Summary of the work performed in the second year

Summary of the work performed in the second year

Project BION

G.A.  213219

Summary 15/4/2008-14/4/2010

 

The main objective of the project is the realization of a new, highly innovative technology for the production of functional molecular assemblies that can perform advanced tasks of information processing, involving learning and decision making, and that can be tailored down to the nanoscale. The underlying concept is the assumption that the paradigms of learning (in particular, Hebbian-type learning rules including spike time dependent plasticity, and other forms of learning in biological cognitive systems) can also apply to complex assemblies of electrochemically controlled, conducting polymeric  non linear elements.

The success of this endeavour is expected to open up a totally new, bottom-up technology for the fabrication of information processing hardware capable of adaptation and decision making.

The work was organized in four  workpackages. In the first reporting period work was performed on Wps 1, 2, 3, with a prevalence of time devoted to Wp1; in the second  period work was performed on Wps 2, 3, 4, with a prevalence of time devoted to WP3.  We wish to highlight two important results:

1. A synaptic-like material memristive circuit has been fabricated and found to have learning properties similar to those involving the Cerebral Giant Cell in the pond snail feeding behaviour neural system.

2. A complex, self-assembled microphase separated matrix of Functionalized Gold Nanoparticles (FGNP) and Polyaniline (PANI) based and Polyethene (PEO) based copolymers has been synthesized and fabricated. This complex micro-heterogeneous matrix contains large numbers of micro-memristors and functionalised gold nanoparticles which are the synthetic analogues of synapses and neurons in bio-systems.

 

Here we present a SEM image of the complex matrix after microphase separation.

 

3. Of the other two alternative matrices foreseen in the Project, the fibre matrix has been also completed, and the cavitand-based molecular recognition one is also in the last stages of finalization.

The best of these material matrices will be optimized and trained using bio-inspired protocols and simulations, to solve complex tasks in a manner as close as possible to the cognitive processes in the cerebral cortex.

The work reported here has led to 22 publications in the second reporting period, of which 16 in print or accepted.

WP1   Fabrication and optimization of electrochemically controlled simple networks

The overall objective of WP1 was the optimization of the electrochemically controlled polymeric heterojuction of the conducting polymer PANI and the solid electrolyte Li-doped PEO which is at the heart of the proposed adaptive network we aim to build [1,2]. The two critical parameters here are the efficiency of the control and the stability in the functioning of the device. For the first parameter, most important  is the conductivity ratio between the  “conducting” and “insulating” phases of PANI. Initially such ratio was about 100 [1]. We succeeded to increase this to the target ratio of 1000 [3]. We used differently synthetized PANI and PEO with varying chain length and polydispersity, and we explored different dopants and doping methods for both polymers to increase the  conductivity of PANI and the conductivity ratio.

Specifically PANI doped with dodecyl benzene sulphonic acid (DBSA) was found optimal to increase the conductivity and stability in the device [4].

PEO samples of  different  molecular weight (12,000, 35,000 and 8,000,000 Da) were used as solid electrolytes; best results were obtained with the high molecular weight PEO.

Lithium tetrafluoroborate (LiBF₄, 98 %),  was found the best  dopant of PEO in  improving the device redox cycling without affecting the mobility of Li ions across the heterojunction  in the single  device. However a  mixture of LiClO₄ with DBSA worked best for statistical networks of the device.

Different fabrication protocols were tested with these materials: for PANI, Langmuir-Schaefer (LS) deposition worked best. Gel casting with successive drying was used for the solid electrolyte.

To improve the device stability after repeated cycling we also explored several assembly configurations, using different combinations of the component materials. We found that the most stable device configuration is the sealed one. With this device we improved the stability in functioning after repeated cycles by at least two orders of magnitude of both the single device (node) and a simple network. 

Finally, we fabricated different types of “statistical” matrices, i.e. a hybrid polymeric networks of PANI and PEO, and also functionalized gold nanoparticles. In particular, we fabricated networks containing statistically distributed PEO and PANI fibers [5].The best results so far were obtained with a statistical matrix containing PEO and PANI fibers, formed by vacuum treatment. We also started the work on FGNPs: the gold surface modification, necessary for the interaction with polymer components to assemble 2D and 3D structures, was obtained by the synthesis of thiol capped Au nanoparticles and Au NPs stabilised by aniline capping [6].This matrix  has shown elementary  adaptative behavior.

Finally, our phenomenological modelling of the device function was put on a firm experimental basis by microscopic studies using  synchrotron radiation X-ray fluorescence [7],  reflectometry and grazing incidence diffraction[8].

