Recurrent processing improves occluded object recognition and gives rise to perceptual hysteresis.

Over the past decades, object recognition has been predominantly studied and modelled as a feedforward process. This notion was supported by the fast response times in psychophysical and neurophysiological experiments and the recent success of deep feedforward neural networks for object recognition. Recently, however, this prevalent view has shifted and recurrent connectivity in the brain is now believed to contribute significantly to object recognition - especially under challenging conditions, including the recognition of partially occluded objects. Moreover, recurrent dynamics might be the key to understanding perceptual phenomena such as perceptual hysteresis. In this work we investigate if and how artificial neural networks can benefit from recurrent connections. We systematically compare architectures comprised of bottom-up, lateral, and top-down connections. To evaluate the impact of recurrent connections for occluded object recognition, we introduce three stereoscopic occluded object datasets, which span the range from classifying partially occluded hand-written digits to recognizing three-dimensional objects. We find that recurrent architectures perform significantly better than parameter-matched feedforward models. An analysis of the hidden representation of the models suggests that occluders are progressively discounted in later time steps of processing. We demonstrate that feedback can correct the initial misclassifications over time and that the recurrent dynamics lead to perceptual hysteresis. Overall, our results emphasize the importance of recurrent feedback for object recognition in difficult situations.

Few-shot learning in deep networks through global prototyping.

Training a deep convolution neural network (CNN) to succeed in visual object classification usually requires a great number of examples. Here, starting from such a pre-learned CNN, we study the task of extending the network to classify additional categories on the basis of only few examples ("few-shot learning"). We find that a simple and fast prototype-based learning procedure in the global feature layers ("Global Prototype Learning", GPL) leads to some remarkably good classification results for a large portion of the new classes. It requires only up to ten examples for the new classes to reach a plateau in performance. To understand this few-shot learning performance resulting from GPL as well as the performance of the original network, we use the t-SNE method (Maaten and Hinton, 2008) to visualize clusters of object category examples. This reveals the strong connection between classification performance and data distribution and explains why some new categories only need few examples for learning while others resist good classification results even when trained with many more examples.

Selective Interareal Synchronization through Gamma Frequency Differences and Slower-Rhythm Gamma Phase Reset.

The communication-through-coherence (CTC) hypothesis states that a sending group of neurons will have a particularly strong effect on a receiving group if both groups oscillate in a phase-locked ("coherent") manner (Fries, 2005 , 2015 ). Here, we consider a situation with two visual stimuli, one in the focus of attention and the other distracting, resulting in two sites of excitation at an early cortical area that project to a common site in a next area. Taking a modeler's perspective, we confirm the workings of a mechanism that was proposed by Bosman et al. ( 2012 ) in the context of providing experimental evidence for the CTC hypothesis: a slightly higher gamma frequency of the attended sending site compared to the distracting site may cause selective interareal synchronization with the receiving site if combined with a slow-rhythm gamma phase reset. We also demonstrate the relevance of a slightly lower intrinsic frequency of the receiving site for this scenario. Moreover, we discuss conditions for a transition from bottom-up to top-down driven phase locking.

Attentional Bias Through Oscillatory Coherence Between Excitatory Activity and Inhibitory Minima.

An implementation of attentional bias is presented for a network model that couples excitatory and inhibitory oscillatory units in a manner that is inspired by the mechanisms that generate cortical gamma oscillations. Attentional biases are implemented as oscillatory coherences between excitatory units that encode the spatial location or features of the target and the pool of inhibitory units. This form of attentional bias is motivated by neurophysiological findings that relate selective attention to spike field coherence. Including also pattern recognition mechanisms, we demonstrate how this implementation of attentional bias leads to selection of an attentional target while suppressing distracters for cases of spatial and feature-based attention. With respect to neurophysiological observations, we argue that the recently found positive correlation between high firing rates and strong gamma locking with attention (Vinck, Womelsdorf, Buffalo, Desimone, & Fries, 2013) may point to an essential mechanism of the brain's attentional selection and suppression processes.

