Can the Brain Do Backpropagation? -Exact Implementation of Backpropagation in Predictive Coding Networks.
Ontology highlight
ABSTRACT:
SUBMITTER: Song Y
PROVIDER: S-EPMC7610561 | biostudies-literature |
REPOSITORIES: biostudies-literature
Similar Datasets
Project description:One of the key problems the brain faces is inferring the state of the world from a sequence of dynamically changing stimuli, and it is not yet clear how the sensory system achieves this task. A well-established computational framework for describing perceptual processes in the brain is provided by the theory of predictive coding. Although the original proposals of predictive coding have discussed temporal prediction, later work developing this theory mostly focused on static stimuli, and key questions on neural implementation and computational properties of temporal predictive coding networks remain open. Here, we address these questions and present a formulation of the temporal predictive coding model that can be naturally implemented in recurrent networks, in which activity dynamics rely only on local inputs to the neurons, and learning only utilises local Hebbian plasticity. Additionally, we show that temporal predictive coding networks can approximate the performance of the Kalman filter in predicting behaviour of linear systems, and behave as a variant of a Kalman filter which does not track its own subjective posterior variance. Importantly, temporal predictive coding networks can achieve similar accuracy as the Kalman filter without performing complex mathematical operations, but just employing simple computations that can be implemented by biological networks. Moreover, when trained with natural dynamic inputs, we found that temporal predictive coding can produce Gabor-like, motion-sensitive receptive fields resembling those observed in real neurons in visual areas. In addition, we demonstrate how the model can be effectively generalized to nonlinear systems. Overall, models presented in this paper show how biologically plausible circuits can predict future stimuli and may guide research on understanding specific neural circuits in brain areas involved in temporal prediction.
| S-EPMC11008833 | biostudies-literature
Project description:Neural network training can be slow and energy-expensive due to the frequent transfer of weight data between digital memory and processing units. Neuromorphic systems can accelerate neural networks by performing multiply-accumulate operations in parallel using nonvolatile analog memory. However, executing the widely used backpropagation training algorithm in multilayer neural networks requires information-and therefore storage-of the partial derivatives of the weight values preventing suitable and scalable implementation in hardware. Here, we propose a hardware implementation of the backpropagation algorithm that progressively updates each layer using in situ stochastic gradient descent, avoiding this storage requirement. We experimentally demonstrate the in situ error calculation and the proposed progressive backpropagation method in a multilayer hardware-implemented neural network. We confirm identical learning characteristics and classification performance compared to conventional backpropagation in software. We show that our approach can be scaled to large and deep neural networks, enabling highly efficient training of advanced artificial intelligence computing systems.
| S-EPMC11244533 | biostudies-literature
Project description:Deep-learning models have become pervasive tools in science and engineering. However, their energy requirements now increasingly limit their scalability1. Deep-learning accelerators2-9 aim to perform deep learning energy-efficiently, usually targeting the inference phase and often by exploiting physical substrates beyond conventional electronics. Approaches so far10-22 have been unable to apply the backpropagation algorithm to train unconventional novel hardware in situ. The advantages of backpropagation have made it the de facto training method for large-scale neural networks, so this deficiency constitutes a major impediment. Here we introduce a hybrid in situ-in silico algorithm, called physics-aware training, that applies backpropagation to train controllable physical systems. Just as deep learning realizes computations with deep neural networks made from layers of mathematical functions, our approach allows us to train deep physical neural networks made from layers of controllable physical systems, even when the physical layers lack any mathematical isomorphism to conventional artificial neural network layers. To demonstrate the universality of our approach, we train diverse physical neural networks based on optics, mechanics and electronics to experimentally perform audio and image classification tasks. Physics-aware training combines the scalability of backpropagation with the automatic mitigation of imperfections and noise achievable with in situ algorithms. Physical neural networks have the potential to perform machine learning faster and more energy-efficiently than conventional electronic processors and, more broadly, can endow physical systems with automatically designed physical functionalities, for example, for robotics23-26, materials27-29 and smart sensors30-32.
