RESUMO
On several key issues we agree with the commentators. Perhaps most importantly, everyone seems to agree that psychology has an important role to play in building better models of human vision, and (most) everyone agrees (including us) that deep neural networks (DNNs) will play an important role in modelling human vision going forward. But there are also disagreements about what models are for, how DNN-human correspondences should be evaluated, the value of alternative modelling approaches, and impact of marketing hype in the literature. In our view, these latter issues are contributing to many unjustified claims regarding DNN-human correspondences in vision and other domains of cognition. We explore all these issues in this response.
Assuntos
Cognição , Redes Neurais de Computação , HumanosRESUMO
Recent work shows that the developmental potential of progenitor cells in the HH10 chick brain changes rapidly, accompanied by subtle changes in morphology. This demands increased temporal resolution for studies of the brain at this stage, necessitating precise and unbiased staging. Here, we investigated whether we could train a deep convolutional neural network to sub-stage HH10 chick brains using a small dataset of 151 expertly labelled images. By augmenting our images with biologically informed transformations and data-driven preprocessing steps, we successfully trained a classifier to sub-stage HH10 brains to 87.1% test accuracy. To determine whether our classifier could be generally applied, we re-trained it using images (269) of randomised control and experimental chick wings, and obtained similarly high test accuracy (86.1%). Saliency analyses revealed that biologically relevant features are used for classification. Our strategy enables training of image classifiers for various applications in developmental biology with limited microscopy data.
Assuntos
Aprendizado Profundo , Animais , Redes Neurais de Computação , Encéfalo , Microscopia , Asas de AnimaisRESUMO
Convolutional neural networks (CNNs) are often described as promising models of human vision, yet they show many differences from human abilities. We focus on a superhuman capacity of top-performing CNNs, namely, their ability to learn very large datasets of random patterns. We verify that human learning on such tasks is extremely limited, even with few stimuli. We argue that the performance difference is due to CNNs' overcapacity and introduce biologically inspired mechanisms to constrain it, while retaining the good test set generalisation to structured images as characteristic of CNNs. We investigate the efficacy of adding noise to hidden units' activations, restricting early convolutional layers with a bottleneck, and using a bounded activation function. Internal noise was the most potent intervention and the only one which, by itself, could reduce random data performance in the tested models to chance levels. We also investigated whether networks with biologically inspired capacity constraints show improved generalisation to out-of-distribution stimuli, however little benefit was observed. Our results suggest that constraining networks with biologically motivated mechanisms paves the way for closer correspondence between network and human performance, but the few manipulations we have tested are only a small step towards that goal.
Assuntos
Aprendizagem , Redes Neurais de Computação , Humanos , Generalização PsicológicaRESUMO
Deep neural networks (DNNs) have had extraordinary successes in classifying photographic images of objects and are often described as the best models of biological vision. This conclusion is largely based on three sets of findings: (1) DNNs are more accurate than any other model in classifying images taken from various datasets, (2) DNNs do the best job in predicting the pattern of human errors in classifying objects taken from various behavioral datasets, and (3) DNNs do the best job in predicting brain signals in response to images taken from various brain datasets (e.g., single cell responses or fMRI data). However, these behavioral and brain datasets do not test hypotheses regarding what features are contributing to good predictions and we show that the predictions may be mediated by DNNs that share little overlap with biological vision. More problematically, we show that DNNs account for almost no results from psychological research. This contradicts the common claim that DNNs are good, let alone the best, models of human object recognition. We argue that theorists interested in developing biologically plausible models of human vision need to direct their attention to explaining psychological findings. More generally, theorists need to build models that explain the results of experiments that manipulate independent variables designed to test hypotheses rather than compete on making the best predictions. We conclude by briefly summarizing various promising modeling approaches that focus on psychological data.
