A neural network model for the mechanism of feature-extraction

We propose a new multilayered neural network model which has the ability of rapid self-organization. This model is a modified version of the cognitron (Fukushima, 1975). It has modifiable inhibitory feedback connections, as well as conventional modifiable excitatory feedforward connections, between the cells of adjoining layers. If a feature-extracting cell in the network is excited by a stimulus which is already familiar to the network, the cell immediately feeds back inhibitory signals to its presynaptic cells in the preceding layer, which suppresses their response. On the other hand, the feature-extracting cell does not respond to an unfamiliar feature, and the responses from its presynaptic cells are therefore not suppressed because they do not receive any feedback inhibition. Modifiable synapses in the new network are reinforced in a way similar to those in the cognitron, and synaptic connections from cells yielding a large sustained output are reinforced. Since familiar stimulus features do not elicit a sustained response from the cells of the network, only circuits which detect novel stimulus features develop. The network therefore quickly acquires favorable pattern-selectivity by the mere repetitive presentation of set of learning patterns.     

A hierarchical neural network model for associative memory

A hierarchical neural network model with feedback interconnections, which has the function of associative memory and the ability to recognize patterns, is proposed. The model consists of a hierarchical multi-layered network to which efferent connections are added, so as to make positive feedback loops in pairs with afferent connections. The cell-layer at the initial stage of the network is the input layer which receives the stimulus input and at the same time works as an output layer for associative recall. The deepest layer is the output layer for pattern-recognition. Pattern-recognition is performed hierarchically by integrating information by converging afferent paths in the network. For the purpose of associative recall, the integrated information is again distributed to lower-order cells by diverging efferent paths. These two operations progress simultaneously in the network. If a fragment of a training pattern is presented to the network which has completed its self-organization, the entire pattern will gradually be recalled in the initial layer. If a stimulus consisting of a number of training patterns superposed is presented, one pattern gradually becomes predominant in the recalled output after competition between the patterns, and the others disappear. At about the same time when the recalled pattern reaches a steady state in he initial layer, in the deepest layer of the network, a response is elicited from the cell corresponding to the category of the finally-recalled pattern. Once a steady state has been reached, the response of the network is automatically extinguished by inhibitory signals from a steadiness-detecting cell. If the same stimulus is still presented after inhibition, a response for another pattern, formerly suppressed, will now appear, because the cells of the network have adaptation characteristics which makes the same response unlikely to recur. Since inhibition occurs repeatedly, the superposed input patterns are recalled one by one in turn.     

Cognitron: A self-organizing multilayered neural network

A new hypothesis for the organization of synapses between neurons is proposed: “The synapse from neuron x to neuron y is reinforced when x fires provided that no neuron in the vicinity of y is firing stronger than y”. By introducing this hypothesis, a new algorithm with which a multilayered neural network is effectively organized can be deduced. A self-organizing multilayered neural network, which is named “cognitron”, is constructed following this algorithm, and is simulated on a digital computer. Unlike the organization of a usual brain models such as a three-layered perceptron, the self-organization of a cognitron progresses favorably without having a “teacher” which instructs in all particulars how the individual cells respond. After repetitive presentations of several stimulus patterns, the cognitron is self-organized in such a way that the receptive fields of the cells become relatively larger in a deeper layer. Each cell in the final layer integrates the information from whole parts of the first layer and selectively responds to a specific stimulus pattern or a feature.     

Neocognitron: A self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position