WP2    Training of the network, and successful application to a biological system

 

The networks and matrices fabricated in WP1 were  tested for simple adaptive behaviour. We started with an eight element deterministic network. This network has demonstrated reproducibly for many cycles  supervised learning, i.e. it changed its output for the same input according to a training voltage procedure.

Applying the improved materials and fabrication procedures developed in WP1, we fabricated several types of statistical matrices.  The best stability and reproducibility was obtained on circuits composed from deterministic elements. Statistical networks, formed by successive vacuum treatment of PEO and PANI solutions, have demonstrated adaptive behaviour. However, the stability of their structure and properties was not high, and more work needs be done on this (see results presented in WP3). In addition, we have fabricated and studied deterministic signal transferring chains composed from 2 or 3 memristors. The results have demonstrated a cascade reinforcement of the chain conductivity with sequential “opening” of each element. The results have also demonstrated the possibility of spontaneous inhibition of the signal pathways with the same elements in appropriate conditions (adequate potential distribution profile). This result is expected to be particularly relevant to the functioning of the complex matrix to be built in WP3.

For the first application to a biological system, biophysical mechanisms underlying learning in the pond snail (Limnea Stagnalis) were studied.  We have found that persistent sodium and potassium channels are the key mechanisms for nonsynaptic learning observed in Conditioned Response [9]. A study of the inferotemporal cortex in sheep yielded results implying  a significant increasing of theta power, but not gamma power inlearning. The developed model [10] is deemed useful to benchmark the adaptive behavior of the fully developed statistical matrix of WP3.      

Another line of work was developing causality analysis tools to tackle experimental data. How to understand complex data sets via reverse engineering approaches is a key issue in computational biology and could be useful for our project as well (to assess how a network is changed before and after learning). We have recently worked on developing such tools [11-13].

One of the most useful tools which we have many years experience on  is the Granger causality analysis. Once we have recordings from the microphase separated materials of WP3, we can apply our methods developed in [14-16] to the signals.

We addressed the problem of finding a model which is biologically plausible, and on the other hand can be implemented in material science. We realized that for a small neuronal network, our team can work out a material version [17]. Two circuits, mimicking homo- and heterosynaptic learning algorithms, were fabricated. The best results were obtained on the circuit corresponding to heterosynaptic learning mechanism. The circuit also corresponds better to the model system developed for the snail. Both the amplitude and DC depolarization were increased about 5 times as the result of learning. This first material synaptic circuit and its performance are described in detail in the report on Deliverable 2.1. However, the generalization of this important result to a large network is hard. Hence the challenging task was to find a neuronal network which is similar to a material network.  We have provided such a theoretical framework in [18-20]. We have also studied  issues related to learning [21-24]. Of particular importance for our work on the synaptic material circuit was [25]. Finally a  paper was submitted discussing the characteristics of our memritors and synaptic material circuits [26].

WP3   Fabrication of the complex statistical matrix and bio-inspired tailoring

The initial activity  concerned mainly the synthesis of FGNPs according to two different protocols and their characterization. In particular:

1. The GNPs functionalised by aniline ligands have been used as reactive monomers in the aniline polymerization to obtain the hybrid Pani-Gold nanoparticles matrix in which the macromolecular chains are linked to the surface of nanoparticles allowing the electronic interaction through a Schottky barrier. Electrical characterization of the material showed  suppression of the conductivity at low applied voltage values, confirming the formation of Schottky junctions.

2. Alkyl thiol capped Au nanoparticles, prepared with tailored diameter distribution and purity, have been used to obtain components for self assembling by modification with cavitand derivatives. We found that best results were obtained for an average diameter of 6nm. Special attention was paid to the purification of these new AuNPs: in particular the removal of tetraoctylammonium bromide surfactant, used in the Au-NPs synthesis, turned out to be pivotal for the subsequent functionalization with cavitands.

3. The first steps of the synthesis of PEO derivatives for self assembling materials and synthesis of polymeric materials or 2-3D  network have been carried out. In particular we prepared  an end-groups functionalized PEO (Me-pyridine-PEO-pyridine-Me) and a /ethylene oxide/styrene block copolymer (PEO-b-PS). Here a most important result was obtained: the sought after microphase separation of the block copolymer – PANI composite was demonstrated with optical and SEM microscopies and by differential scanning calorimetry. Furthermore, the electrical characterization showed the presence in the matrix of adequate numbers of interfaces PANI-PEO, i.e. micro-memristors which allow adaptation, hence learning.

