Training Recurrent Networks Research Articles

Contrastive signal-dependent plasticity: Self-supervised learning in spiking neural circuits.

Brain-inspired machine intelligence research seeks to develop computational models that emulate the information processing and adaptability that distinguishes biological systems of neurons. This has led to the development of spiking neural networks, a class of models that promisingly addresses the biological implausibility and the lack of energy efficiency inherent to modern-day deep neural networks. In this work, we address the challenge of designing neurobiologically motivated schemes for adjusting the synapses of spiking networks and propose contrastive signal-dependent plasticity, a process which generalizes ideas behind self-supervised learning to facilitate local adaptation in architectures of event-based neuronal layers that operate in parallel. Our experimental simulations demonstrate a consistent advantage over other biologically plausible approaches when training recurrent spiking networks, crucially side-stepping the need for extra structure such as feedback synapses.

Start small: Training controllable game level generators without training data by learning at multiple sizes

A level generator is a tool that generates game levels from noise. Training a generator without a dataset suffers from feedback sparsity, since it is unlikely to generate a playable level via random exploration. A common solution is shaped rewards, which guides the generator to achieve subgoals towards level playability, but they consume effort to design and require game-specific domain knowledge. This paper proposes a novel approach to train generators without datasets or shaped rewards by learning at multiple level sizes starting from small sizes and up to the desired sizes. The denser feedback at small sizes negates the need for shaped rewards. Additionally, the generators learn to build levels at various sizes, including sizes they were not trained for. We apply our approach to train recurrent autoregressive generative flow networks (GFlowNets) for controllable level generation. We also adapt diversity sampling to be compatible with GFlowNets. The results show that our generators create diverse playable levels at various sizes for Sokoban, Zelda, and Danger Dave. When compared with controllable reinforcement learning level generators for Sokoban, the results show that our generators achieve better controllability and competitive diversity, while being 9× faster at training and level generation.

Functional Bayesian Filter

We present a general nonlinear Bayesian filter for high-dimensional state estimation using the theory of reproducing kernel Hilbert space (RKHS). By applying the kernel method and the representer theorem to perform linear quadratic estimation in a functional space, we derive a Bayesian recursive state estimator for a general nonlinear dynamical system in the original input space. Unlike existing nonlinear extensions of the Kalman filter where the system dynamics are assumed known, the state-space representation for the Functional Bayesian Filter (FBF) is completely learned online from measurement data in the form of an infinite impulse response (IIR) filter or recurrent network in the RKHS, with universal approximation property. Using a positive definite kernel function satisfying Mercer’s conditions to compute and evolve information quantities, the FBF exploits both the statistical and time-domain information about the signal, extracts higher-order moments, and preserves the properties of covariances without the ill effects due to conventional arithmetic operations. We apply this novel kernel adaptive filtering (KAF) to recurrent network training, chaotic time-series estimation and cooperative filtering using Gaussian and non-Gaussian noises, and inverse kinematics modeling. Simulation results show FBF outperforms existing Kalman-based algorithms.

Extreme neural machines

Recurrent neural networks can solve a variety of computational tasks and produce patterns of activity that capture key properties of brain circuits. However, learning rules designed to train these models are time-consuming and prone to inaccuracies when tuning connection weights located deep within the network. Here, we describe a rapid one-shot learning rule to train recurrent networks composed of biologically-grounded neurons. First, inputs to the model are compressed onto a smaller number of recurrent neurons. Then, a non-iterative rule adjusts the output weights of these neurons based on a target signal. The model learned to reproduce natural images, sequential patterns, as well as a high-resolution movie scene. Together, results provide a novel avenue for one-shot learning in biologically realistic recurrent networks and open a path to solving complex tasks by merging brain-inspired models with rapid optimization rules.

Do Neural Networks for Segmentation Understand Insideness?

The insideness problem is an aspect of image segmentation that consists of determining which pixels are inside and outside a region. Deep neural networks (DNNs) excel in segmentation benchmarks, but it is unclear if they have the ability to solve the insideness problem as it requires evaluating long-range spatial dependencies. In this letter, we analyze the insideness problem in isolation, without texture or semantic cues, such that other aspects of segmentation do not interfere in the analysis. We demonstrate that DNNs for segmentation with few units have sufficient complexity to solve the insideness for any curve. Yet such DNNs have severe problems with learning general solutions. Only recurrent networks trained with small images learn solutions that generalize well to almost any curve. Recurrent networks can decompose the evaluation of long-range dependencies into a sequence of local operations, and learning with small images alleviates the common difficulties of training recurrent networks with a large number of unrolling steps.