A neural network model for a mechanism of visual pattern recognition is proposed in this paper. The network is self-organized by learning without a teacher, and acquires an ability to recognize stimulus patterns based on the geometrical similarity (Gestalt) of their shapes without affected by their positions. This network is given a nickname neocognitron. After completion of self-organization, the network has a structure similar to the hierarchy model of the visual nervous system proposed by Hubel and Wiesel. The network consits of an input layer (photoreceptor array) followed by a cascade connection of a number of modular structures, each of which is composed of two layers of cells connected in a cascade. The first layer of each module consists of S-cells, which show characteristics similar to simple cells or lower order hypercomplex cells, and the second layer consists of C-cells similar to complex cells or higher order hypercomplex cells. The afferent synapses to each S-cell have plasticity and are modifiable. The network has an ability of unsupervised learning: We do not need any teacher during the process of self-organization, and it is only needed to present a set of stimulus patterns repeatedly to the input layer of the network. The network has been simulated on a digital computer. After repetitive presentation of a set of stimulus patterns, each stimulus pattern has become to elicit an output only from one of the C-cell of the last layer, and conversely, this C-cell has become selectively responsive only to that stimulus pattern. That is, none of the C-cells of the last layer responds to more than one stimulus pattern. The response of the C-cells of the last layer is not affected by the pattern's position at all. Neither is it affected by a small change in shape nor in size of the stimulus pattern.     

A self-organizing neural network sharing features of the mammalian visual system

This paper describes a neural network model whose structure is designed to closely fit neuroanatomical and-physiological data, and not to be most suitable for rigorous mathematical analysis.It is shown by computer simulation that a process of self-organization that departs from a fixed retinotopic order at peripheral layers and includes hebbian modifications of synaptic connectivity at higher processing levels leads to a system that is capable of mimicking various functions of visual systems:In the initial state the overall structure of the network is preset, individual connections at higher levels are randomly selected and their strength is initialized with random numbers.For this model the outcome of the self-organization process is determined by the stimulation during the developmental phase. Depending on the type of stimuli used the model can either develop towards a featureselective preprocessor stage in a complex vision system or towards a subsystem for associative recall of abstract patterns.This flexibility supports the hypothesis that the principles embodied are rather universal and can account for the development of various nervous system structures.Presented at teh 9th Cybernetics-Congress, Göttingen, March 1986     

Network Self-Organization Explains the Statistics and Dynamics of Synaptic Connection Strengths in Cortex

The information processing abilities of neural circuits arise from their synaptic connection patterns. Understanding the laws governing these connectivity patterns is essential for understanding brain function. The overall distribution of synaptic strengths of local excitatory connections in cortex and hippocampus is long-tailed, exhibiting a small number of synaptic connections of very large efficacy. At the same time, new synaptic connections are constantly being created and individual synaptic connection strengths show substantial fluctuations across time. It remains unclear through what mechanisms these properties of neural circuits arise and how they contribute to learning and memory. In this study we show that fundamental characteristics of excitatory synaptic connections in cortex and hippocampus can be explained as a consequence of self-organization in a recurrent network combining spike-timing-dependent plasticity (STDP), structural plasticity and different forms of homeostatic plasticity. In the network, associative synaptic plasticity in the form of STDP induces a rich-get-richer dynamics among synapses, while homeostatic mechanisms induce competition. Under distinctly different initial conditions, the ensuing self-organization produces long-tailed synaptic strength distributions matching experimental findings. We show that this self-organization can take place with a purely additive STDP mechanism and that multiplicative weight dynamics emerge as a consequence of network interactions. The observed patterns of fluctuation of synaptic strengths, including elimination and generation of synaptic connections and long-term persistence of strong connections, are consistent with the dynamics of dendritic spines found in rat hippocampus. Beyond this, the model predicts an approximately power-law scaling of the lifetimes of newly established synaptic connection strengths during development. Our results suggest that the combined action of multiple forms of neuronal plasticity plays an essential role in the formation and maintenance of cortical circuits.     

Fast adaptive formation of orthogonalizing filters and associative memory in recurrent networks of neuron-like elements

A demonstration is given that an orthogonalizing filter for patterns is formed adaptively and very rapidly in a network of neuron-like elements with internal feedback connections. It is here assumed that the feedback gain is variable, and proportional to the correlation matrix of the output pattern vectors. The time-dependent signal transfer properties of the complete system are described by a system matrix which satisfies a matrix Bernoulli differential equation; solutions of this equation are outlined. The asymptotic value of the system matrix is shown to correspond to the orthogonal projection operator on the space that is complementary to the space spanned by all of the earlier input pattern vectors. Such a system then acts as a filter, which optimally extracts the amount that is new in an input pattern with respect to all old patterns. It also has features that are directly attributable to a distributed associative memory that is optimally selective.     

