Elsevier

Neurocomputing

Volume 151, Part 1, 3 March 2015, Pages 69-77
Neurocomputing

Regulation of specialists and generalists by neural variability improves pattern recognition performance

https://doi.org/10.1016/j.neucom.2014.09.073Get rights and content

Abstract

To analyze the impact of neural threshold variability in the mushroom body (MB) for pattern recognition, we used a computational model based on the olfactory system of insects. This model is a single-hidden-layer neural network (SLN) where the input layer represents the antennal lobe (AL). The remaining layers are in the MBs that are formed by the Kenyon cell (KC) layer and the output neurons that are responsible for odor learning. The binary code obtained for each odorant in the output layer by unsupervised learning was used to measure the classification error. This classification error allows us to identify the neural variability paradigm that achieves a better odor classification. The neural variability is provided by the neural threshold of activation. We compare two hypotheses: a unique threshold for all the neurons in the MB layer, which leads to no variability (homogeneity), and different thresholds for each MB layer (heterogeneity). The results show that when there is threshold variability, odor classification performance improves. Neural variability induces populations of neurons that are specialists and generalists. Specialist neurons respond to fewer stimulus than the generalists. The proper combination of these two neuron types leads to performance improvement in the bioinspired classifier.

Introduction

The olfactory system of insects is made of a complex neural machinery made of at least four processing layers [1] capable of classifying a large number of odorants from an unlimited number of stimuli that are highly variable [2] (different gas concentrations, mixtures, etc). The main reasons to chose the olfactory system of insects are the simplicity of the structural organization [3], [4], [5], [6], [7], [8], [9], [10], the nature of the neural coding [[2], [11], [12], [13], [14], [15], [16], [17], [18], [19]], the advent of the genetic manipulation techniques that isolate brain areas during the formation of memories [20], [21], [22], [23], and the extensive odor conditioning experiments that shed light into the dynamics of learning during discrimination tasks [24], [25], [26], [27], [28], [29]. Olfactory systems implement simple mechanisms to realize a quick and stable odorant discrimination [30], a goal we want to achieve through computer modeling. Our focus in this work is on neural variability. The driving question is how neural heterogeneity impacts system performance in pattern recognition.

Neural diversity is widespread in the brain, even within the same neural types there is a large heterogeneity in the intrinsic properties and the connectivity patterns, one hypothesis that explains this puzzling observation is functional differentiation within the same types [31]. Another explication is the hypothesis of homeostatic regulation of neural systems, in particular in the olfactory system [32], [33], [34], [35]. However, as we show in this paper, neural heterogeneity can be very beneficial in terms of improving performance in pattern recognition tasks.

Typical models of the olfactory system use very little variability in the excitability in the neurons, implemented by fixed neural thresholds. However, recent applied research on artificial noses determined that using heterogeneous detection thresholds for different odorants, you can improve gas discrimination [36], [37]. This is one of the motivations why we study neuron threshold variability in the information process achieved by the neural olfactory system. Additionally, it has been reported that neural thresholds vary in olfactory receptor neurons (ORN) [38] and in Kenyon cells (KCs) [39]. Neural variability in the form of a broad distribution of thresholds is a generic property of neurons in the brain.

To investigate if neural threshold variability increases odorant classification performance, we use a simple model of the olfactory system [40], [41] based on McCulloch–Pitts neurons [42]. The insect olfactory pathway starts at the antenna, where a massive number of receptors encoding the odor stimulus in a high-dimensional space. This information is then sent to the AL for additional processing. The AL exhibits complex dynamics produced by the interaction of its excitatory and inhibitory neural populations [43], [44], [13]. The excitatory cells are called projection neurons, PNs, because they only transmit the result of AL computation to deeper regions. Moreover, recordings from the AL in the locust indicate that the activity in the projections of the excitatory neurons of the Locust remains nearly constant despite large variations of the odor concentration [45]. Therefore, a gain control mechanism [46], [47] controlling neuronal activity in the AL is likely to exist [48]. The projection neurons deliver the AL output to a very large number cells of Kenyon of the MB using a fan-out connectivity that increases the separability between different odor encodings. This fan-out phase combined with the sparse firing for these KCs [39], [49], [50] facilitates the odorant discrimination process realized in a fan-in phase by output neurons, which are involved in memory formation and storage [51], [52], [20].

