Biology subjects

Benjamin, A. S.

Publications and source records attributed to Benjamin, A. S..

5 recordsLinked to original sources

bioRxiv2025

TissueFormer: a neural network for labeling tissue from grouped single-cell RNA profiles

BackgroundSingle-cell RNA sequencing technologies have enabled unprecedented insights into gene expression and opened new pathways for diagnostics and tissue annotation. At present, most computational approaches for interpreting single-cell data predict labels or properties based on isolated single-cell transcriptomic profiles. This approach overlooks the cellular composition within a sample, which is often critical for inferring tissue identity or other sample-level phenotypes. ResultsTo address this limitation, we introduce TissueFormer, a Transformer-based neural network that infers population-level labels from groups of single-cell RNA profiles while retaining single-cell resolution. We applied TissueFormer to two tasks: predicting COVID-19 severity from single-cell RNA sequencing of blood samples, and predicting cortical area identity from spatial transcriptomic data in mouse brains. TissueFormer outperformed single-cell foundation models and machine learning methods applied to pseudobulk and cell type composition. ConclusionsTissueFormers higher performance promises more accurate diagnostics and enables the automated construction of high-resolution brain region maps in individual mice directly from spatial transcriptomic data. Applied to mice with developmental perturbations to visual input, these maps revealed a significant reduction in predicted visual cortex area, illustrating how individual differences in neuroanatomy can be quantified. More broadly, TissueFormer provides a framework for predicting any population-level phenotypes which are influenced by cellular diversity and tissue-level organization.

bioinformatics↗

bioRxiv2024

Modeling Attention and Binding in the Brain through Bidirectional Recurrent Gating

Attention is a key component of the visual system, essential for perception, learning, and memory. Attention can also be seen as a solution to the binding problem: concurrent attention to all parts of an entity allows separating it from the rest. However, the rich models of attention in computational neuroscience are generally not scaled to real-world problems and there are thus many behavioral and neural phenomena that current models cannot explain. Here, we propose a bidirectional recurrent model of attention that is inspired by the emerging understanding of biological object-based attention and modern neural networks for image segmentation. It conceptualizes recurrent connections as a multi-stage internal gating process where bottom-up connections transmit features, while top-down and lateral connections transmit attentional gating signals. Our model can recognize and segment simple stimuli such as digits as well as objects in natural images and is able to be prompted with object labels, attributes or locations. It can learn to perform a range of behavioral findings, such as object binding, selective attention, inhibition of return, and visual search. It also replicates a variety of neural findings, including increased activity for attended objects, features, and locations, attention-invariant tuning, and relatively late onset attention. Most importantly, our proposed model unifies decades of cognitive and neurophysiological findings of visual attention into a single principled architecture. Our results highlight that the ability to selectively and dynamically focus on specific parts of stimulus streams can help artificial neural networks to better generalize and align with human brains.

neuroscience↗

bioRxiv2022

Efficient neural codes naturally emerge through gradient descent learning

Animal sensory systems are more sensitive to common features in the environment than uncommon features. For example, small deviations from the more frequently encountered horizontal orientations can be more easily detected than small deviations from the less frequent diagonal ones. Here we find that artificial neural networks trained to recognize objects also have patterns of sensitivity that match the statistics of features in images. To interpret these findings, we show mathematically that learning with gradient descent in deep neural networks preferentially creates representations that are more sensitive to common features, a hallmark of efficient coding. This result suggests that efficient coding naturally emerges from gradient-like learning on natural stimuli.

neuroscience↗

bioRxiv2019

Hue tuning curves in V4 change with visual context

Neurons are often probed by presenting a set of stimuli that vary along one dimension (e.g. color) and quantifying how this stimulus property affect neural activity. An open question, in particular where higher-level areas are involved, is how much tuning measured with one stimulus set reveals about tuning to a new set. Here we ask this question by estimating tuning to hue in macaque V4 from a set of natural scenes and a set of simple color stimuli. We found that hue tuning was strong in each dataset but was not correlated across the datasets, a finding expected if neurons have strong mixed selectivity. We also show how such mixed selectivity may be useful for transmitting information about multiple dimensions of the world. Our finding suggest that tuning in higher visual areas measured with simple stimuli may thus not generalize to naturalistic stimuli. New & NoteworthyVisual cortex is often investigated by mapping neural tuning to variables selected by the researcher such as color. How much does this approach tell us a neurons general role in vision? Here we show that for strongly hue-tuned neurons in V4, estimating hue tuning from artificial stimuli does not reveal the hue tuning in the context of natural scenes. We show how models of optimal information processing suggest that such mixed selectivity maximizes information transmission.

neuroscience↗

bioRxiv2017

Modern machine learning far outperforms GLMs at predicting spikes

Neuroscience has long focused on finding encoding models that effectively ask \"what predicts neural spiking?\" and generalized linear models (GLMs) are a typical approach. It is often unknown how much of explainable neural activity is captured, or missed, when fitting a GLM. Here we compared the predictive performance of GLMs to three leading machine learning methods: feedforward neural networks, gradient boosted trees (using XGBoost), and stacked ensembles that combine the predictions of several methods. We predicted spike counts in macaque motor (M1) and somatosensory (S1) cortices from standard representations of reaching kinematics, and in rat hippocampal cells from open field location and orientation. In general, the modern methods (particularly XGBoost and the ensemble) produced more accurate spike predictions and were less sensitive to the preprocessing of features. This discrepancy in performance suggests that standard feature sets may often relate to neural activity in a nonlinear manner not captured by GLMs. Encoding models built with machine learning techniques, which can be largely automated, more accurately predict spikes and can offer meaningful benchmarks for simpler models.

neuroscience↗