A diversity of interneurons and Hebbian plasticity facilitate rapid compressible learning in the hippocampus.

The hippocampus is able to rapidly learn incoming information, even if that information is only observed once. Furthermore, this information can be replayed in a compressed format in either forward or reverse modes during sharp wave-ripples (SPW-Rs). We leveraged state-of-the-art techniques in training recurrent spiking networks to demonstrate how primarily interneuron networks can achieve the following: (1) generate internal theta sequences to bind externally elicited spikes in the presence of inhibition from the medial septum; (2) compress learned spike sequences in the form of a SPW-R when septal inhibition is removed; (3) generate and refine high-frequency assemblies during SPW-R-mediated compression; and (4) regulate the inter-SPW interval timing between SPW-Rs in ripple clusters. From the fast timescale of neurons to the slow timescale of behaviors, interneuron networks serve as the scaffolding for one-shot learning by replaying, reversing, refining, and regulating spike sequences.

Effects of singular value spectrum on the performance of echo state network

Reservoir computing (RC), as an efficient and simple paradigm of training recurrent networks, has attracted considerable attention for its superior performance on many tasks. Echo state network (ESN) is one of typical representatives of RC, whose good performance mainly depends on the selection of right parameters. Hence, it is necessary for users to understand the inter-relationship between the performance and the parameters of ESN. Using computational experiments and statistical analysis, this paper systematically analyzes the effects of singular value spectrum on the performance of ESN, such as memory capacity, robustness, generalization performance, dynamics and so on. To simultaneously control the singular value spectrum and the sparsity when designing an ESN, two algorithms are introduced to construct the reservoir with the predefined singular value spectrum and sparsity, and their benefits are assessed on some benchmark problems.

A Geometrical Analysis of Global Stability in Trained Feedback Networks.

Recurrent neural networks have been extensively studied in the context of neuroscience and machine learning due to their ability to implement complex computations. While substantial progress in designing effective learning algorithms has been achieved, a full understanding of trained recurrent networks is still lacking. Specifically, the mechanisms that allow computations to emerge from the underlying recurrent dynamics are largely unknown. Here we focus on a simple yet underexplored computational setup: a feedback architecture trained to associate a stationary output to a stationary input. As a starting point, we derive an approximate analytical description of global dynamics in trained networks, which assumes uncorrelated connectivity weights in the feedback and in the random bulk. The resulting mean-field theory suggests that the task admits several classes of solutions, which imply different stability properties. Different classes are characterized in terms of the geometrical arrangement of the readout with respect to the input vectors, defined in the high-dimensional space spanned by the network population. We find that such an approximate theoretical approach can be used to understand how standard training techniques implement the input-output task in finite-size feedback networks. In particular, our simplified description captures the local and the global stability properties of the target solution, and thus predicts training performance.

Learning phase transitions from dynamics

We propose the use of recurrent neural networks for classifying phases of matter based on the dynamics of experimentally accessible observables. We demonstrate this approach by training recurrent networks on the magnetization traces of two distinct models of one-dimensional disordered and interacting spin chains. The obtained phase diagram for a well-studied model of the many-body localization transition shows excellent agreement with previously known results obtained from time-independent entanglement spectra. For a periodically-driven model featuring an inherently dynamical time-crystalline phase, the phase diagram that our network traces in a previously-unexplored regime coincides with an order parameter for its expected phases.

Full-FORCE: A target-based method for training recurrent networks.

Trained recurrent networks are powerful tools for modeling dynamic neural computations. We present a target-based method for modifying the full connectivity matrix of a recurrent network to train it to perform tasks involving temporally complex input/output transformations. The method introduces a second network during training to provide suitable “target” dynamics useful for performing the task. Because it exploits the full recurrent connectivity, the method produces networks that perform tasks with fewer neurons and greater noise robustness than traditional least-squares (FORCE) approaches. In addition, we show how introducing additional input signals into the target-generating network, which act as task hints, greatly extends the range of tasks that can be learned and provides control over the complexity and nature of the dynamics of the trained, task-performing network.

Stabilizing patterns in time: Neural network approach.