Assuntos
Redes Neurais de Computação , Percepção Visual , Humanos , Percepção Visual/fisiologia , Visão Ocular , Encéfalo/diagnóstico por imagem , Encéfalo/fisiologia , Imageamento por Ressonância Magnética/métodosRESUMO
Deep Convolutional Neural Networks (DNNs) have achieved superhuman accuracy on standard image classification benchmarks. Their success has reignited significant interest in their use as models of the primate visual system, bolstered by claims of their architectural and representational similarities. However, closer scrutiny of these models suggests that they rely on various forms of shortcut learning to achieve their impressive performance, such as using texture rather than shape information. Such superficial solutions to image recognition have been shown to make DNNs brittle in the face of more challenging tests such as noise-perturbed or out-of-distribution images, casting doubt on their similarity to their biological counterparts. In the present work, we demonstrate that adding fixed biological filter banks, in particular banks of Gabor filters, helps to constrain the networks to avoid reliance on shortcuts, making them develop more structured internal representations and more tolerance to noise. Importantly, they also gained around 20-35% improved accuracy when generalising to our novel out-of-distribution test image sets over standard end-to-end trained architectures. We take these findings to suggest that these properties of the primate visual system should be incorporated into DNNs to make them more able to cope with real-world vision and better capture some of the more impressive aspects of human visual perception such as generalisation.
Assuntos
Redes Neurais de Computação , Percepção Visual , Animais , Generalização Psicológica , Reconhecimento Psicológico , Visão OcularRESUMO
Clinical classification is essential for estimating disease prevalence but is difficult, often requiring complex investigations. The widespread availability of population level genetic data makes novel genetic stratification techniques a highly attractive alternative. We propose a generalizable mathematical framework for determining disease prevalence within a cohort using genetic risk scores. We compare and evaluate methods based on the means of genetic risk scores' distributions; the Earth Mover's Distance between distributions; a linear combination of kernel density estimates of distributions; and an Excess method. We demonstrate the performance of genetic stratification to produce robust prevalence estimates. Specifically, we show that robust estimates of prevalence are still possible even with rarer diseases, smaller cohort sizes and less discriminative genetic risk scores, highlighting the general utility of these approaches. Genetic stratification techniques offer exciting new research tools, enabling unbiased insights into disease prevalence and clinical characteristics unhampered by clinical classification criteria.
Assuntos
Algoritmos , Diabetes Mellitus Tipo 1/genética , Diabetes Mellitus Tipo 2/genética , Predisposição Genética para Doença/genética , Modelos Genéticos , Estudos de Coortes , Simulação por Computador , Diabetes Mellitus Tipo 1/diagnóstico , Diabetes Mellitus Tipo 1/epidemiologia , Diabetes Mellitus Tipo 2/diagnóstico , Diabetes Mellitus Tipo 2/epidemiologia , Humanos , Herança Multifatorial/genética , Polimorfismo de Nucleotídeo Único , Prevalência , Reprodutibilidade dos Testes , Fatores de Risco , Sensibilidade e EspecificidadeRESUMO
Wnt signaling regulates cell proliferation and cell differentiation as well as migration and polarity during development. However, it is still unclear how the Wnt ligand distribution is precisely controlled to fulfil these functions. Here, we show that the planar cell polarity protein Vangl2 regulates the distribution of Wnt by cytonemes. In zebrafish epiblast cells, mouse intestinal telocytes and human gastric cancer cells, Vangl2 activation generates extremely long cytonemes, which branch and deliver Wnt protein to multiple cells. The Vangl2-activated cytonemes increase Wnt/ß-catenin signaling in the surrounding cells. Concordantly, Vangl2 inhibition causes fewer and shorter cytonemes to be formed and reduces paracrine Wnt/ß-catenin signaling. A mathematical model simulating these Vangl2 functions on cytonemes in zebrafish gastrulation predicts a shift of the signaling gradient, altered tissue patterning, and a loss of tissue domain sharpness. We confirmed these predictions during anteroposterior patterning in the zebrafish neural plate. In summary, we demonstrate that Vangl2 is fundamental to paracrine Wnt/ß-catenin signaling by controlling cytoneme behaviour.