The binding problem.

UNLABELLED: The brain processes information in a distributed manner so that features of the sensory input are detected at different sites and subsets of these features are integrated into objects. The notion of 'binding' refers to the corresponding integration process, leading to perception of these objects as entities, and 'the binding problem' either refers to the scientific challenge of identifying mechanisms that may achieve binding or to the difficulty that mind and brain may have with binding in certain situations. This review concentrates on binding of properties in visual perception, but other varieties of the binding problem are also mentioned. The binding problem is reviewed from psychological, neurobiological, and computational perspectives. For further resources related to this article, please visit the WIREs website. CONFLICT OF INTEREST: The author has declared no conflicts of interest for this article.

Temporal coding: the relevance of classical activity, its relation to pattern frequency bands, and a remark on recoding of excitatory drive into phase shifts.

We consider an oscillatory network model that is obtained as complex-valued generalization of the classical Cohen-Grossberg-Hopfield (CGH) model. Apart from a synchronizing mechanism, a stronger and/or more coherent input to a unit in the network implies a higher phase velocity of this unit. This constitutes the desynchronizing mechanism, referred to as acceleration. The units' activity of the classical model translates into the amplitudes of the phase model oscillators. This allows to associate classical and temporal coding with amplitude and phase dynamics, respectively. We discuss how the two dynamics act together to achieve the unambiguous pattern recognition that avoids the superposition problem. With respect to coherence, dominating patterns may take coherent states also if only a subset of its units is on-state. The competition for coherence, introduced by acceleration, realizes a kind of feature counting that identifies the dominating pattern as the pattern with the most on-state units. This dominating but possibly only partially active pattern may take a coherent state with a frequency level that is related to the number of on-state units. We also speculate on neurophysiological findings, related to observed phase differences between optimally and suboptimally activated neurons, that may indicate the presence of acceleration.

Temporal coding: assembly formation through constructive interference.

Temporal coding is studied for an oscillatory neural network model with synchronization and acceleration. The latter mechanism refers to increasing (decreasing) the phase velocity of each unit for stronger (weaker) or more coherent (decoherent) input from the other units. It has been demonstrated that acceleration generates the desynchronization that is needed for self-organized segmentation of two overlapping patterns. In this letter, we continue the discussion of this remarkable feature, giving also an example with several overlapping patterns. Due to acceleration, Hebbian memory implies a frequency spectrum for pure pattern states, defined as coherent patterns with decoherent overlapping patterns. With reference to this frequency spectrum and related frequency bands, the process of pattern retrieval, corresponding to the formation of temporal coding assemblies, is described as resulting from constructive interference (with frequency differences due to acceleration) and phase locking (due to synchronization).

Oscillatory neural networks with self-organized segmentation of overlapping patterns.

Temporal coding is considered with an oscillatory network model that generalizes the Cohen-Grossberg-Hopfield model. It is assumed that the frequency of oscillating units increases with stronger and more coherent input. We refer to this mechanism as acceleration. In the context of Hebbian memory, synchronization and acceleration take complementary roles, and their combined effect on the storage of patterns is profound. Acceleration implies the desynchronization that is needed for self-organized segmention of two overlapping patterns. The superposition problem is thereby solved even without including competition couplings. With respect to brain dynamics, we point to analogies with oscillation spindles in the gamma range and responses to perceptual rivalries.

Oscillatory networks: pattern recognition without a superposition catastrophe.

Using an oscillatory network model that combines classical network models with phase dynamics, we demonstrate how the superposition catastrophe of pattern recognition may be avoided in the context of phase models. The model is designed to meet two requirements: on and off states should correspond, respectively, to high and low phase velocities, and patterns should be retrieved in coherent mode. Nonoverlapping patterns can be simultaneously active with mutually different phases. For overlapping patterns, competition can be used to reduce coherence to a subset of patterns. The model thereby solves the superposition problem.