix Σ as given in Eq 5. PCA is a statistical dimensionality reduction technique which can be used to find directions of maximum variability. This result is consistent with the presence of unique area-specific circuitry in the visual cortex, which shapes visually driven or resting state activity in these areas. Since this is a multi-class classification problem, softmax non-linear function is used as the activation function for the output layer.