4. Gold nanoparticles (3nm diameter) functionalized on the surface with phosphonate cavitands have been prepared and characterized. These hybrid materials exhibited precise molecular recognition properties toward methylalkylammonium and methylpyridinium guests. Methylpyridinium terminated PEO was connected to the nanoparticles to form reversible networks. The formation/dissociation of these networks, driven by molecular recognition events, was proven [27, 28], and was fully characterized at the molecular level for the PEO-AuNP network. In doing so, we realized that a larger number of molecular recognition events[29,30] between guest-functionalized PEO and host-functionalized AuNP is necessary to achieve the desired microsegregation effects in the alternative matrix. Thus we prepared a cavitand functionalized polystyrene as matrix to interact with methylpyridinium-decorated polymer. We therefore synthetized a phosphonate cavitand monomer, presenting at the lower rim a styrene unit and polymerized it with styrene. The resulting supramolecular copolymer[31] will be a model for testing the microsegregation conditions to be applied to the AuNP-polymer matrix.

            Significant activity was also performed on the fibrillar statistical networks on solid porous and fibrillar templates using the newly synthesized composite materials. The results have demonstrated the improved stability and reproducibility of adaptation properties with respect to those obtained on free-standing fibrillar networks.

            In order to make a direct comparison of the properties of the realized networks with those of nervous systems, two additional lines of the research were established. The first was a mathematical modeling of the organization and properties of deterministic and statistical networks composed of nodes with properties of the organic memristor. For the deterministic case, we chose to study a network with two hidden layers of nodes, one input layer and one output layer. Nodes in each layer would be connected by memristors to every node in the previous and the next layer. The training would be finalized to achieve an input-output pairing network, in which injecting a current into one input or applying a voltage between one input and the output layer would create a strong output current in only one specific output, and all the inputs would be paired to different outputs. The statistical network simulation is described in WP4.

The second line was based on the study of the close connection between optical measurements and conductivity changes in the memristive device. Total internal reflection mode ellipsometry (TIRE), and reflection spectrometry which correlated the microscopic structure with the optical and electronic states of molecular layers of PANI through the doping induced conductivity changes [32]. On this basis we developed  a contactless method of space-temporal mapping of the network conductivity changes [33]. The method was based on the spectroscopic changes associated with training (green form of PANI is conducting, while blue form is insulating). The validity of the method was demonstrated on single elements. We plan to apply it to the mapping of the network during learning. Such maps will be compared with maps of the potential distribution in simulated memristor and neuron networks.

 Since the aim of BION is to imitate features of the nervous system which are crucial for information processing and learning, an important  task was to use  the relevant knowledge on this topic. Basic knowledge on the nervous system, relevant already from WP1 on, was transferred during several meetings (see also [34]) and the features of the  molecular network to be fabricated in WP3 were adapted accordingly. In addition, optimization of neurological data has been carried out  by quantitative studies on the internal  connectivity of the mammalian cerebral cortex [35,36]. These are highly relevant for the formation of the cortex-like matrix envisaged in WP4.

Some of our results have been reviewed for outreach in [37].

 

WP4   Discrimination and learning in  the biologically inspired supramolecular device.

We reviewed some of our results and put them on the international context in [38]

We implemented a platform to simulate and evaluate adaptive properties of stochastic memristor networks. In this platform, network dynamics can be parametrically simulated with the control at different granularity levels, going from a single memristor to large-scale configurations. Using this simulation platform we have shown that memristor networks stimulated with random noise follow a stable (or semistable) behavior that diverges from its initial state depending on the stimulation history. Interestingly, by changing the connectivity patterns (from random to distance-dependent), we observed differences in the dynamics of the network that depend on the type of input used. The results have demonstrated the possibility of adaptations and learning in complex statistical memristor networks.

            Using a new morphing technique, we created realistic three-dimensional morphed faces that linearly span the continuum between humans and monkeys (“species” continuum) [39]. This work will provide a training protocol for the image recognition task in WP4.

            Now that the fully phase separated complex matrix is available to the project, it is possible to pass from the simple electrode configurations used so far to highly complex miniaturized configurations. For this we have developed lithography processes for the formation of microelectrodes arrays and patterning of active channels. The size of elements we are using is within 100 – 10000 nm, for now, but we plan to bring it further down to the nanoscale.

Publications

 

WP1

[1] V. Erokhin and M.P. Fontana, “Electrochemically controlled polymeric device: A memristor (and more) found two years ago”, ArXiv: 0807.033 (2008).

