In the olfactory system of male moths, a specialized subset of neurons detects and processes the main component of the sex pheromone emitted by females. It is composed of several thousand first-order olfactory receptor neurons (ORNs), all expressing the same pheromone receptor, that contact synaptically a few tens of second-order projection neurons (PNs) within a single restricted brain area. The functional simplicity of this system makes it a favorable model for studying the factors that contribute to its exquisite sensitivity and speed. Sensory information--primarily the identity and intensity of the stimulus--is encoded as the firing rate of the action potentials, and possibly as the latency of the neuron response. We found that over all their dynamic range, PNs respond with a shorter latency and a higher firing rate than most ORNs. Modelling showed that the increased sensitivity of PNs can be explained by the ORN-to-PN convergent architecture alone, whereas their faster response also requires cell-to-cell heterogeneity of the ORN population. So, far from being detrimental to signal detection, the ORN heterogeneity is exploited by PNs, and results in two different schemes of population coding based either on the response of a few extreme neurons (latency) or on the average response of many (firing rate). Moreover, ORN-to-PN transformations are linear for latency and nonlinear for firing rate, suggesting that latency could be involved in concentration-invariant coding of the pheromone blend and that sensitivity at low concentrations is achieved at the expense of precise encoding at high concentrations.

Male moths aiming to locate pheromone-releasing females rely on stimulus-adapted search maneuvers complicated by a discontinuous distribution of pheromone patches. They alternate sequences of upwind surge when perceiving the pheromone and cross- or downwind casting when the odor is lost. We compare four search strategies: three reactive versus one cognitive. The former consist of pre-programmed movement sequences triggered by pheromone detections while the latter uses Bayesian inference to build spatial probability maps. Based on the analysis of triphasic responses of antennal lobe neurons (On, inhibition, Off), we propose three reactive strategies. One combines upwind surge (representing the On response to a pheromone detection) and spiral casting, only. The other two additionally include crosswind (zigzag) casting representing the Off phase. As cognitive strategy we use the infotaxis algorithm which was developed for searching in a turbulent medium. Detection events in the electroantennogram of a moth attached to a robot indirectly control this cyborg, depending on the strategy in use. The recorded trajectories are analyzed with regard to success rates, efficiency, and other features. In addition, we qualitatively compare our robotic trajectories to behavioral search paths. Reactive searching is more efficient (yielding shorter trajectories) for higher pheromone doses whereas cognitive searching works better for lower doses. With respect to our experimental conditions (2 m from starting position to pheromone source), reactive searching with crosswind zigzag yields the shortest trajectories (for comparable success rates). Assuming that the neuronal Off response represents a short-term memory, zigzagging is an efficient movement to relocate a recently lost pheromone plume. Accordingly, such reactive strategies offer an interesting alternative to complex cognitive searching.

Existing gas recognition techniques rely on complex signal processing techniques. This paper presents a simple bio-inspired gas recognition technique, exploiting fundamental characteristics of the organization of the olfactory pathway. The technique was validated using an-in house custom-fabricated tin gas sensor array together with three target gases: ethanol, methane, and carbon monoxide. Experimental results show that the proposed approach provides high accuracy, enabling the concept of a fully integrated electronic nose.

In traditional event-driven strategies, spike timings are analytically given or calculated with arbitrary precision (up to machine precision). Exact computation is possible only for simplified neuron models, mainly the leaky integrate-and-fire model. In a recent paper, Zheng, Tonnelier, and Martinez (2009) introduced an approximate event-driven strategy, named voltage stepping, that allows the generic simulation of nonlinear spiking neurons. Promising results were achieved in the simulation of single quadratic integrate-and-fire neurons. Here, we assess the performance of voltage stepping in network simulations by considering more complex neurons (quadratic integrate-and-fire neurons with adaptation) coupled with multiple synapses. To handle the discrete nature of synaptic interactions, we recast voltage stepping in a general framework, the discrete event system specification. The efficiency of the method is assessed through simulations and comparisons with a modified time-stepping scheme of the Runge-Kutta type. We demonstrated numerically that the original order of voltage stepping is preserved when simulating connected spiking neurons, independent of the network activity and connectivity.