| S-EPMC8791835 | biostudies-literature
Project description:Two observations about the cortex have puzzled neuroscientists for a long time. First, neural responses are highly variable. Second, the level of excitation and inhibition received by each neuron is tightly balanced at all times. Here, we demonstrate that both properties are necessary consequences of neural networks that represent information efficiently in their spikes. We illustrate this insight with spiking networks that represent dynamical variables. Our approach is based on two assumptions: We assume that information about dynamical variables can be read out linearly from neural spike trains, and we assume that neurons only fire a spike if that improves the representation of the dynamical variables. Based on these assumptions, we derive a network of leaky integrate-and-fire neurons that is able to implement arbitrary linear dynamical systems. We show that the membrane voltage of the neurons is equivalent to a prediction error about a common population-level signal. Among other things, our approach allows us to construct an integrator network of spiking neurons that is robust against many perturbations. Most importantly, neural variability in our networks cannot be equated to noise. Despite exhibiting the same single unit properties as widely used population code models (e.g. tuning curves, Poisson distributed spike trains), balanced networks are orders of magnitudes more reliable. Our approach suggests that spikes do matter when considering how the brain computes, and that the reliability of cortical representations could have been strongly underestimated.
| S-EPMC3828152 | biostudies-literature
Project description:Controlling complex networks is of paramount importance in science and engineering. Despite the recent development of structural controllability theory, we continue to lack a framework to control undirected complex networks, especially given link weights. Here we introduce an exact controllability paradigm based on the maximum multiplicity to identify the minimum set of driver nodes required to achieve full control of networks with arbitrary structures and link-weight distributions. The framework reproduces the structural controllability of directed networks characterized by structural matrices. We explore the controllability of a large number of real and model networks, finding that dense networks with identical weights are difficult to be controlled. An efficient and accurate tool is offered to assess the controllability of large sparse and dense networks. The exact controllability framework enables a comprehensive understanding of the impact of network properties on controllability, a fundamental problem towards our ultimate control of complex systems.
| S-EPMC3945876 | biostudies-other
Project description:The development of state-of-the-art convolutional neural networks (CNN) has allowed researchers to perform plant classification tasks previously thought impossible and rely on human judgment. Researchers often develop complex CNN models to achieve better performances, introducing over-parameterization and forcing the model to overfit on a training dataset. The most popular process for evaluating overfitting in a deep learning model is using accuracy and loss curves. Train and loss curves may help understand the performance of a model but do not provide guidance on how the model could be modified to attain better performance. In this article, we analyzed the relation between the features learned by a model and its capacity and showed that a model with higher representational capacity might learn many subtle features that may negatively affect its performance. Next, we showed that the shallow layers of a deep learning model learn more diverse features than the ones learned by the deeper layers. Finally, we propose SSIM cut curve, a new way to select the depth of a CNN model by using the pairwise similarity matrix between the visualization of the features learned at different depths by using Guided Backpropagation. We showed that our proposed method could potentially pave a new way to select a better CNN model.
| S-EPMC9132261 | biostudies-literature
Project description:While most gene transcription yields RNA transcripts that code for proteins, a sizable proportion of the genome generates RNA transcripts that do not code for proteins, but may have important regulatory functions. The brain-derived neurotrophic factor (BDNF) gene, a key regulator of neuronal activity, is overlapped by a primate-specific, antisense long noncoding RNA (lncRNA) called BDNFOS. We demonstrate reciprocal patterns of BDNF and BDNFOS transcription in highly active regions of human neocortex removed as a treatment for intractable seizures. A genome-wide analysis of activity-dependent coding and noncoding human transcription using a custom lncRNA microarray identified 1288 differentially expressed lncRNAs, of which 26 had expression profiles that matched activity-dependent coding genes and an additional 8 were adjacent to or overlapping with differentially expressed protein-coding genes. The functions of most of these protein-coding partner genes, such as ARC, include long-term potentiation, synaptic activity, and memory. The nuclear lncRNAs NEAT1, MALAT1, and RPPH1, composing an RNAse P-dependent lncRNA-maturation pathway, were also upregulated. As a means to replicate human neuronal activity, repeated depolarization of SY5Y cells resulted in sustained CREB activation and produced an inverse pattern of BDNF-BDNFOS co-expression that was not achieved with a single depolarization. RNAi-mediated knockdown of BDNFOS in human SY5Y cells increased BDNF expression, suggesting that BDNFOS directly downregulates BDNF. Temporal expression patterns of other lncRNA-messenger RNA pairs validated the effect of chronic neuronal activity on the transcriptome and implied various lncRNA regulatory mechanisms. lncRNAs, some of which are unique to primates, thus appear to have potentially important regulatory roles in activity-dependent human brain plasticity.
