Explore the vast range of titles available.

Deep learning has led to significant advances in artificial intelligence, in part, by adopting strategies motivated by neurophysiology. However, it is unclear whether deep learning could occur in the real brain. Here, we show that a deep learning algorithm that utilizes multi-compartment neurons might help us to understand how the neocortex optimizes cost functions. Like neocortical pyramidal neurons, neurons in our model receive sensory information and higher-order feedback in electrotonically segregated compartments. Thanks to this segregation, neurons in different layers of the network can coordinate synaptic weight updates. As a result, the network learns to categorize images better than a single layer network. Furthermore, we show that our algorithm takes advantage of multilayer architectures to identify useful higher-order representations—the hallmark of deep learning. This work demonstrates that deep learning can be achieved using segregated dendritic compartments, which may help to explain the morphology of neocortical pyramidal neurons. Artificial intelligence has made major progress in recent years thanks to a technique known as deep learning, which works by mimicking the human brain. When computers employ deep learning, they learn by using networks made up of many layers of simulated neurons. Deep learning has opened the door to computers with human – or even super-human – levels of skill in recognizing images, processing speech and controlling vehicles. But many neuroscientists are skeptical about whether the brain itself performs deep learning. The patterns of activity that occur in computer networks during deep learning resemble those seen in human brains. But some features of deep learning seem incompatible with how the brain works. Moreover, neurons in artificial networks are much simpler than our own neurons. For instance, in the region of the brain responsible for thinking and planning, most neurons have complex tree-like shapes. Each cell has ‘roots’ deep inside the brain and ‘branches’ close to the surface. By contrast, simulated neurons have a uniform structure. To find out whether networks made up of more realistic simulated neurons could be used to make deep learning more biologically realistic, Guerguiev et al. designed artificial neurons with two compartments, similar to the ‘roots’ and ‘branches’. The network learned to recognize hand-written digits more easily when it had many layers than when it had only a few. This shows that artificial neurons more like those in the brain can enable deep learning. It even suggests that our own neurons may have evolved their shape to support this process. If confirmed, the link between neuronal shape and deep learning could help us develop better brain-computer interfaces. These allow people to use their brain activity to control devices such as artificial limbs. Despite advances in computing, we are still superior to computers when it comes to learning. Understanding how our own brains show deep learning could thus help us develop better, more human-like artificial intelligence in the future.

A major issue in the application of deep learning is the definition of a proper architecture for the learning machine at hand, in such a way that the model is neither excessively large (which results in overfitting the training data) nor too small (which limits the learning and modeling capabilities of the automatic learner). Facing this issue boosted the development of algorithms for automatically growing and pruning the architectures as part of the learning process. The paper introduces a novel approach to growing the architecture of deep neural networks, called downward-growing neural network (DGNN). The approach can be applied to arbitrary feed-forward deep neural networks. Groups of neurons that negatively affect the performance of the network are selected and grown with the aim of improving the learning and generalization capabilities of the resulting machine. The growing process is realized via replacement of these groups of neurons with sub-networks that are trained relying on ad hoc target propagation techniques. In so doing, the growth process takes place simultaneously in both the depth and width of the DGNN architecture. We assess empirically the effectiveness of the DGNN on several UCI datasets, where the DGNN significantly improves the average accuracy over a range of established deep neural network approaches and over two popular growing algorithms, namely, the AdaNet and the cascade correlation neural network.