Assuntos
Proteínas de Membrana/metabolismo , Pseudópodes/metabolismo , Via de Sinalização Wnt , Animais , Animais Geneticamente Modificados , Padronização Corporal , Embrião não Mamífero/metabolismo , Ativação Enzimática , Fibroblastos/metabolismo , Gastrulação , Células HEK293 , Humanos , Proteínas Quinases JNK Ativadas por Mitógeno/metabolismo , Camundongos Endogâmicos C57BL , Placa Neural/embriologia , Placa Neural/metabolismo , Neurogênese , Comunicação Parácrina , Análise de Sistemas , Telócitos/metabolismo , Peixe-Zebra/embriologia , Peixe-Zebra/metabolismoRESUMO
Computational science has been greatly improved by the use of containers for packaging software and data dependencies. In a scholarly context, the main drivers for using these containers are transparency and support of reproducibility; in turn, a workflow's reproducibility can be greatly affected by the choices that are made with respect to building containers. In many cases, the build process for the container's image is created from instructions provided in a Dockerfile format. In support of this approach, we present a set of rules to help researchers write understandable Dockerfiles for typical data science workflows. By following the rules in this article, researchers can create containers suitable for sharing with fellow scientists, for including in scholarly communication such as education or scientific papers, and for effective and sustainable personal workflows.
Assuntos
Ciência de Dados , Guias como Assunto , Linguagens de Programação , Software , Algoritmos , Reprodutibilidade dos TestesRESUMO
Selectivity remains a challenge for rapid optical vapor sensing via light reflected from porous silicon photonic crystals. This work highlights a method to increase optical vapor selectivity of porous silicon rugate filters by analyzing additive spectra from two rugate filter substrates with different functionalities, an oxidized and carbonized surface. Individually, both porous silicon rugate filters demonstrated sensitivity but not selectivity toward the vapor analytes. However, differences in peak shift trends between the two substrates suggested differences in vapor affinities for the surfaces. By adding the two spectra, improvements to selectivity relative to the individual surfaces were observed even at low vapor pressures and for analytes of similar polarity, refractive index, and concentration. These results are expected to contribute toward optical vapor selectivity improvements in one-dimensional porous silicon photonic crystals.
RESUMO
When deep convolutional neural networks (CNNs) are trained "end-to-end" on raw data, some of the feature detectors they develop in their early layers resemble the representations found in early visual cortex. This result has been used to draw parallels between deep learning systems and human visual perception. In this study, we show that when CNNs are trained end-to-end they learn to classify images based on whatever feature is predictive of a category within the dataset. This can lead to bizarre results where CNNs learn idiosyncratic features such as high-frequency noise-like masks. In the extreme case, our results demonstrate image categorisation on the basis of a single pixel. Such features are extremely unlikely to play any role in human object recognition, where experiments have repeatedly shown a strong preference for shape. Through a series of empirical studies with standard high-performance CNNs, we show that these networks do not develop a shape-bias merely through regularisation methods or more ecologically plausible training regimes. These results raise doubts over the assumption that simply learning end-to-end in standard CNNs leads to the emergence of similar representations to the human visual system. In the second part of the paper, we show that CNNs are less reliant on these idiosyncratic features when we forgo end-to-end learning and introduce hard-wired Gabor filters designed to mimic early visual processing in V1.
Assuntos
Redes Neurais de Computação , Percepção Visual , HumanosRESUMO
Chaste (Cancer, Heart And Soft Tissue Environment) is an open source simulation package for the numerical solution of mathematical models arising in physiology and biology. To date, Chaste development has been driven primarily by applications that include continuum modelling of cardiac electrophysiology ('Cardiac Chaste'), discrete cell-based modelling of soft tissues ('Cell-based Chaste'), and modelling of ventilation in lungs ('Lung Chaste'). Cardiac Chaste addresses the need for a high-performance, generic, and verified simulation framework for cardiac electrophysiology that is freely available to the scientific community. Cardiac chaste provides a software package capable of realistic heart simulations that is efficient, rigorously tested, and runs on HPC platforms. Cell-based Chaste addresses the need for efficient and verified implementations of cell-based modelling frameworks, providing a set of extensible tools for simulating biological tissues. Computational modelling, along with live imaging techniques, plays an important role in understanding the processes of tissue growth and repair. A wide range of cell-based modelling frameworks have been developed that have each been successfully applied in a range of biological applications. Cell-based Chaste includes implementations of the cellular automaton model, the cellular Potts model, cell-centre models with cell representations as overlapping spheres or Voronoi tessellations, and the vertex model. Lung Chaste addresses the need for a novel, generic and efficient lung modelling software package that is both tested and verified. It aims to couple biophysically-detailed models of airway mechanics with organ-scale ventilation models in a package that is freely available to the scientific community. Chaste is designed to be modular and extensible, providing libraries for common scientific computing infrastructure such as linear algebra operations, finite element meshes, and ordinary and partial differential equation solvers. This infrastructure is used by libraries for specific applications, such as continuum mechanics, cardiac models, and cell-based models. The software engineering techniques used to develop Chaste are intended to ensure code quality, re-usability and reliability. Primary applications of the software include cardiac and respiratory physiology, cancer and developmental biology.
