Sparsity through evolutionary pruning prevents neuronal networks from overfitting

Mutual Correlation Network for few-shot learning

Most metric-based Few-Shot Learning (FSL) methods focus on learning good embeddings of images. However, these methods either lack the ability to explore the cross-correlation (i.e., correlated information) between image pairs or explore limited consensus among the correlation map constrained by the limited receptive field of CNN. We propose a Mutual Correlation Network (MCNet) to explore global consensus among the correlation map by using the self-attention mechanism which has a global receptive field. Our MCNet contains two core modules: (1) a multi-level embedding module that generates multi-level embeddings for an image pair which capture hierarchical semantics, and (2) a mutual correlation module that refines correlation map of two embeddings and generates more robust relational embeddings. Extensive experiments show that our MCNet achieves competitive results on four widely-used few-shot classification benchmarks miniImageNet, tieredImageNet, CUB-200-2011, and CIFAR-FS. Code is available at https://github.com/DRGreat/MCNet.

Quantifying and Maximizing the Information Flux in Recurrent Neural Networks

Free-running recurrent neural networks (RNNs), especially probabilistic models, generate an ongoing information flux that can be quantified with the mutual information I[x→(t),x→(t+1)] between subsequent system states x→. Although previous studies have shown that I depends on the statistics of the network's connection weights, it is unclear how to maximize I systematically and how to quantify the flux in large systems where computing the mutual information becomes intractable. Here, we address these questions using Boltzmann machines as model systems. We find that in networks with moderately strong connections, the mutual information I is approximately a monotonic transformation of the root-mean-square averaged Pearson correlations between neuron pairs, a quantity that can be efficiently computed even in large systems. Furthermore, evolutionary maximization of I[x→(t),x→(t+1)] reveals a general design principle for the weight matrices enabling the systematic construction of systems with a high spontaneous information flux. Finally, we simultaneously maximize information flux and the mean period length of cyclic attractors in the state-space of these dynamical networks. Our results are potentially useful for the construction of RNNs that serve as short-time memories or pattern generators.

Predictive coding and stochastic resonance as fundamental principles of auditory phantom perception

Mechanistic insight is achieved only when experiments are employed to test formal or computational models. Furthermore, in analogy to lesion studies, phantom perception may serve as a vehicle to understand the fundamental processing principles underlying healthy auditory perception. With a special focus on tinnitus-as the prime example of auditory phantom perception-we review recent work at the intersection of artificial intelligence, psychology and neuroscience. In particular, we discuss why everyone with tinnitus suffers from (at least hidden) hearing loss, but not everyone with hearing loss suffers from tinnitus. We argue that intrinsic neural noise is generated and amplified along the auditory pathway as a compensatory mechanism to restore normal hearing based on adaptive stochastic resonance. The neural noise increase can then be misinterpreted as auditory input and perceived as tinnitus. This mechanism can be formalized in the Bayesian brain framework, where the percept (posterior) assimilates a prior prediction (brain's expectations) and likelihood (bottom-up neural signal). A higher mean and lower variance (i.e. enhanced precision) of the likelihood shifts the posterior, evincing a misinterpretation of sensory evidence, which may be further confounded by plastic changes in the brain that underwrite prior predictions. Hence, two fundamental processing principles provide the most explanatory power for the emergence of auditory phantom perceptions: predictive coding as a top-down and adaptive stochastic resonance as a complementary bottom-up mechanism. We conclude that both principles also play a crucial role in healthy auditory perception. Finally, in the context of neuroscience-inspired artificial intelligence, both processing principles may serve to improve contemporary machine learning techniques.

Neural correlates of linguistic collocations during continuous speech perception

Language is fundamentally predictable, both on a higher schematic level as well as low-level lexical items. Regarding predictability on a lexical level, collocations are frequent co-occurrences of words that are often characterized by high strength of association. So far, psycho- and neurolinguistic studies have mostly employed highly artificial experimental paradigms in the investigation of collocations by focusing on the processing of single words or isolated sentences. In contrast, here we analyze EEG brain responses recorded during stimulation with continuous speech, i.e., audio books. We find that the N400 response to collocations is significantly different from that of non-collocations, whereas the effect varies with respect to cortical region (anterior/posterior) and laterality (left/right). Our results are in line with studies using continuous speech, and they mostly contradict those using artificial paradigms and stimuli. To the best of our knowledge, this is the first neurolinguistic study on collocations using continuous speech stimulation.

