Biology subjects

Khosla, M.

Publications and source records attributed to Khosla, M..

12 recordsLinked to original sources

bioRxiv2026

Apparent food selectivity reflects multiple non-food image properties

Three recent publications have reported selective fMRI responses to images of food in the human ventral visual pathway. However, all three studies were based primarily on the Natural Scenes Dataset (NSD), in which high-level categories like food are correlated with other image properties such as the color and size of objects, and the distance of the scene. To test whether the reported food selectivity might reflect these or other correlates of food images rather than (or in addition to) food, we constructed novel stimuli that manipulated these properties on both food and non-food images, and collected data from new subjects in a data-rich design across two experiments pre-registered in OSF. We used a localizer paradigm based on a subset of NSD images to infer the "food component" in new subjects and then measured responses of this component to our new stimuli. In Experiment 1, we found that the food component response magnitude i) had a higher response to color than to greyscale images but showed no interaction between food and color, ii) showed a reduced preference for food over non-food when both were within reaching distance, and most importantly iii) was no higher for food than non-food when both were visually matched and presented as Cutouts on a white background. In Experiment 2, we found that the response of the food component could not be explained by object distance, real-world size, or mid-level visual image statistics. However, the response to non-food images with "gooey" material properties was as high as the response to food. Across both experiments, we consistently found a very low response (at or below fixation baseline) to NSD non-food images, which were predominantly outdoor scenes. Together, our results prompt a revision of prior claims including our own, indicating that the previously reported food component is better characterized as food-biased rather than strictly food-selective, and is driven in part by contextual or material features that are also present in non-food images. Our findings further highlight the importance of supplementing studies based on naturalistic images with experiments that unconfound image properties with carefully designed stimuli.

neuroscience↗

bioRxiv2024

Conclusions about Neural Network to Brain Alignment are Profoundly Impacted by the Similarity Measure

Deep neural networks are widely used to model biological perception and behavior, making the similarities and differences between artificial and biological systems consequential. If a principle (e.g. self-supervised learning) produces a model resembling biology, this is taken as evidence the same principle shapes the biological system. But what it means for an artificial system to be similar to a biological one is complex. A popular approach compares representations of identical stimuli using a similarity measure like Representational Similarity Analysis. Yet scientific questions rarely specify which measure is appropriate, raising a key question: do conclusions depend on this choice? Focusing on vision, we show that measure choice influences both hierarchical correspondence between systems and the ranking of which artificial systems are most biological. Reanalyzing prior studies, we find the choice hugely consequential: models best under one measure are often worst under another, and prior conclusions can flip or dissolve. Different metrics capture fundamentally different aspects of similarity, warranting healthy skepticism toward such comparisons.

neuroscience↗

bioRxiv2024

Privileged representational axes in biological and artificial neural networks

How do neurons code information? Recent work emphasizes properties of population codes, such as their geometry and decodable information, using measures that are blind to the native tunings (or axes) of neural responses. But might these representational axes matter, with some privileged systematically over others? To find out, we developed methods to test for alignment of neural tuning across brains and deep convolutional neural networks (DCNNs). Across both vision and audition, both brains and DCNNs consistently favored certain axes for representing the natural world. Moreover, the representational axes of DCNNs trained on natural inputs were aligned to those in perceptual cortices, such that axis-sensitive model-brain similarity metrics better differentiated competing models of biological sensory systems. We further show that coding schemes that privilege certain axes can reduce downstream wiring costs and improve generalization. These results motivate a new framework for understanding neural tuning in biological and artificial networks and its computational benefits.

neuroscience↗

bioRxiv2022

A Highly Selective Response to Food in Human Visual Cortex Revealed by Hypothesis-Free Voxel Decomposition

Prior work has identified cortical regions selectively responsive to specific categories of visual stimuli. However, this hypothesis-driven work cannot reveal how prominent these category selectivities are in the overall functional organization of visual cortex, or what others might exist that scientists have not thought to look for. Further, standard voxel-wise tests cannot detect distinct neural selectivities that coexist within voxels. To overcome these limitations, we used data-driven voxel decomposition methods to identify the main components underlying fMRI responses to thousands of complex photographic images (Allen et al 2021). Our hypothesis-neutral analysis rediscovered components selective for faces, places, bodies, and words, validating our method and showing that these selectivities are dominant features of the ventral visual pathway. The analysis also revealed an unexpected component with a distinct anatomical distribution that responded highly selectively to images of food. Alternative accounts based on low to mid-level visual features like color, shape or texture failed to account for the food selectivity of this component. High-throughput testing and control experiments with matched stimuli on a highly accurate computational model of this component confirm its selectivity for food. We registered our methods and hypotheses before replicating them on held-out participants and in a novel dataset. These findings demonstrate the power of data-driven methods, and show that the dominant neural responses of the ventral visual pathway include not only selectivities for faces, scenes, bodies, and words, but also the visually heterogeneous category of food, thus constraining accounts of when and why functional specialization arises in the cortex.