[2] A. Smerieri, T. Berzina, V. Erokhin, and M.P. Fontana, “Polymeric electrochemical element for adaptive networks: Pulse mode”, J. Appl. Phys., 104, 114513 (2008).

[3] T. Berzina, A. Smerieri, M. Bernabò, A. Pucci, G. Ruggeri, V. Erokhin and M.P. Fontana, “Optimization of an organic memristor as an adaptive memory element”, J. Appl. Phys., 105, 124515  (2009).

[4] T. Berzina, A. Smerieri, G. Ruggeri, M. Bernabo’, V. Erokhin, and M.P. Fontana, “Role of the solid electrolyte composition on the performance of organic memristor”, Mater. Sci. Engineer. C, 30, 407-411 (2010).

[5] V. Erokhin, T. Berzina, S. Erokhina, and M.P. Fontana, “Organic memristors and adaptive networks”, in A. Schmid et al, NanoNet 2009 , LNICST 20, pp. 210–221, Springer, (2009).

[6] V. Erokhin, A. Schüz, and M. P. Fontana, “Organic memristor and bio-inspired information processing”, Int. J. Unconventional Computing, 6, 15-32  (2010).

 [7] T. Berzina, S. Erokhina, P. Camorani, O. Konovalov, V. Erokhin, and M.P. Fontana, “Electrochemical control of the conductivity in an organic memristor: A time-resolved X-ray fluorescence study of ionic drift as a function of applied voltage”, ACS Appl. Mater. Interfaces 1, 2115-2118 (2009).

[8] L. Cristofolini, M. P. Fontana, O. Konovalov, T. Berzina,A. Smerieri, “Doping-Induced Conductivity Transition in Molecular Layers of Polyaniline: Detailed Structural Study” Langmuir (Letter) 25, 12429-34 (2009).

 

WP2

[9] ES Nikitin, DV Vavoulis, Feng J.F.,M O’Shea, PR Benjamin & G Kemenes “ Persistent sodium current is a non-synaptic substrate for long-term memory” Current Biology  18, 1221-1226 (2008)

[10] Kendrick KM, Y Zhan, H Fischer, A U Nicol, XJ Zhang, Feng J.F. “Learning alters theta-nested gamma oscillations in inferotemporal cortex” Nature Precedings hdl: 10101/ npre. 2009.3151.1. (2009)

[11] Zou CL, Feng J.F. “Granger causality vs. Dynamic Bayesian network inference: A Comparative Study” BMC Bioinformatics 10, 122 (2009)

[12] Feng J.F. Yi DY, Krishna R, Guo SX, Buchanan-Wollaston V. “ Listen to genes: dealing with microarray data in the frequency domain” PLoS One 4(4), (2009)

[13] Guo SX, Wu J.F.H.,  Ding MZ, Feng J.F.  “Uncovering interactions in the frequency domain” PLoS Comput Biol” 4(5), (2008)

[14] Ge T., Kendrick K., J.F. Feng, “A Unified Dynamic and Granger Causal Model Approach Demonstrates Brain Hemispheric Differences During Face Recognition Learning” PLoS Comp. Biol. 5, 11 (2009)

[15] Zhang XJ, Leng G., J.F. Feng,, (2009). Coherent Peptide-mediated Brain Activities of Neuronal Networks Controlled by Subcellular Signaling Pathway: Experimental and Modelling Results J. Biotechnology 10.1016/j.jbiotec.2010.01.003.

[16] Zou ZL, Ladroue C, Gou SX J.F. Feng,, (2010). Indentifying interactions in the time and freqeuncy domains in local and global networks BMC Bioinformatics (under revision).

[17] Erokhin V, Berzina T, Camorani P, Smerieri A, Vavoulis D, Feng JF , and M.P. Fontana (2010). Material Memristor Circuits with Synaptic Plasticity: Learning and Memory PNAS (under revision).

[18] Lu WL J.F. Feng,, (2010). On Gaussian Random Neuronal Field Model: Moment Neuronal Network Approach IJCNN2010 (in press).

[19] Ge T, Lu WL, J.F. Feng, (2010). Recover Synaptic Topology from Spike Trains IJCNN2010 (in press).

[20] Lu W.L., Rossoni E., J.F. Feng,, (2010). Toward a theory of random neuronal field model NeuroImage doi: 10.1016/j.neuroimage.2010.02.075.