| S-EPMC3522156 | biostudies-literature
Project description:BACKGROUND: Receptors and scaffold proteins usually possess a high number of distinct binding domains inducing the formation of large multiprotein signaling complexes. Due to combinatorial reasons the number of distinguishable species grows exponentially with the number of binding domains and can easily reach several millions. Even by including only a limited number of components and binding domains the resulting models are very large and hardly manageable. A novel model reduction technique allows the significant reduction and modularization of these models. RESULTS: We introduce methods that extend and complete the already introduced approach. For instance, we provide techniques to handle the formation of multi-scaffold complexes as well as receptor dimerization. Furthermore, we discuss a new modeling approach that allows the direct generation of exactly reduced model structures. The developed methods are used to reduce a model of EGF and insulin receptor crosstalk comprising 5,182 ordinary differential equations (ODEs) to a model with 87 ODEs. CONCLUSION: The methods, presented in this contribution, significantly enhance the available methods to exactly reduce models of combinatorial reaction networks.
| S-EPMC2570670 | biostudies-literature
Project description:Research has demonstrated the importance of economic forecasts for financial decisions at the aggregate economic level. However, little is known about the psychological and neurophysiological mechanisms that economic forecasts activate at the level of individual decision-making. In the present study, we used event-related brain potentials (ERPs) to test the hypothesis that economic forecasts influence individuals' internal model of the economy and their subsequent decision behavior. Using a simple economic decision-making game, the Balloon Analogue of Risk Task (BART) and predictive messages about possible economic changes in the game before each block, we test the idea that brain potentials time-locked to decision outcomes can vary as a function of exposure to economic forecasts. Behavioural results indicate that economic forecasts influenced the amount of risk that participants were willing to take. Analyses of brain potentials indicated parametric increases of the N1, P2, P3a, and P3b amplitudes as a function of the level of risk in subsequent inflation steps in the BART. Mismatches between economic forecasts and decision outcomes in the BART (i.e., reward prediction errors) were reflected in the amplitude of the P2, P3a, and P3b, suggesting increased attentional processing of unexpected outcomes. These electrophysiological results corroborate the idea that economic messages may indeed influence people's beliefs about the economy and bias their subsequent financial decision-making. Our findings present a first important step in the development of a low-level neurophysiological model that may help to explain the self-fulfilling prophecy effect of economic news in the larger economy.
| S-EPMC7497516 | biostudies-literature
Project description:Predictive coding theory suggests the brain anticipates sensory information using prior knowledge. While this theory has been extensively researched within individual sensory modalities, evidence for predictive processing across sensory modalities is limited. Here, we examine how crossmodal knowledge is represented and learned in the brain, by identifying the hierarchical networks underlying crossmodal predictions when information of one sensory modality leads to a prediction in another modality. We record electroencephalogram (EEG) during a crossmodal audiovisual local-global oddball paradigm, in which the predictability of transitions between tones and images are manipulated at both the stimulus and sequence levels. To dissect the complex predictive signals in our EEG data, we employed a model-fitting approach to untangle neural interactions across modalities and hierarchies. The model-fitting result demonstrates that audiovisual integration occurs at both the levels of individual stimulus interactions and multi-stimulus sequences. Furthermore, we identify the spatio-spectro-temporal signatures of prediction-error signals across hierarchies and modalities, and reveal that auditory and visual prediction errors are rapidly redirected to the central-parietal electrodes during learning through alpha-band interactions. Our study suggests a crossmodal predictive coding mechanism where unimodal predictions are processed by distributed brain networks to form crossmodal knowledge.
| S-EPMC11316022 | biostudies-literature