Dealing with high-throughput information systems is becoming an everyday problem in many fields of science, as technological advances improve our ability to gather data. In particular, the information encoding problem in high-dimensional spaces is a crucial aspect to consider. In fact, biological systems are known to be very efficient at encoding and processing high-dimensional information. Here we propose a biologically-based solution that mimics the neural processing performed by the Antennal Lobe of insects. Based on our understanding of this system, our model exploits plausible neural mechanisms to transform the massive and high-dimensional spatial and temporal input of the olfactory receptor neurons into a neural population encoding based on synchrony and frequency, consistent with known physiology. We demonstrate the capabilities of our Antennal Lobe model in the context of a classification task of different olfactory stimuli of varying concentrations. We show that the generated neural representation conveys both the identity and the concentration of each stimuli.

In this paper, we propose a biologically inspired architecture for a Hopfield-like associative memory applied to artificial olfaction. The proposed algorithm captures the projection between two neural layers of the insect olfactory system (Antennal Lobe and Mushroom Body) with a kernel based projection. We have tested its classification performance as a function of the size of the training set and the time elapsed since training and compared it with that obtained with a Support Vector Machine.

Electronic nose (EN) system based on an array of gas sensors and its recognition techniques have been widely investigated in recent decades. However, the integration of a portable EN system is not mature yet. We present an integrated EN system based on an array of gas sensor fabricated in our in-house foundry. A data acquisition technique based on a resistance-to-time conversion is proposed enabling to remove the need for an analog-to-digital converter enabling to reduce the power consumption and area. This ADC-free system is implemented in a multi-chip platform and enabling gas sensing and recognition functionalities. This system is compact, low cost and low power which is suitable for portable applications. Both sensing and recognition operations have been successfully demonstrated in our demo system.

This paper presents a robust gas identification technique for tin oxide (SnO2) gas sensors. The proposed technique generates a unique spike pattern or signature for each sensed gas, irrespective of its concentration. The proposed gas identification technique is insensitive to drift in the sensor baseline resistance. Furthermore, its calibration requires a single measurement to be made for each targeted gas. The proposed spike-based gas identification technique has been implemented in TSMC 0.18 mum CMOS technology and validated using experimental data from a fabricated in-house 4 times 4 SnO2gas sensor array. Reported results reveal a 10% increase in correct gas detection rate.

It has been proposed that synchronized neural assemblies in the antennal lobe of insects encode the identity of olfactory stimuli. In response to an odor, some projection neurons exhibit synchronous firing, phase-locked to the oscillations of the field potential, whereas others do not. Experimental data indicate that neural synchronization and field oscillations are induced by fast GABA(A)-type inhibition, but it remains unclear how desynchronization occurs. We hypothesize that slow inhibition plays a key role in desynchronizing projection neurons. Because synaptic noise is believed to be the dominant factor that limits neuronal reliability, we consider a computational model of the antennal lobe in which a population of oscillatory neurons interact through unreliable GABA(A) and GABA(B) inhibitory synapses. From theoretical analysis and extensive computer simulations, we show that transmission failures at slow GABA(B) synapses make the neural response unpredictable. Depending on the balance between GABA(A) and GABA(B) inputs, particular neurons may either synchronize or desynchronize. These findings suggest a wiring scheme that triggers stimulus-specific synchronized assemblies. Inhibitory connections are set by Hebbian learning and selectively activated by stimulus patterns to form a spiking associative memory whose storage capacity is comparable to that of classical binary-coded models. We conclude that fast inhibition acts in concert with slow inhibition to reformat the glomerular input into odor-specific synchronized neural assemblies.