[21] Kang J, Robinson HPC, J.F. Feng, “Minimal mechanism for decoding input temporal frequencies -modelling and experimental approach” PLoS One ,5, 3 (2010)

[22] Kang J, Wu JH, Smerieri A, J.F. Feng, “Weber's Law Implies Sub-Poisson Neural Discharge” Eur. J. Neurosci. ,36, 1006-1018 (2010)

[23] Smerieri A, Rolls ET, J.F. Feng,, (2010). Decision reaction time, slow inhibition, and theta rhythms J. Neurosci.(under revision).

[24] E. Rossoni, J. Kang, Feng J.F.  Controlling precise movement with stochastic signals, Biol. Cybern. 102, 441-450 (2010)

[25] D Vavoulis, ES Nikitin, I Kemenes, V Marra, J.F. Feng,, P. R. Benjamin and G. Kemenes, (2010). Balanced plasticity and stability of the electrical properties of a molluscan modulatory interneuron after classical conditioning: a computational study Frontiers in Behaviour Neuroscience (accepted).

[26]V. Erokhin, M.P. Fontana (2010). Comment on “An Organic Nanoparticle Transistor Behaving as a Biological Spiking Synapse”, Adv. Func. Mater.  (submitted).

 

WP3

[27]  B. Gadenne, M. Semeraro, R. M. Yebeutchou, F. Tancini, L. Pirondini, E. Dalcanale, A. Credi,
“Electrochemically Controlled Formation/Dissociation of Phosphonate Cavitand-Methylpyridinium Complexes” Chem.Eur. J., 14 8964-8971 (2008)

[28]  R. M. Yebeutchou, F. Tancini, N. Demitri, S. Geremia, R. Mendichi, E. Dalcanale, “Host-Guest Driven Self-Assembly of Linear and Star Supramolecular Polymers” Angew. Chem. Int. Ed. Engl., 47, 4504-4508 (2008)

[29] E. Biavardi, M. Favazza, A. Motta, I. L. Fragalà, C. Massera, L. Prodi, M. Montalti, M. Melegari, G. G. Condorelli, E. Dalcanale: Molecular Recognition on a Cavitand-Functionalized Silicon Surface J. Am. Chem. Soc., 131, 7447-7455 (2009)

[30] F. Tancini, D. Genovese, M. Montalti, L. Cristofolini, L. Nasi, L. Prodi, E. Dalcanale: Hierarchical Self-Assembly on Silicon J. Am. Chem. Soc., 132, 4781-4789 (2010)

[31] For an example of cavitand-based supramolecular polymer see: F. Tancini, E. Rampazzo, E. Dalcanale: Interplay between Cyclization and Polymerization in Ditopic Cavitand Monomers Aust. J. Chem,, 63, 646-652 (2010)

  [32]  L. Cristofolini, M.P. Fontana, P. Camorani, T. Berzina, A. Nabok, “Doping-induced conductivity transitions in molecular layers of polyaniline: optical studies of electronic state changes.” Langmuir 26, 5829-35 (2010)

 [33]  Paolo Camorani, Tatiana Berzina, Victor Erokhin, Anteo Smerieri  and M.P. Fontana, Adaptive polymeric system for Hebbian type learning , Phys. Rev. Letters (submitted)

 [34] Schüz, A.: Neuroanatomy. Scholarpedia 3(3), 3158, Scholarpedia.org, San Diego, CA, USA (2008) [Note: Online-Resource]

[35] Voges, N., A. Schüz, A. Aertsen and S. Rotter: A modeler‘s view on the spatial structure of intrinsic horizontal connectivity in the neocortex. Progress in Neurobiology (accepted)

[36] Schüz, A. and F. Sultan: Brain connectivity and brain size. New Encyclopedia of Neuroscience 2, 317-326. (Eds.) Squire, L. R. et al, Elsevier, Amsterdam, Netherlands ( 2009)

[37]V. Erokhin, T. Berzina, A. Smerieri, P. Camorani, S. Erokhina, M.P. Fontana (2010). “Bio-inspired adaptive networks based on organic memristors”, Nano Commun. Networks, (accepted).

 

WP4

[38] V. Erokhin, M.P. Fontana (2010). Thin film electrochemical memristive systems for bio-inspired computing, J. Comput. Theor. Nanosci., submitted revised version.

[39] R. Sigala, J. Schultz, N. K. Logothetis, G. Rainer, (2010) ““Own-species” bias in the categorical representation of a human/monkey continuum in the human and non-human primate temporal lobe” Neuroscience 2010.

 

More information of the project, including references to our published work, news and highlights can be found at our Project internet site http://www.fp7-bion.eu