We focus on the AL and MB (model in Fig. 1), where the input to single-hidden-layer neural network (SLN) is the AL activity, which is connected to MB through a non-specific connectivity matrix [50]. The reason for this non-specific connectivity matrix is due to the individual connection variability of insects of the same species [53], [54]. The other layers of the SLN, hidden and output, are composed by KC and output neurons, respectively. These are connected by a connectivity matrix that implements Hebbian-like learning [52], [55].

Our goal is to analyze, first, how information is processed in the olfactory system and, second, the role of threshold variability in this system. Hence, we compare the existence of threshold variability (heterogeneous thresholds) with their absence (homogeneous threshold) to determine whether this improves odorant classification. To this end, we measure the classification error obtained in the output layer after applying unsupervised learning. A correctly classified odorant always generates the same output pattern class A for a given input pattern class A.

We conclude that odorant classification can improve with neuron threshold variability or heterogeneity, leading us to label neurons as generalists or specialists [56], [57]. Moreover, the classification performance is closely related to sparse activity of the KC population [39], [58] which can be regulated by neural thresholds too in addition to the connectivity degrees [50].

Section snippets

Neuron model

In locusts, activity patterns in the AL are practically time-discretized by a periodic feedforward inhibition onto the MB calyxes [59] with very low KC activity [39]. Thus, the information is represented by time-discrete, sparse activity patterns with the KCs locked on the 50 ms local field potential oscillation cycle. Because of these neurons are inactive most of the time, but being activated, their neuronal response is produced by the coincidence of concurrent spikes followed by a reset, we

Odorants

We use odorant patterns with different complexity degrees (Fig. 3). The simplest odorants are orthogonal with no overlapping activity. Character odorants have overlapping and contain binary information (input neurons are active or inactive 0,1). Finally, the more complex odorants are extracted from data provided by electronic noses. These are real numbers (input neurons have different degrees of activation) and we refer to them as real odorants. We use 100 patterns of each kind of odorant. This

Threshold selection

Every neuron is capable of firing and transmitting information to the next neuron when it exceeds a certain threshold. To select the value of these thresholds, we use the concept of limit threshold. A limit threshold is the number of stimuli received in a neuron for a given odorant. This is the minimum threshold which prevents a neuron from spiking for a given odorant.

To compare the use of a unique threshold for all neurons in the MB layer (homogeneous thresholds) with the use of different

Results

The results are divided into two parts. First, we compare the paradigm that provides better classification results that use homogeneous thresholds (no threshold variability) or heterogeneous ones (threshold variability). We show the results for different sets of odorants and different connection probabilities for the hidden layer, pc. Finally, we present the results for a particular pc, which shows the relationship between classification error and spikes rate for different odorants.

Conclusions and discussions

The main objective of this work is to analyze how information is processed by investigating the role of threshold variability. To accomplish this, we used a simple model of the olfactory system based on a connectionist model that uses random connections. Thus the model focuses on the AL and MB, where the input to single-hidden-layer neural network (SLN) is the AL activity, which is connected to MB through a non-specific connectivity matrix. The other layers of the SLN, hidden and output, are

Acknowledgments

This work was supported by the Spanish Government Project TIN2010-19607 and Predoctoral Research Grant BES-2011-049274. R.H. acknowledges partial support by NIDCD R01DC011422.

Aaron Montero is a IT engineer with a Master׳s degree in Computer Engineering and a Master׳s degree in Automatic Control and Robotics. Currently, he is a Ph.D. student at the Group of Biological Neurocomputation, Universidad Autonoma de Madrid (Spain). His research area includes neuroscience, artificial intelligence and pattern recognition. His thesis focuses on the study how information is processed in the olfactory system. This study involves the analysis of the olfactory system

References (63)

  • E.C. Marin et al.

    Representation of the glomerular olfactory map in the (Drosophila) brain

    Cell

    (2002)
  • N.K. Tanaka et al.

    Integration of chemosensory pathways in the Drosophila second-order olfactory centers

    Curr. Biol.

    (2004)
  • A. Vergara et al.

    Chemical gas sensor drift compensation using classifier ensembles

    Sensors Actuators B: Chem.

    (2012)
  • M. Heisenberg

    Mushroom body memoir: from maps to models

    Nat. Rev. Neurosci.

    (2003)
  • C.G. Galizia et al.