We develop biologically plausible training mechanisms for self-supervised learning (SSL) in deep networks. Specifically, by biologically plausible training we mean (i) All updates of weights are based on current activities of pre-synaptic units and current, or activity retrieved from short term memory of post-synaptic units, including at the top-most error computing layer, (ii) Complex computations such as normalization, inner products and division are avoided (iii) Asymmetric connections between units, (iv) Most learning is carried out in an unsupervised manner. SSL with a contrastive loss satisfies the third condition as it does not require labeled data and it introduces robustness to observed perturbations of objects, which occur naturally as objects or observers move in 3d and with variable lighting over time. We propose a contrastive hinge based loss whose error involves simple local computations satisfying (ii), as opposed to the standard contrastive losses employed in the literature, which do not lend themselves easily to implementation in a network architecture due to complex computations involving ratios and inner products. Furthermore we show that learning can be performed with one of two more plausible alternatives to backpropagation that satisfy conditions (i) and (ii). The first is difference target propagation (DTP), which trains network parameters using target-based local losses and employs a Hebbian learning rule (Hebb, 1949), thus overcoming the biologically implausible symmetric weight problem in backpropagation. The second is layer-wise learning, where each layer is directly connected to a layer computing the loss error. The layers are either updated sequentially in a greedy fashion (GLL) or in random order (RLL), and each training stage involves a single hidden layer network. Backpropagation through one layer needed for each such network can either be altered with fixed random feedback (RF) weights as proposed inLillicrap et al. (2016), or using updated random feedback (URF) as in Amit (2019). Both methods represent alternatives to the symmetric weight issue of backpropagation. By training convolutional neural networks (CNNs) with SSL and DTP, GLL or RLL, we find that our proposed framework achieves comparable performance to standard BP learning downstream linear classifier evaluation of the learned embeddings.

A key driver of mammalian intelligence is the ability to represent incoming sensory information across multiple abstraction levels. For example, in the visual ventral stream, incoming signals are first represented as low-level edge filters and then transformed into high-level object representations. Similar hierarchical structures routinely emerge in artificial neural networks (ANNs) trained for object recognition tasks, suggesting that similar structures may underlie biological neural networks. However, the classical ANN training algorithm, backpropagation, is considered biologically implausible, and thus alternative biologically plausible training methods have been developed such as Equilibrium Propagation, Deep Feedback Control, Supervised Predictive Coding, and Dendritic Error Backpropagation. Several of those models propose that local errors are calculated for each neuron by comparing apical and somatic activities. Notwithstanding, from a neuroscience perspective, it is not clear how a neuron could compare compartmental signals. Here, we propose a solution to this problem in that we let the apical feedback signal change the postsynaptic firing rate and combine this with a differential Hebbian update, a rate-based version of classical spiking time-dependent plasticity (STDP). We prove that weight updates of this form minimize two alternative loss functions that we prove to be equivalent to the error-based losses used in machine learning: the inference latency and the amount of top-down feedback necessary. Moreover, we show that the use of differential Hebbian updates works similarly well in other feedback-based deep learning frameworks such as Predictive Coding or Equilibrium Propagation. Finally, our work removes a key requirement of biologically plausible models for deep learning and proposes a learning mechanism that would explain how temporal Hebbian learning rules can implement supervised hierarchical learning.

What can artificial intelligence learn from neuroscience, and vice versa?

It is common statistical learning practice to build models, nowadays largely connectionist models, on very large, static, and fully annotated datasets of identically and independently distributed samples. However, the nature of this setup raises several questions that uncover the brittleness of the models fit to these datasets. What if the data contains few, if any, labels? Labels are generally difficult to come by and require intensive human labor if a dataset is to be fully and properly given ground truth. Alternatively, what if the data is sequential in nature and contains dependencies that span large gaps of time? In the task of modeling the characters of text, a model needs to learn how to spell words as well as how to arrange them in a coherent order in order to produce a meaningful sentence. Or finally, what if the distribution to be learned is dynamic and samples are presented over time, either drawn from a stream or a sequence of task datasets? A model must learn to predict the future well and yet retain previously acquired knowledge. In these settings, traditional machine learning approaches no longer directly apply. Motivated by this issue, we will ultimately address the issue of lifelong learning and the nature of systems that adapt themselves to such distributions. The goal of this thesis is to propose a new family of algorithms that will enable connectionist systems to operate in the lifelong learning setting. In this dissertation, we will specifically: 1) propose a class of deep neural architectures, as well as their learning and inference procedures, that are capable of learning from data with few class labels, 2) develop a novel framework for incorporating simple and effective forms of long-term memory into recurrent neural networks so that they are better able to capture longer-term dependencies in sequential data, and 3) propose a new family of learning algorithms inspired and motivated by neuro-cognitive theory and develop an architecture that can learn without unfolding over time. Finally, we will investigate the challenging problem of sequential and cumulative learning. The work carried forth in this dissertation is meant to serve as a crucial stepping stone towards tackling the greatest challenge facing machine learning and artificial intelligence research efforts–developing an agent that can continually learn.