This paper presents a 4 times 4 logarithmic spike-timing encoding scheme used to translate the output of an integrated tin oxide gas sensor array into spike sequence, which is exploited to perform gas recognition. Hydrogen, ethanol and carbon monoxide were used to characterize the gas sensor array. The collected data were then used to test the proposed circuit for spike encoding and gas recognition. Simulation results illustrate that a particular analyte gas generates a unique spike pattern with certain spike ordering sequence, which is independent of the gas concentration. This unique spike sequence can thus be used to recognize different gases. In addition, the concentration information can also be extracted from the time-to-the-first spike in the sequence making it possible to perform not only gas/odor recognition but quantification as well.

Event-driven strategies have been used to simulate spiking neural networks exactly. Previous work is limited to linear integrate-and-fire neurons. In this note, we extend event-driven schemes to a class of nonlinear integrate-and-fire models. Results are presented for the quadratic integrate-and-fire model with instantaneous or exponential synaptic currents. Extensions to conductance-based currents and exponential integrate-and-fire neurons are discussed.

In the insect olfactory system, odor-evoked transient synchronization of antennal lobe (AL) projection neurons (PNs) is phase-locked to the oscillations of the local field potential. Sensory information is contained in the spatiotemporal synchronization pattern formed by the identities of the phase-locked PNs. This article investigates the role of feedback inhibition from the local neurons (LNs) in this coding. First, experimental biological results are reproduced with a reduced computational spiking neural network model of the AL. Second, the low complexity of the model leads to a mathematical analysis from which a lower bound on the phase-locking probability is derived. Parameters involved in the bound indicate that PN phase locking depends not only on the number of LN-evoked inhibitory postsynaptic potentials (IPSPs) previously received, but also on their temporal jitter. If the inhibition received by a PN at the current oscillatory cycle is both perfectly balanced (i.e., equal to the mean inhibitory drive) and precise (without any jitter), then the PN will be phase-locked at the next oscillatory cycle with probability one.

This paper describes a hardware implementation of threshold network ensembles (TNE) for classification applications. We first describe the algorithm and compare its performance with those of individual classifiers such as binary neural network and support vector machine (SVM). The effect of limited precision on the performance of threshold network ensembles is also investigated. The proposed multi-precision architecture is then mapped into a scalable systolic architecture implemented first on a single VLSI chip. The modularity and the easy programability of the basic chip has made possible the extension of the architecture to a low cost multi-chip solution. We propose a 3D packaged circuit in which 12 basic chips have been integrated into a very compact volume of (2 /spl times/ 2 /spl times/ 0.7)cm/sup 3/. Successful operation of the 3D prototype is demonstrated through experimental test results of the chip.

A bagging ensemble consists of a set of classifiers trained independently and combined by a majority vote. Such a combination improves generalization performance but can require large amounts of memory and computation, a serious drawback for addressing portable real-time pattern recognition applications. We report here a compact three-dimensional (3D) multiprecision very large-scale integration (VLSI) implementation of a bagging ensemble. In our circuit, individual classifiers are decision trees implemented as threshold networks - one layer of threshold logic units (TLUs) followed by combinatorial logic functions. The hardware was fabricated using 0.7-/spl mu/m CMOS technology and packaged using MCM-V micro-packaging technology. The 3D chip implements up to 192 TLUs operating at a speed of up to 48 GCPPS and implemented in a volume of (/spl omega/ /spl times/ L /spl times/ h) = (2 /spl times/ 2 /spl times/ 0.7) cm/sup 3/. The 3D circuit features a high level of programmability and flexibility offering the possibility to make an efficient use of the hardware resources in order to reduce the power consumption. Successful operation of the 3D chip for various precisions and ensemble sizes is demonstrated through an electronic nose application.

We derive a new method for solving nonlinear blind source separation (BSS) problems by exploiting second-order statistics in a kernel induced feature space. This paper extends a new and efficient closed-form linear algorithm to the nonlinear domain using the kernel trick originally applied in support vector machines (SVMs). This technique could likewise be applied to other linear covariance-based source separation algorithms. Experiments on realistic nonlinear mixtures of speech signals, gas multisensor data, and visual disparity data illustrate the applicability of our approach.