    Probing the olfactory code

    Nat. Neurosci.

    (2000)
  • C.G. Galizia et al.

    A digital 3D atlas of the honeybee antennal lobe based on optical sections acquired using confocal microscopy

    Cell Tissue Res.

    (1999)
  • M. Mizunami et al.

    Mushroom bodies of the cockroachtheir participation in place memory

    J. Comp. Neurol.

    (1998)
  • K. Ito et al.

    The organization of extrinsic neurons and their implications in the functional roles of the mushroom bodies in Drosophila melanogaster Meigen

    Learn Mem.

    (1998)
  • S.M. Farris et al.

    Development of laminar organization in the mushroom bodies of the cockroachKenyon cell proliferation, outgrowth, and maturation

    J. Comp. Neurol.

    (2001)
  • N.J. Strausfeld

    Organization of the honey bee mushroom bodyrepresentation of the calyx within the vertical and gamma lobes

    J. Comp. Neurol.

    (2002)
  • C.G. Galizia et al.

    Physiological and morphological characterization of honeybee olfactory neurons combining electrophysiology, calcium imaging and confocal microscopy

    J. Comp. Physiol. A

    (2004)
  • J.G. Hildebrand et al.

    Mechanisms of olfactory discriminationconverging evidence for common principles across phyla

    Annu. Rev. Neurosci.

    (1997)
  • R.F. Galan et al.

    Sensory memory for odors is encoded in spontaneous correlated activity between olfactory glomeruli

    Neural. Comput.

    (2006)
  • G. Laurent

    A systems perspective on early olfactory coding

    Science

    (1999)
  • G. Laurent

    Olfactory network dynamics and the coding of multidimensional signals

    Nat. Rev. Neurosci.

    (2002)
  • T. Faber et al.

    Associative learning modifies neural representations of odors in the insect brain

    Nature Neurosci.

    (1999)
  • R. Abel et al.

    Structure and response patterns of olfactory interneurons in the honeybee Apis mellifera

    J. Comp. Neurol.

    (2001)
  • N.J. Vickers et al.

    Odour-plume dynamics influence the brain׳s olfactory code

    Nature

    (2001)
  • K.C. Daly et al.

    Detailed characterization of local field potential oscillations and their relationship to spike timing in the antennal lobe of the moth manduca sexta

    Front. Neuroeng.

    (2011)
  • T. Zars et al.

    Localization of a short-term memory in drosophila

    Science

    (2000)
  • T. Zars et al.

    Tissue-specific expression of a type I adenylyl cyclase rescues the rutabaga mutant memory defectin search of the engram

    Learn Mem.

    (2000)
  • Cited by (0)

    Aaron Montero is a IT engineer with a Master׳s degree in Computer Engineering and a Master׳s degree in Automatic Control and Robotics. Currently, he is a Ph.D. student at the Group of Biological Neurocomputation, Universidad Autonoma de Madrid (Spain). His research area includes neuroscience, artificial intelligence and pattern recognition. His thesis focuses on the study how information is processed in the olfactory system. This study involves the analysis of the olfactory system characteristics that allows odor pattern recognition.

    Ramón Huerta (Ph.D., 1994 – Universidad Autónoma de Madrid) is a research scientist at the BioCircuits Institute, UC San Diego. Prior his current appointment, he was an associate professor at the Universidad Autónoma de Madrid (Spain). His areas of expertise include dynamic systems, artificial intelligence, and neuroscience. His work deals with the development algorithms for the discrimination and quantification of complex multidimensional time series, model building to understand the information processing in the brain, and chemical sensing and machine olfaction applications based on bio-inspired technology. Huerta׳s research work gathers in a publication record of over 100 articles in peer-reviewed journals at the intersection of computer science, physics, and biology.

    Francisco de Borja Rodríguez Ortiz received his degree in Applied Physics in 1992. He then worked at the Instituto de ingeniería del conocimiento from Universidad Autónoma de Madrid (UAM) until 1995. He received his Ph.D. in Computer Science in 1999 from UAM. He has worked at the Nijmegen University in Holland, at the Institute for Nonlinear Science in University of California San Diego, at Centro de Neurociencias Integradas in Santiago de Chile and at Instituto de Física de São Carlos in Universidad de São Paulo. Since 2002 he is “Professor Titular” at the Escuela Politécnica Superior, UAM.

    View full text