Feedback enhances feedforward figure-ground segmentation by changing firing mode

In the visual cortex, feedback projections are conjectured to be crucial in figure-ground segregation. However, the precise function of feedback herein is unclear. Here we tested a hypothetical model of reentrant feedback. We used a previous developed 2-layered feedforward spiking network that is able to segregate figure from ground and included feedback connections. Our computer model data show that without feedback, neurons respond with regular low-frequency (～9 Hz) bursting to a figure-ground stimulus. After including feedback the firing pattern changed into a regular (tonic) spiking pattern. In this state, we found an extra enhancement of figure responses and a further suppression of background responses resulting in a stronger figure-ground signal. Such push-pull effect was confirmed by comparing the figure-ground responses with the responses to a homogenous texture. We propose that feedback controls figure-ground segregation by influencing the neural firing patterns of feedforward projecting neurons.     

Self-organization of day cycle and hierarchical associative memory in “live” neural network

The live neural network model is proposed on the basis of live neuron model and optimal learning rule. By means of numerical simulation the initial stages of neural network self-organization have been shown: (1) the formation of two activity forms, which are identified with sleep and awaking, and (2) the self-organization of hierarchical associative memory when feeding a receptor excitation to the neural network. The energetic profit of self-organization is demonstrated. The formation of neural ensembles, playing the role of generalized neurons, is obtained.     

Associative search network: A reinforcement learning associative memory

An associative memory system is presented which does not require a teacher to provide the desired associations. For each input key it conducts a search for the output pattern which optimizes an external payoff or reinforcement signal. The associative search network (ASN) combines pattern recognition and function optimization capabilities in a simple and effective way. We define the associative search problem, discuss conditions under which the associative search network is capable of solving it, and present results from computer simulations. The synthesis of sensory-motor control surfaces is discussed as an example of the associative search problem.     

The Importance of Lateral Connections in the Parietal Cortex for Generating Motor Plans

Substantial evidence has highlighted the significant role of associative brain areas, such as the posterior parietal cortex (PPC) in transforming multimodal sensory information into motor plans. However, little is known about how different sensory information, which can have different delays or be absent, combines to produce a motor plan, such as executing a reaching movement. To address these issues, we constructed four biologically plausible network architectures to simulate PPC: 1) feedforward from sensory input to the PPC to a motor output area, 2) feedforward with the addition of an efference copy from the motor area, 3) feedforward with the addition of lateral or recurrent connectivity across PPC neurons, and 4) feedforward plus efference copy, and lateral connections. Using an evolutionary strategy, the connectivity of these network architectures was evolved to execute visually guided movements, where the target stimulus provided visual input for the entirety of each trial. The models were then tested on a memory guided motor task, where the visual target disappeared after a short duration. Sensory input to the neural networks had sensory delays consistent with results from monkey studies. We found that lateral connections within the PPC resulted in smoother movements and were necessary for accurate movements in the absence of visual input. The addition of lateral connections resulted in velocity profiles consistent with those observed in human and non-human primate visually guided studies of reaching, and allowed for smooth, rapid, and accurate movements under all conditions. In contrast, Feedforward or Feedback architectures were insufficient to overcome these challenges. Our results suggest that intrinsic lateral connections are critical for executing accurate, smooth motor plans.     

Recognition of general patterns using neural networks

The Hopfield model of neural network stores memory in its symmetric synaptic connections and can only learn to recognize sets of nearly orthogonal patterns. A new algorithm is put forth to permit the recognition of general (non-orthogonal) patterns. The algorithm specifies the construction of the new network's memory matrix T _ij, which is, in general, asymmetrical and contains the Hopfield neural network (Hopfield 1982) as a special case. We find further that in addition to this new algorithm for general pattern recognition, there exists in fact a large class of T _ij memory matrices which permit the recognition of non-orthogonal patterns. The general form of this class of T _ij memory matrix is presented, and the projection matrix neural network (Personnaz et al. 1985) is found as a special case of this general form. This general form of memory matrix extends the library of memory matrices which allow a neural network to recognize non-orthogonal patterns. A neural network which followed this general form of memory matrix was modeled on a computer and successfully recognized a set of non-orthogonal patterns. The new network also showed a tolerance for altered and incomplete data. Through this new method, general patterns may be taught to the neural network.     