Recurrent and feedback networks are capable of holding dynamic memories. Nonetheless, training a network for that task is challenging. In order to do so, one should face non-linear propagation of errors in the system. Small deviations from the desired dynamics due to error or inherent noise might have a dramatic effect in the future. A method to cope with these difficulties is thus needed. In this work we focus on recurrent networks with linear activation functions and binary output unit. We characterize its ability to reproduce a temporal sequence of actions over its output unit. We suggest casting the temporal learning problem to a perceptron problem. In the discrete case a finite margin appears, providing the network, to some extent, robustness to noise, for which it performs perfectly (i.e. producing a desired sequence for an arbitrary number of cycles flawlessly). In the continuous case the margin approaches zero when the output unit changes its state, hence the network is only able to reproduce the sequence with slight jitters. Numerical simulation suggest that in the discrete time case, the longest sequence that can be learned scales, at best, as square root of the network size. A dramatic effect occurs when learning several short sequences in parallel, that is, their total length substantially exceeds the length of the longest single sequence the network can learn. This model easily generalizes to an arbitrary number of output units, which boost its performance. This effect is demonstrated by considering two practical examples for sequence learning. This work suggests a way to overcome stability problems for training recurrent networks and further quantifies the performance of a network under the specific learning scheme.

A Review of Heuristic Global Optimization Based Artificial Neural Network Training Approahes

Artificial Neural Networks have earned popularity in recent years because of their ability to approximate nonlinear functions. Training a neural network involves minimizing the mean square error between the target and network output. The error surface is nonconvex and highly multimodal. Finding the minimum of a multimodal function is a NP complete problem and cannot be solved completely. Thus application of heuristic global optimization algorithms that computes a good global minimum to neural network training is of interest. This paper reviews the various heuristic global optimization algorithms used for training feedforward neural networks and recurrent neural networks. The training algorithms are compared in terms of the learning rate, convergence speed and accuracy of the output produced by the neural network. The paper concludes by suggesting directions for novel ANN training algorithms based on recent advances in global optimization.

Dynamic Control of Response Criterion in Premotor Cortex during Perceptual Detection under Temporal Uncertainty

Under uncertainty, the brain uses previous knowledge to transform sensory inputs into the percepts on which decisions are based. When the uncertainty lies in the timing of sensory evidence, however, the mechanism underlying the use of previously acquired temporal information remains unknown. We study this issue in monkeys performing a detection task with variable stimulation times. We use the neural correlates of false alarms to infer the subject's response criterion and find that it modulates over the course of a trial. Analysis of premotor cortex activity shows that this modulation is represented by the dynamics of population responses. A trained recurrent network model reproduces the experimental findings and demonstrates a neural mechanism to benefit from temporal expectations in perceptual detection. Previous knowledge about the probability of stimulation over time can be intrinsically encoded in the neural population dynamics, allowing a flexible control of the response criterion over time.

On the correlation between reservoir metrics and performance for time series classification under the influence of synaptic plasticity.

Reservoir computing provides a simpler paradigm of training recurrent networks by initialising and adapting the recurrent connections separately to a supervised linear readout. This creates a problem, though. As the recurrent weights and topology are now separated from adapting to the task, there is a burden on the reservoir designer to construct an effective network that happens to produce state vectors that can be mapped linearly into the desired outputs. Guidance in forming a reservoir can be through the use of some established metrics which link a number of theoretical properties of the reservoir computing paradigm to quantitative measures that can be used to evaluate the effectiveness of a given design. We provide a comprehensive empirical study of four metrics; class separation, kernel quality, Lyapunov's exponent and spectral radius. These metrics are each compared over a number of repeated runs, for different reservoir computing set-ups that include three types of network topology and three mechanisms of weight adaptation through synaptic plasticity. Each combination of these methods is tested on two time-series classification problems. We find that the two metrics that correlate most strongly with the classification performance are Lyapunov's exponent and kernel quality. It is also evident in the comparisons that these two metrics both measure a similar property of the reservoir dynamics. We also find that class separation and spectral radius are both less reliable and less effective in predicting performance.

Transferring Learning from External to Internal Weights in Echo-State Networks with Sparse Connectivity

Modifying weights within a recurrent network to improve performance on a task has proven to be difficult. Echo-state networks in which modification is restricted to the weights of connections onto network outputs provide an easier alternative, but at the expense of modifying the typically sparse architecture of the network by including feedback from the output back into the network. We derive methods for using the values of the output weights from a trained echo-state network to set recurrent weights within the network. The result of this “transfer of learning” is a recurrent network that performs the task without requiring the output feedback present in the original network. We also discuss a hybrid version in which online learning is applied to both output and recurrent weights. Both approaches provide efficient ways of training recurrent networks to perform complex tasks. Through an analysis of the conditions required to make transfer of learning work, we define the concept of a “self-sensing” network state, and we compare and contrast this with compressed sensing.