neuroscience↗

bioRxiv2022

High-level visual areas act like domain-general filters with strong selectivity and functional specialization

Neuroscientific studies rely heavily on a-priori hypotheses, which can bias results toward existing theories. Here, we use a hypothesis-neutral approach to study category selectivity in higher visual cortex. Using only stimulus images and their associated fMRI activity, we constrain randomly initialized neural networks to predict voxel activity. Despite no category-level supervision, the units in the trained networks act as detectors for semantic concepts like faces or words, providing solid empirical support for categorical selectivity. Importantly, this selectivity is maintained when training the networks without images that contain the preferred category, strongly suggesting that selectivity is not domain-specific machinery, but sensitivity to generic patterns that characterize preferred categories. The ability of the models representations to transfer to perceptual tasks further reveals the functional role of their selective responses. Finally, our models show selectivity only for a limited number of categories, all previously identified, suggesting that the essential categories are already known. TeaserModels trained solely to predict fMRI activity from images reveal strong category selectivity in higher visual areas, even without exposure to these categories in training.

neuroscience↗

bioRxiv2022

Visual Perception of 3D Space and Shape in Time - Part I: 2D Space Perception by 2D Linear Translation

Visual perception plays a critical role in navigating space and extracting useful semantic information crucial to survival. To identify distant landmarks, we constantly shift gaze vectors through saccades, while still maintaining the visual perception of stable allocentric space. How can we sustain stable allocentric space so effortlessly? To solve this question, we have developed a new concept of NHT (Neural Holography Tomography). This model states that retinotopy is invisible (not available to consciousness) and must be converted to a time code by traveling alpha brainwaves to perceive objects consciously. According to this framework, if identical alpha phases are continually assigned to a landmark, we perceive its exact and consistent allocentric location. To test this hypothesis, we designed reaction time (RT) experiments to observe evidence of the predicted space-to-time conversion. Various visual stimuli were generated at a wide range of eccentricities either on a large TV (up to 40{degrees}) or by LED strips on a hemispherical dome (up to 60{degrees}). Participants were instructed to report the observed patterns promptly under either covert (no eye movement) or overt (with eye movement) conditions. As predicted, stimuli presented at the center of fixation always produced the fastest RTs. The additional RT delay was precisely proportional to the eccentricity of the peripheral stimulus presentation. Furthermore, both covert and overt attention protocols created the same RT delays, and trajectories of saccadic eye motions were in parallel to the overt RT vs. eccentricity. These findings strongly support our NHT model, in which the observed RT-eccentricity dependence is indicative of the spatiotemporal conversion required for maintaining a stable allocentric frame of reference. That is, we perceive space by time.

neuroscience↗

bioRxiv2022

Visual Perception of 3D Space and Shape in Time - Part III: 2D Shape Recognition by Log-Scaling

Human vision has a remarkable ability to recognize complex 3D objects such as faces that appear with any size and 3D orientations at any 3D location. If we initially memorize a face only with a normalized size and viewed from directly head on, the direct comparison between the one-sized memory and a new incoming image would demand tremendous mental frame translations in 7D. How can we perform such a demanding task so promptly and reliably as we experience the objects in the world around us? Intriguingly, our primary visual cortex exhibits a 2D retinotopy with a log-polar coordinate system, where scaling up/down of shape is converted to linear frame translation. As a result, mental scaling can be performed by linearly translating the memory or the perceptual image until they overlap with each other. According to our new model of NHT (Neural Holography Tomography), alpha brainwaves traveling at a constant speed can conduct this linear translation. With this scheme, every scaling up/down by a factor of two should take the same amount of extra mental time to recognize a smaller/larger face. To test this hypothesis, we designed a reaction time (RT) experiment, where participants were first asked to memorize sets of unfamiliar faces with a given specific size (4{degrees} or 8{degrees}). Following the memorization phase, similar stimuli with a wide range of sizes (from 1{degrees} to 32{degrees}) were presented, and RTs were recorded. As predicted, the increase in RT was proportional to the scaling factor in the log scale. Furthermore, we observed that RTs were fastest for 8{degrees} faces even if the memorized face was 4{degrees}. This supports our hypothesis that we always memorize faces at the exact size of ~8 {degrees}. To our surprise, the increases in RT were also consistent with the mentally-estimated depth sensation, which indicates that the apparent size of the recognized face can create a proper depth sensation.

neuroscience↗

bioRxiv2021

Predicting Individual Task Contrasts From Resting-state Functional Connectivity using a Surface-based Convolutional Network