A Model of Self-Organizing Head-Centered Visual Responses in Primate Parietal Areas

We present a hypothesis for how head-centered visual representations in primate parietal areas could self-organize through visually-guided learning, and test this hypothesis using a neural network model. The model consists of a competitive output layer of neurons that receives afferent synaptic connections from a population of input neurons with eye position gain modulated retinal receptive fields. The synaptic connections in the model are trained with an associative trace learning rule which has the effect of encouraging output neurons to learn to respond to subsets of input patterns that tend to occur close together in time. This network architecture and synaptic learning rule is hypothesized to promote the development of head-centered output neurons during periods of time when the head remains fixed while the eyes move. This hypothesis is demonstrated to be feasible, and each of the core model components described is tested and found to be individually necessary for successful self-organization.     

A self-organizing model of walking patterns of insects

It is well known that the motor systems of animals are controlled by a hierarchy consisting of a brain, central pattern generator, and effector organs. An animal's walking patterns change depending on its walking velocities, even when it has been decerebrated, which indicates that the walking patterns may, in fact, be generated in the subregions of the neural systems of the central pattern generator and the effector organs. In order to explain the self-organization of the walking pattern in response to changing circumstances, our model incorporates the following ideas: (1) the brain sends only a few commands to the central pattern generator (CPG) which act as constraints to self-organize the walking patterns in the CPG; (2) the neural network of the CPG is composed of oscillating elements such as the KYS oscillator, which has been shown to simulate effectively the diversity of the neural activities; and (3) we have introduced a rule to coordinate leg movement, in which the excitatory and inhibitory interactions among the neurons act to optimize the efficiency of the energy transduction of the effector organs. We describe this mechanism as the least dissatisfaction for the greatest number of elements, which is a self-organization rule in the generation of walking patterns. By this rule, each leg tends to share the load as efficiently as possible under any circumstances. Using this self-organizing model, we discuss the control mechanism of walking patterns.     

A neural network model for selective attention in visual pattern recognition

A neural network model of the mechanism of selective attention in visual pattern recognition is proposed and simulated on a digital computer.When a complex figure consisting of two patterns or more is presented to the model, it is segmented into individual patterns, and each pattern is recognized separately. Even if one of the patterns to which the model is paying selective attention is affected by noise or defects, the model can recall the complete pattern from which the noise has been eliminated and the defects corrected. It is not necessary for perfect recall that the stimulus pattern should be identical in shape to the training pattern. Even though the pattern is distorted in shape or changed in size, it can be correctly recognized and the missing portions restored.The model consists of a hierarchical neural network which has efferent as well as afferent connections between cells. The afferent and the efferent signals interact with each other in the network: the efferent signals, that is, the signals for selective attention, have a facilitating effect on the afferent ones, and, at the same time, the afferent signals gate efferent signal flow. When some feature in the stimulus is not extracted in the afferent paths, the threshold for detection of that feature is automatically lowered by decreasing the efficiency of inhibition, and the model tries to extract even vague traces of the undetected feature.     

Recognition of visual stimuli from multiple neuronal activity in monkey visual cortex

Response patterns recorded with 30 microelectrodes from area 17 of anaesthetized monkeys are analysed. A proportion of the patterns are used to define prototype response patterns. These in turn are used to recognize the stimulus from further non-averaged response patterns. In comparison, recognition by a feedforward neural network is much slower, and slightly inferior. The excitation time structure, with a resolution of about 20 ms, is found to contribute strongly to the recognition. There is some inter-ocular recognition for oriented moving bars, and for on and off phases of switched lights, but none for colours. Generalizations over some stimulus parameters (i.e. cases of confusion) are examined: If small jerking shapes are incorrectly recognized, in general the jerk direction often is the correct one. The onset of a response can most easily be found by determining the dissimilarity relative to spontaneous activity in a sliding window.     