The impact of sparsity in low-rank recurrent neural networks

Neural population dynamics are often highly coordinated, allowing task-related computations to be understood as neural trajectories through low-dimensional subspaces. How the network connectivity and input structure give rise to such activity can be investigated with the aid of low-rank recurrent neural networks, a recently-developed class of computational models which offer a rich theoretical framework linking the underlying connectivity structure to emergent low-dimensional dynamics. This framework has so far relied on the assumption of all-to-all connectivity, yet cortical networks are known to be highly sparse. Here we investigate the dynamics of low-rank recurrent networks in which the connections are randomly sparsified, which makes the network connectivity formally full-rank. We first analyse the impact of sparsity on the eigenvalue spectrum of low-rank connectivity matrices, and use this to examine the implications for the dynamics. We find that in the presence of sparsity, the eigenspectra in the complex plane consist of a continuous bulk and isolated outliers, a form analogous to the eigenspectra of connectivity matrices composed of a low-rank and a full-rank random component. This analogy allows us to characterise distinct dynamical regimes of the sparsified low-rank network as a function of key network parameters. Altogether, we find that the low-dimensional dynamics induced by low-rank connectivity structure are preserved even at high levels of sparsity, and can therefore support rich and robust computations even in networks sparsified to a biologically-realistic extent.

In large networks of neurons, the activity displayed by the population depends on the strength of the connections between each neuron. In cortical regions engaged in cognitive tasks, this population activity is often seen to be highly coordinated and low-dimensional. A recent line of theoretical work explores how such coordinated activity can arise in a network of neurons in which the matrix defining the connections is constrained to be mathematically low-rank. Until now, this connectivity structure has only been explored in fully-connected networks, in which every neuron is connected to every other. However, in the brain, network connections are often highly sparse, in the sense that most neurons do not share direct connections. Here, we test the robustness of the theoretical framework of low-rank networks to the reality of sparsity present in biological networks. By mathematically analysing the impact of removing connections, we find that the low-dimensional dynamics previously found in dense low-rank networks can in fact persist even at very high levels of sparsity. This has promising implications for the proposal that complex cortical computations which appear to rely on low-dimensional dynamics may be underpinned by a network which has a fundamentally low-rank structure, albeit with only a small fraction of possible connections present.

Neural network based successor representations to form cognitive maps of space and language

How does the mind organize thoughts? The hippocampal-entorhinal complex is thought to support domain-general representation and processing of structural knowledge of arbitrary state, feature and concept spaces. In particular, it enables the formation of cognitive maps, and navigation on these maps, thereby broadly contributing to cognition. It has been proposed that the concept of multi-scale successor representations provides an explanation of the underlying computations performed by place and grid cells. Here, we present a neural network based approach to learn such representations, and its application to different scenarios: a spatial exploration task based on supervised learning, a spatial navigation task based on reinforcement learning, and a non-spatial task where linguistic constructions have to be inferred by observing sample sentences. In all scenarios, the neural network correctly learns and approximates the underlying structure by building successor representations. Furthermore, the resulting neural firing patterns are strikingly similar to experimentally observed place and grid cell firing patterns. We conclude that cognitive maps and neural network-based successor representations of structured knowledge provide a promising way to overcome some of the short comings of deep learning towards artificial general intelligence.

Intrinsic Noise Improves Speech Recognition in a Computational Model of the Auditory Pathway

Noise is generally considered to harm information processing performance. However, in the context of stochastic resonance, noise has been shown to improve signal detection of weak sub- threshold signals, and it has been proposed that the brain might actively exploit this phenomenon. Especially within the auditory system, recent studies suggest that intrinsic noise plays a key role in signal processing and might even correspond to increased spontaneous neuronal firing rates observed in early processing stages of the auditory brain stem and cortex after hearing loss. Here we present a computational model of the auditory pathway based on a deep neural network, trained on speech recognition. We simulate different levels of hearing loss and investigate the effect of intrinsic noise. Remarkably, speech recognition after hearing loss actually improves with additional intrinsic noise. This surprising result indicates that intrinsic noise might not only play a crucial role in human auditory processing, but might even be beneficial for contemporary machine learning approaches.

Dynamics and Information Import in Recurrent Neural Networks

Recurrent neural networks (RNNs) are complex dynamical systems, capable of ongoing activity without any driving input. The long-term behavior of free-running RNNs, described by periodic, chaotic and fixed point attractors, is controlled by the statistics of the neural connection weights, such as the density d of non-zero connections, or the balance b between excitatory and inhibitory connections. However, for information processing purposes, RNNs need to receive external input signals, and it is not clear which of the dynamical regimes is optimal for this information import. We use both the average correlations C and the mutual information I between the momentary input vector and the next system state vector as quantitative measures of information import and analyze their dependence on the balance and density of the network. Remarkably, both resulting phase diagrams C(b, d) and I(b, d) are highly consistent, pointing to a link between the dynamical systems and the information-processing approach to complex systems. Information import is maximal not at the "edge of chaos," which is optimally suited for computation, but surprisingly in the low-density chaotic regime and at the border between the chaotic and fixed point regime. Moreover, we find a completely new type of resonance phenomenon, which we call "Import Resonance" (IR), where the information import shows a maximum, i.e., a peak-like dependence on the coupling strength between the RNN and its external input. IR complements previously found Recurrence Resonance (RR), where correlation and mutual information of successive system states peak for a certain amplitude of noise added to the system. Both IR and RR can be exploited to optimize information processing in artificial neural networks and might also play a crucial role in biological neural systems.