A generalized LSTM-like training algorithm for second-order recurrent neural networks

The long short term memory (LSTM) is a second-order recurrent neural network architecture that excels at storing sequential short-term memories and retrieving them many time-steps later. LSTM's original training algorithm provides the important properties of spatial and temporal locality, which are missing from other training approaches, at the cost of limiting its applicability to a small set of network architectures. Here we introduce the generalized long short-term memory(LSTM-g) training algorithm, which provides LSTM-like locality while being applicable without modification to a much wider range of second-order network architectures. With LSTM-g, all units have an identical set of operating instructions for both activation and learning, subject only to the configuration of their local environment in the network; this is in contrast to the original LSTM training algorithm, where each type of unit has its own activation and training instructions. When applied to LSTM architectures with peephole connections, LSTM-g takes advantage of an additional source of back-propagated error which can enable better performance than the original algorithm. Enabled by the broad architectural applicability of LSTM-g, we demonstrate that training recurrent networks engineered for specific tasks can produce better results than single-layer networks. We conclude that LSTM-g has the potential to both improve the performance and broaden the applicability of spatially and temporally local gradient-based training algorithms for recurrent neural networks.

Nonmonotone BFGS-trained recurrent neural networks for temporal sequence processing

In this paper we propose a nonmonotone approach to recurrent neural networks training for temporal sequence processing applications. This approach allows learning performance to deteriorate in some iterations, nevertheless the network’s performance is improved over time. A self-scaling BFGS is equipped with an adaptive nonmonotone technique that employs approximations of the Lipschitz constant and is tested on a set of sequence processing problems. Simulation results show that the proposed algorithm outperforms the BFGS as well as other methods previously applied to these sequences, providing an effective modification that is capable of training recurrent networks of various architectures.

Spurious Valleys in the Error Surface of Recurrent Networks—Analysis and Avoidance

This paper gives a detailed analysis of the error surfaces of certain recurrent networks and explains some difficulties encountered in training recurrent networks. We show that these error surfaces contain many spurious valleys, and we analyze the mechanisms that cause the valleys to appear. We demonstrate that the principle mechanism can be understood through the analysis of the roots of random polynomials. This paper also provides suggestions for improvements in batch training procedures that can help avoid the difficulties caused by spurious valleys, thereby improving training speed and reliability.

Training Recurrent Networks by Evolino

In recent years, gradient-based LSTM recurrent neural networks (RNNs) solved many previously RNN-unlearnable tasks. Sometimes, however, gradient information is of little use for training RNNs, due to numerous local minima. For such cases, we present a novel method: EVOlution of systems with LINear Outputs (Evolino). Evolino evolves weights to the nonlinear, hidden nodes of RNNs while computing optimal linear mappings from hidden state to output, using methods such as pseudo-inverse-based linear regression. If we instead use quadratic programming to maximize the margin, we obtain the first evolutionary recurrent support vector machines. We show that Evolino-based LSTM can solve tasks that Echo State nets (Jaeger, 2004a) cannot and achieves higher accuracy in certain continuous function generation tasks than conventional gradient descent RNNs, including gradient-based LSTM.

Learning in the multiple class random neural network

Spiked recurrent neural networks with "multiple classes" of signals have been recently introduced by Gelenbe and Fourneau (1999), as an extension of the recurrent spiked random neural network introduced by Gelenbe (1989). These new networks can represent interconnected neurons, which simultaneously process multiple streams of data such as the color information of images, or networks which simultaneously process streams of data from multiple sensors. This paper introduces a learning algorithm which applies both to recurrent and feedforward multiple signal class random neural networks (MCRNNs). It is based on gradient descent optimization of a cost function. The algorithm exploits the analytical properties of the MCRNN and requires the solution of a system of nC linear and nC nonlinear equations (where C is the number of signal classes and n is the number of neurons) each time the network learns a new input-output pair. Thus, the algorithm is of O([nC]/sup 3/) complexity for the recurrent case, and O([nC]/sup 2/) for a feedforward MCRNN. Finally, we apply this learning algorithm to color texture modeling (learning), based on learning the weights of a recurrent network directly from the color texture image. The same trained recurrent network is then used to generate a synthetic texture that imitates the original. This approach is illustrated with various synthetic and natural textures.