Computational neural networks: Enhancing supervised learning algorithms via self-organization

A neural network processing scheme is proposed which utilizes a self-organizing Kohonen feature map as the front end to a feedforward classifier network. The results of a series of benchmarking studies based upon artificial statistical pattern recognition tasks indicate that the proposed architecture performs significantly better than conventional feedforward classifier networks when the decision regions are disjoint. This is attributed to the fact that the self-organization process allows internal units in the succeeding classifier network to be sensitive to a specific set of features in the input space at the outset of training.     

Discrimination training with multimodal stimuli changes activity in the mushroom body of the hawkmoth Manduca sexta

Background

The mushroom bodies of the insect brain play an important role in olfactory processing, associative learning and memory. The mushroom bodies show odor-specific spatial patterns of activity and are also influenced by visual stimuli.

Methodology/Principal Findings

Functional imaging was used to investigate changes in the in vivo responses of the mushroom body of the hawkmoth Manduca sexta during multimodal discrimination training. A visual and an odour stimulus were presented either together or individually. Initially, mushroom body activation patterns were identical to the odour stimulus and the multimodal stimulus. After training, however, the mushroom body response to the rewarded multimodal stimulus was significantly lower than the response to the unrewarded unimodal odour stimulus, indicating that the coding of the stimuli had changed as a result of training. The opposite pattern was seen when only the unimodal odour stimulus was rewarded. In this case, the mushroom body was more strongly activated by the multimodal stimuli after training. When no stimuli were rewarded, the mushroom body activity decreased for both the multimodal and unimodal odour stimuli. There was no measurable response to the unimodal visual stimulus in any of the experiments. These results can be explained using a connectionist model where the mushroom body is assumed to be excited by olfactory stimulus components, and suppressed by multimodal configurations.

Conclusions

Discrimination training with multimodal stimuli consisting of visual and odour cues leads to stimulus specific changes in the in vivo responses of the mushroom body of the hawkmoth.     

Adaptive synaptogenesis constructs networks that maintain information and reduce statistical dependence

This report demonstrates the effectiveness of two processes in constructing simple feedforward networks which perform good transformations on their inputs. Good transformations are characterized by the minimization of two information measures: the information loss incurred with the transformation and the statistical dependency of the output. The two processes build appropriate synaptic connections in initially unconnected networks. The first process, synaptogenesis, creates new synaptic connections; the second process, associative synaptic modification, adjusts the connection strength of existing synapses. Synaptogenesis produces additional innervation for each output neuron until each output neuron achieves a firing rate of approximately 0.50. Associative modification of existing synaptic connections lends robustness to network construction by adjusting suboptimal choices of initial synaptic weights. Networks constructed using synaptogenesis and synaptic modification successfully preserve the information content of a variety of inputs. By recording a high-dimensional input into an output of much smaller dimension, these networks drastically reduce the statistical dependence of neuronal representations. Networks constructed with synaptogenesis and associative modification perform good transformations over a wide range of neuron firing thresholds.     

A neural network model rapidly learning gains and gating of reflexes necessary to adapt to an arm's dynamics

Effects of dynamic coupling, gravity, inertia and the mechanical impedances of the segments of a multi-jointed arm are shown to be neutralizable through a reflex-like operating three layer static feedforward network. The network requires the proprioceptively mediated actual state variables (here angular velocity and position) of each arm segment. Added neural integrators (and/or differentiators) can make the network exhibit dynamic properties. Then, actual feedback is not necessary and the network can operate in a pure feedforward fashion. Feedforward of an additional load can easily be implemented into the network using descendent gating, and a negative feedback control loop added to the feedforward control reduces errors due to external noise. A training, which combines a least squared error based simultaneous learning rule (LSQ-rule) with a self-imitation algorithm based on direct inverse modeling, enables the network to acquire the whole inverse dynamics, limb parameters included, during one short training movement. The considerations presented also hold for multi-jointed manipulators.