Merging pruning and neuroevolution: towards robust and efficient controllers for modular soft robots

Artificial neural networks (ANNs) can be employed as controllers for robotic agents. Their structure is often complex, with many neurons and connections, especially when the robots have many sensors and actuators distributed across their bodies and/or when high expressive power is desirable. Pruning (removing neurons or connections) reduces the complexity of the ANN, thus increasing its energy efficiency, and has been reported to improve the generalization capability, in some cases. In addition, it is well-known that pruning in biological neural networks plays a fundamental role in the development of brains and their ability to learn. In this study, we consider the evolutionary optimization of neural controllers for the case study of Voxel-based soft robots, a kind of modular, bio-inspired soft robots, applying pruning during fitness evaluation. For a locomotion task, and for centralized as well as distributed controllers, we experimentally characterize the effect of different forms of pruning on after-pruning effectiveness, life-long effectiveness, adaptability to new terrains, and behavior. We find that incorporating some forms of pruning in neuroevolution leads to almost equally effective controllers as those evolved without pruning, with the benefit of higher robustness to pruning. We also observe occasional improvements in generalization ability.

A Hybrid Finite Element-Machine Learning Backward Training Approach to Analyze the Optimal Machining Conditions

As machining processes are complex in nature due to the involvement of large plastic strains occurring at higher strain rates, and simultaneous thermal softening of material, it is necessary for manufacturers to have some manner of determining whether the inputs will achieve the desired outputs within the limitations of available resources. However, finite element simulations—the most common means to analyze and understand the machining of high-performance materials under various cutting conditions and environments—require high amounts of processing power and time in order to output reliable and accurate results which can lead to delays in the initiation of manufacture. The objective of this study is to reduce the time required prior to fabrication to determine how available inputs will affect the desired outputs and machining parameters. This study proposes a hybrid predictive methodology where finite element simulation data and machine learning are combined by feeding the time series output data generated by Finite Element Modeling to an Artificial Neural Network in order to acquire reliable predictions of optimal and/or expected machining inputs (depending on the application of the proposed approach) using what we describe as a backwards training model. The trained network was then fed a test dataset from the simulations, and the results acquired show a high degree of accuracy with regards to cutting force and depth of cut, whereas the predicted/expected feed rate was wildly inaccurate. This is believed to be due to either a limited dataset or the much stronger effect that cutting speed and depth of cut have on power, cutting forces, etc., as opposed to the feed rate. It shows great promise for further research to be performed for implementation in manufacturing facilities for the generation of optimal inputs or the real-time monitoring of input conditions to ensure machining conditions do not vary beyond the norm during the machining process.

Integration of Leaky-Integrate-and-Fire Neurons in Standard Machine Learning Architectures to Generate Hybrid Networks: A Surrogate Gradient Approach

Up to now, modern machine learning (ML) has been based on approximating big data sets with high-dimensional functions, taking advantage of huge computational resources. We show that biologically inspired neuron models such as the leaky-integrate-and-fire (LIF) neuron provide novel and efficient ways of information processing. They can be integrated in machine learning models and are a potential target to improve ML performance. Thus, we have derived simple update rules for LIF units to numerically integrate the differential equations. We apply a surrogate gradient approach to train the LIF units via backpropagation. We demonstrate that tuning the leak term of the LIF neurons can be used to run the neurons in different operating modes, such as simple signal integrators or coincidence detectors. Furthermore, we show that the constant surrogate gradient, in combination with tuning the leak term of the LIF units, can be used to achieve the learning dynamics of more complex surrogate gradients. To prove the validity of our method, we applied it to established image data sets (the Oxford 102 flower data set, MNIST), implemented various network architectures, used several input data encodings and demonstrated that the method is suitable to achieve state-of-the-art classification performance. We provide our method as well as further surrogate gradient methods to train spiking neural networks via backpropagation as an open-source KERAS package to make it available to the neuroscience and machine learning community. To increase the interpretability of the underlying effects and thus make a small step toward opening the black box of machine learning, we provide interactive illustrations, with the possibility of systematically monitoring the effects of parameter changes on the learning characteristics.

Analysis and visualization of sleep stages based on deep neural networks

Automatic sleep stage scoring based on deep neural networks has come into focus of sleep researchers and physicians, as a reliable method able to objectively classify sleep stages would save human resources and simplify clinical routines. Due to novel open-source software libraries for machine learning, in combination with enormous recent progress in hardware development, a paradigm shift in the field of sleep research towards automatic diagnostics might be imminent. We argue that modern machine learning techniques are not just a tool to perform automatic sleep stage classification, but are also a creative approach to find hidden properties of sleep physiology. We have already developed and established algorithms to visualize and cluster EEG data, facilitating first assessments on sleep health in terms of sleep-apnea and consequently reduced daytime vigilance. In the following study, we further analyze cortical activity during sleep by determining the probabilities of momentary sleep stages, represented as hypnodensity graphs and then computing vectorial cross-correlations of different EEG channels. We can show that this measure serves to estimate the period length of sleep cycles and thus can help to find disturbances due to pathological conditions.