RESUMO
We present a reconfigurable neural processor for real-time simulation and prediction of opto-neural behaviour. We combined a detailed Hodgkin-Huxley CA3 neuron integrated with a four-state Channelrhodopsin-2 (ChR2) model into reconfigurable silicon hardware. Our architecture consists of a Field Programmable Gated Array (FPGA) with a custom-built computing data-path, a separate data management system and a memory approach based router. Advancements over previous work include the incorporation of short and long-term calcium and light-dependent ion channels in reconfigurable hardware. Also, the developed processor is computationally efficient, requiring only 0.03 ms processing time per sub-frame for a single neuron and 9.7 ms for a fully connected network of 500 neurons with a given FPGA frequency of 56.7 MHz. It can therefore be utilized for exploration of closed loop processing and tuning of biologically realistic optogenetic circuitry.
Assuntos
Canais Iônicos/química , Modelos Neurológicos , Redes Neurais de Computação , Optogenética , SilícioRESUMO
Optogenetics has become a key tool for understanding the function of neural circuits and controlling their behavior. An array of directly light driven opsins have been genetically isolated from several families of organisms, with a wide range of temporal and spectral properties. In order to characterize, understand and apply these opsins, we present an integrated suite of open-source, multi-scale computational tools called PyRhO. The purpose of developing PyRhO is three-fold: (i) to characterize new (and existing) opsins by automatically fitting a minimal set of experimental data to three-, four-, or six-state kinetic models, (ii) to simulate these models at the channel, neuron and network levels, and (iii) provide functional insights through model selection and virtual experiments in silico. The module is written in Python with an additional IPython/Jupyter notebook based GUI, allowing models to be fit, simulations to be run and results to be shared through simply interacting with a webpage. The seamless integration of model fitting algorithms with simulation environments (including NEURON and Brian2) for these virtual opsins will enable neuroscientists to gain a comprehensive understanding of their behavior and rapidly identify the most suitable variant for application in a particular biological system. This process may thereby guide not only experimental design and opsin choice but also alterations of the opsin genetic code in a neuro-engineering feed-back loop. In this way, we expect PyRhO will help to significantly advance optogenetics as a tool for transforming biological sciences.
RESUMO
Neurons in successive stages of the primate ventral visual pathway encode the spatial structure of visual objects. In this paper, we investigate through computer simulation how these cell firing properties may develop through unsupervised visually-guided learning. Individual neurons in the model are shown to exploit statistical regularity and temporal continuity of the visual inputs during training to learn firing properties that are similar to neurons in V4 and TEO. Neurons in V4 encode the conformation of boundary contour elements at a particular position within an object regardless of the location of the object on the retina, while neurons in TEO integrate information from multiple boundary contour elements. This representation goes beyond mere object recognition, in which neurons simply respond to the presence of a whole object, but provides an essential foundation from which the brain is subsequently able to recognize the whole object.
RESUMO
Learning to recognise objects and faces is an important and challenging problem tackled by the primate ventral visual system. One major difficulty lies in recognising an object despite profound differences in the retinal images it projects, due to changes in view, scale, position and other identity-preserving transformations. Several models of the ventral visual system have been successful in coping with these issues, but have typically been privileged by exposure to only one object at a time. In natural scenes, however, the challenges of object recognition are typically further compounded by the presence of several objects which should be perceived as distinct entities. In the present work, we explore one possible mechanism by which the visual system may overcome these two difficulties simultaneously, through segmenting unseen (artificial) stimuli using information about their category encoded in plastic lateral connections. We demonstrate that these experience-guided lateral interactions robustly organise input representations into perceptual cycles, allowing feed-forward connections trained with spike-timing-dependent plasticity to form independent, translation-invariant output representations. We present these simulations as a functional explanation for the role of plasticity in the lateral connectivity of visual cortex.