Task-based and resting-state represent the two most common experimental paradigms of functional neuroimaging. While resting-state offers a flexible and scalable approach for characterizing brain function, task-based techniques provide superior localization. In this paper, we build on recent deep learning methods to create a model that predicts task-based contrast maps from resting-state fMRI scans. Specifically, we propose BrainSurfCNN, a surface-based fully-convolutional neural network model that works with a representation of the brains cortical sheet. Our model achieves state of the art predictive accuracy on independent test data from the Human Connectome Project and yields individual-level predicted maps that are on par with the target-repeat reliability of the measured contrast maps. We also demonstrate that BrainSurfCNN can generalize remarkably well to novel domains with limited training data.

neuroscience↗

bioRxiv2021

A multitask transfer learning framework for novel virus-human protein interactions

AO_SCPLOWBSTRACTC_SCPLOWUnderstanding the interaction patterns between a particular virus and human proteins plays a crucial role in unveiling the underlying mechanism of viral infection. This could further help in developing treatments of viral diseases. The main issues in tackling it as a machine learning problem is the scarcity of training data as well input information of the viral proteins. We overcome these limitations by exploiting powerful statistical protein representations derived from a corpus of around 24 Million protein sequences in a multi task framework. Our experiments on 7 varied benchmark datasets support the superiority of our approach.

bioinformatics↗

bioRxiv2020

Cortical response to naturalistic stimuli is largely predictable with deep neural networks

Naturalistic stimuli, such as movies, activate a substantial portion of the human brain, invoking a response shared across individuals. Encoding models that predict the neural response to a given stimulus can be very useful for studying brain function. However, existing neural encoding models focus on limited aspects of naturalistic stimuli, ignoring the complex and dynamic interactions of modalities in this inherently context-rich paradigm. Using movie watching data from the Human Connectome Project (HCP, N = 158) database, we build group-level models of neural activity that incorporate several inductive biases about information processing in the brain, including hierarchical processing, assimilation over longer timescales and multi-sensory auditory-visual interactions. We demonstrate how incorporating this joint information leads to remarkable prediction performance across large areas of the cortex, well beyond the visual and auditory cortices into multi-sensory sites and frontal cortex. Furthermore, we illustrate that encoding models learn high-level concepts that generalize remarkably well to alternate task-bound paradigms. Taken together, our findings underscore the potential of neural encoding models as a powerful tool for studying brain function in ecologically valid conditions.

neuroscience↗

bioRxiv2020

A consistent evaluation of miRNA-disease association prediction models

MotivationA variety of machine learning based approaches have been applied to predicting miRNA-disease association. Although promising, the evaluation set up to measure prediction performance is inconsistent making it difficult to assess the actual progress. A more acute problem is that most of the models overlook the problem of data leakage due to the use of precomputed miRNA and disease similarity features. ResultsWe unearth a crucial problem of data leakage in evaluation of machine learning models for miRNA-disease association prediction. In particular, information from test set, in the form of precomputed input features for miRNA and disease, is used during training of the model. Moreover, we point out problems in the widely used performance metrics used in model evaluation. While resolving the issues of data leakage and model evaluation, we perform an indepth study of 3 recent models along with our proposed 9 variants of these models. Our proposed variants have resulted in improvements in Average Precision scores (as compared to original models) by approximately 287.7% and 36.7% on HMDDv2.0 (AP:0.504) and HMDDv3.0 (AP: 0.216) datasets respectively. Availability and ImplementationWe release a unified evaluation framework including all models and datasets at https://git.l3s.uni-hannover.de/dong/simplifying_mirna_disease.

bioinformatics↗

bioRxiv2019

Revisiting Feature Selection with Data Complexity for Biomedicine

The identification of biomarkers or predictive features that are indicative of a specific biological or disease state is a major research topic in biomedical applications. Several feature selection(FS) methods ranging from simple univariate methods to recent deep-learning methods have been proposed to select a minimal set of the most predictive features. However, there still lacks the answer to the question of "which method to use when". In this paper, we study the performance of feature selection methods with respect to the underlying datasets statistics and their data complexity measures. We perform a comparative study of 11 feature selection methods over 27 publicly available datasets evaluated over a range of number of selected features using classification as the downstream task. We take the first step towards understanding the FS methods performance from the viewpoint of data complexity. Specifically, we (empirically) show that as regard to classification, the performance of all studied feature selection methods is highly correlated with the error rate of a nearest neighbor based classifier. We also argue about the non-suitability of studied complexity measures to determine the optimal number of relevant features. While looking closely at several other aspects, we also provide recommendations for choosing a particular FS method for a given dataset.

bioinformatics↗