A concept of atmospheric turbulence characterization using laser light backscattered off a moving unresolved target or a moving target with a glint is considered and analyzed through wave-optics numerical simulations. The technique is based on analysis of the autocorrelation function and variance of the power signal measured by the target-in-the-loop atmospheric sensing (TILAS) system composed of a single-mode-fiber-based optical transceiver and the moving target. It is shown that the TILAS received power signal autocorrelation function strongly depends on the turbulence distribution and is weakly sensitive to the turbulence strength, while the signal variance equally depends on these parameters. Assuming the atmospheric turbulence model can be represented by a single spatially localized turbulence layer and the target position and speed are known independently, consecutive analysis of the autocorrelation function and variance of the TILAS signal allows evaluation of both the turbulence layer strength and position along the optical propagation path. It is also demonstrated that the autocorrelation function can potentially be used for the atmospheric turbulence outer scale estimation.

In an attempt to exploit the automatic feature extraction ability of biologically-inspired deep learning models, and enhance the learning of target features, we propose a novel deep learning algorithm. This is based on a deep convolutional neural network (DCNN) trained with an improved cost function, and combined with a support vector machine (SVM). Specifically, class separation information, which explicitly facilitates intra-class compactness and inter-class separability in the process of learning features, is added to an improved cost function as a regularization term, to enhance the DCNN’s feature extraction ability. The enhanced DCNN is applied to learn the features of Synthetic Aperture Radar (SAR) images, and the SVM is utilized to map features into output labels. Simulation experiments are performed using benchmark SAR image data from the Moving and Stationary Target Acquisition and Recognition (MSTAR) database. Comparative results demonstrate the effectiveness of our proposed method, with an average accuracy of 99% on ten types of targets, including variants and articulated targets. We conclude that our proposed DCNN method has significant potential to be exploited for SAR image target recognition, and can serve as a new benchmark for the research community.

Alzheimer’s Disease (AD) contributes to most cases of dementia. Its prominent neuropathological features are the extracellular neuritic plaques and intercellular neurofibrillary tangles composed of aggregated β-amyloid (Aβ) and hyperphosphorylated tau protein, respectively. In the past few decades, disease-modifying therapy targeting Aβ has been the focus of AD drug development. Even though it is encouraging that two of these drugs have recently received accelerated US Food and Drug Administration approval for AD treatment, their efficacy or long-term safety is controversial. Tau has received increasing attention as a potential therapeutic target, since evidence indicates that tau pathology is more associated with cognitive dysfunction. Moreover, inflammation, especially neuroinflammation, accompanies AD pathological processes and is also linked to cognitive deficits. Accumulating evidence indicates that inflammation has a complex and tight interplay with tau pathology. Here, we review recent evidence on the interaction between tau pathology, focusing on tau post-translational modification and dissemination, and neuroinflammatory responses, including glial cell activation and inflammatory signaling pathways. Then, we summarize the latest clinical trials targeting tau and neuroinflammation. Sustained and increased inflammatory responses in glial cells and neurons are pivotal cellular drivers and regulators of the exacerbation of tau pathology, which further contributes to its worsening by aggravating inflammatory responses. Unraveling the precise mechanisms underlying the relationship between tau pathology and neuroinflammation will provide new insights into the discovery and clinical translation of therapeutic targets for AD and other tau-related diseases (tauopathies). Targeting multiple pathologies and precision therapy strategies will be the crucial direction for developing drugs for AD and other tauopathies.

Recent progress in artificial intelligence has renewed interest in building systems that learn and think like people. Many advances have come from using deep neural networks trained end-to-end in tasks such as object recognition, video games, and board games, achieving performance that equals or even beats that of humans in some respects. Despite their biological inspiration and performance achievements, these systems differ from human intelligence in crucial ways. We review progress in cognitive science suggesting that truly human-like learning and thinking machines will have to reach beyond current engineering trends in both what they learn and how they learn it. Specifically, we argue that these machines should (1) build causal models of the world that support explanation and understanding, rather than merely solving pattern recognition problems; (2) ground learning in intuitive theories of physics and psychology to support and enrich the knowledge that is learned; and (3) harness compositionality and learning-to-learn to rapidly acquire and generalize knowledge to new tasks and situations. We suggest concrete challenges and promising routes toward these goals that can combine the strengths of recent neural network advances with more structured cognitive models.