In Slow Feature Analysis (SFA [1]), it has been demonstrated that high-order invariant properties can be extracted by projecting in(cid:173) puts into a nonlinear space and computing the slowest changing features in this space; this has been proposed as a simple general model for learning nonlinear invariances in the visual system. How(cid:173) ever, this method is highly constrained by the curse of dimension(cid:173) ality which limits it to simple theoretical simulations. This paper demonstrates that by using a different but closely-related objective function for extracting slowly varying features ([2, 3]), and then ex(cid:173) ploiting the kernel trick, this curse can be avoided. Using this new method we show that both the complex cell properties of transla(cid:173) tion invariance and disparity coding can be learnt simultaneously from natural images when complex cells are driven by simple cells also learnt from the image. The notion of maximising an objective function based upon the temporal pre(cid:173) dictability of output has been progressively applied in modelling the development of invariances in the visual system. F6ldiak used it indirectly via a Hebbian trace rule for modelling the development of translation invariance in complex cells [4] (closely related to many other models [5,6,7]); this rule has been used to maximise invariance as one component of a hierarchical system for object and face recognition [8]. On the other hand, similar functions have been maximised directly in networks for extracting linear [2] and nonlinear [9, 1] visual invariances. Direct maximisation of such functions have recently been used to model complex cells [10] and as an alternative to maximising sparseness/independence in modelling simple cells [11]. Slow Feature Analysis [1] combines many of the best properties of these methods to provide a good general nonlinear model. That is, it uses an objective function that minimises the first-order temporal derivative of the outputs; it provides a closed(cid:173) form solution which maximises this function by projecting inputs into a nonlinear http://www.loria.fr/equipes/cortex/ space; it exploits sphering (or PCA-whitening) of the data to ensure that all outputs have unit variance and are uncorrelated. However, the method suffers from the curse of dimensionality in that the nonlinear feature space soon becomes very large as the input dimension grows, and yet this feature space must be represented explicitly in order for the essential sphering to occur. The alternative that we propose here is to use the objective function of Stone [2, 9], that maximises output variance over a long period whilst minimising variance over a shorter period; in the linear case, this can be implemented by a biologically plausible mixture of Hebbian and anti-Hebbian learning on the same synapses [2]. In recent work, Stone has proposed a closed-form solution for maximising this function in the linear domain of blind source separation that does not involve data-sphering. This paper describes how this method can be kernelised. The use of the "kernel trick" allows projection of inputs into a nonlinear kernel induced feature space of very high (possibly infinite) dimension which is never explicitly represented or accessed. This leads to an efficient method that maps to an architecture that could be biologically implemented either by Sigma-Pi neurons, or fixed REF networks (as described for SFA [1]). We demonstrate that using this method to extract features that vary slowly in natural images leads to the development of both the complex-cell properties of translation invariance and disparity coding simultaneously. 1 Finding Slow Features with kernels Given I time-series vectors X i F _ V _ L.i Yi 2 L.i Yi

In this paper, a neural competitive learning tree is introduced as a computationally attractive scheme for adaptive density estimation and novelty detection. The learning rule yields equiprobable quantization of the input space and provides an adaptive focusing mechanism capable of tracking time-varying distributions. It is shown by simulation that the neural tree performs reasonably well while being much faster than any of the other competitive learning algorithms.

This letter presents a novel unsupervised competitive learning rule called the boundary adaptation rule (BAR), for scalar quantization. It is shown both mathematically and by simulations that BAR converges to equiprobable quantizations of univariate probability density functions and that, in this way, it outperforms other unsupervised competitive learning rules.

A novel unsupervised learning rule, called Boundary Adaptation Rule (BAR), is introduced for scalar quantization. It is shown that the rule maximizes information-theoretic entropy and thus yields equiprobable quantizations of univariate probability density functions. It is shown by simulations that BAR outperforms other unsupervised competitive learning rules in generating equiprobable quantizations. It is also shown that our rule can do better or worse than the Lloyd I algorithm in minimizing average mean square error, depending on the input distribution. Finally, an application to adaptive non-uniform analog to digital (A/D) conversion is considered.