Assuntos
Aprendizagem/fisiologia , Modelos Neurológicos , Rede Nervosa/fisiologia , Plasticidade Neuronal/fisiologia , Reconhecimento Visual de Modelos/fisiologia , Córtex Visual/fisiologia , Potenciais de Ação , Animais , Simulação por Computador , Sinais (Psicologia) , Interneurônios/fisiologia , Redes Neurais de Computação , Neurônios/fisiologia , Estimulação Luminosa , PrimatasRESUMO
Over successive stages, the ventral visual system of the primate brain develops neurons that respond selectively to particular objects or faces with translation, size and view invariance. The powerful neural representations found in Inferotemporal cortex form a remarkably rapid and robust basis for object recognition which belies the difficulties faced by the system when learning in natural visual environments. A central issue in understanding the process of biological object recognition is how these neurons learn to form separate representations of objects from complex visual scenes composed of multiple objects. We show how a one-layer competitive network comprised of 'spiking' neurons is able to learn separate transformation-invariant representations (exemplified by one-dimensional translations) of visual objects that are always seen together moving in lock-step, but separated in space. This is achieved by combining 'Mexican hat' functional lateral connectivity with cell firing-rate adaptation to temporally segment input representations of competing stimuli through anti-phase oscillations (perceptual cycles). These spiking dynamics are quickly and reliably generated, enabling selective modification of the feed-forward connections to neurons in the next layer through Spike-Time-Dependent Plasticity (STDP), resulting in separate translation-invariant representations of each stimulus. Variations in key properties of the model are investigated with respect to the network's ability to develop appropriate input representations and subsequently output representations through STDP. Contrary to earlier rate-coded models of this learning process, this work shows how spiking neural networks may learn about more than one stimulus together without suffering from the 'superposition catastrophe'. We take these results to suggest that spiking dynamics are key to understanding biological visual object recognition.
Assuntos
Percepção de Forma/fisiologia , Aprendizagem/fisiologia , Neurônios/fisiologia , Reconhecimento Visual de Modelos/fisiologia , Vias Visuais/fisiologia , Animais , Simulação por Computador , Modelos Neurológicos , Estimulação Luminosa , PrimatasRESUMO
We show how hand-centred visual representations could develop in the primate posterior parietal and premotor cortices during visually guided learning in a self-organizing neural network model. The model incorporates trace learning in the feed-forward synaptic connections between successive neuronal layers. Trace learning encourages neurons to learn to respond to input images that tend to occur close together in time. We assume that sequences of eye movements are performed around individual scenes containing a fixed hand-object configuration. Trace learning will then encourage individual cells to learn to respond to particular hand-object configurations across different retinal locations. The plausibility of this hypothesis is demonstrated in computer simulations.
Assuntos
Simulação por Computador , Redes Neurais de Computação , Software , Algoritmos , Animais , Encéfalo/fisiologia , Mãos/fisiologia , Aprendizagem , Modelos Biológicos , Rede Nervosa/fisiologia , Primatas/fisiologia , Percepção VisualRESUMO
The ventral visual pathway achieves object and face recognition by building transformation-invariant representations from elementary visual features. In previous computer simulation studies with rate-coded neural networks, the development of transformation-invariant representations has been demonstrated using either of two biologically plausible learning mechanisms, Trace learning and Continuous Transformation (CT) learning. However, it has not previously been investigated how transformation-invariant representations may be learned in a more biologically accurate spiking neural network. A key issue is how the synaptic connection strengths in such a spiking network might self-organize through Spike-Time Dependent Plasticity (STDP) where the change in synaptic strength is dependent on the relative times of the spikes emitted by the presynaptic and postsynaptic neurons rather than simply correlated activity driving changes in synaptic efficacy. Here we present simulations with conductance-based integrate-and-fire (IF) neurons using a STDP learning rule to address these gaps in our understanding. It is demonstrated that with the appropriate selection of model parameters and training regime, the spiking network model can utilize either Trace-like or CT-like learning mechanisms to achieve transform-invariant representations.