Neurocomputational Modelling of Tactile Perception for the Development of Artificial Sense of Touch.

Neurocomputational Modelling of Tactile

Perception for the Development of Artificial Sense

of Touch

Thesis submitted in the fulfillment of the requirements for the award of the PhD degree in BioRobotics

June 2017

Candidate

Supervisor

Udaya Bhaskar Rongala

Calogero Maria Oddo

Co - Supervisor

Acknowledgement

Dad, Mom, Sister & Brother, trust you have on me and the support you give me cannot be quantified. Less said, any page of my life book wouldn’t exist without your love and support. The freedom you gave me, let me be mad and take the decisions I made, and achieve what I achieved.

Sara, thanks would be underrated for the support you have been throughout this journey, by bearing me and believing in me more than I do. It’s been a journey, where you filled my soul with lot of life & love. Along with everything, thanks for giving me a beautiful Italian family.

Paoletti’s & Miglioranzi’s, thanks for all the love & support you gave me as a family. Marzia & Marco, you made me feel like a fourth crazy person along with the other three at your home. Guido, you are a true inspiration.

Friends, I would be a wired depressed asocial Indian guy in Italy without anyone of you guys. You literally made Italy home for me.

Calogero M Oddo, thanks for trusting in me and giving the right freedom to explore and gain all the knowledge that I encompass today in this field of research. You guided me through a defined direction whenever I wander with my ideas. Thanks for believing in me and making this PhD possible and what it is today.

Alberto Mazzoni*, thanks for your boundless enthusiasm to share the vast knowl-edge you have. I couldn’t ask for anyone better as a guide.

Henrik J¨orntell, the organic approach you followed in educating and guiding me was literal magic. You have been a patient science guru for me through out the last two years. Working with you made me confident in what I do and what I can do. Thanks for everything.

Federica Radici, I would have quit my PhD in first year without you. I cant imagine handling any bureaucracy without your support. Thanks for fighting along with me and making it possible every year.

Administration, thanks to everyone in administration department, you never made me doubt how will I pass next month with my finances.

Neurocomputational Modelling of Tactile

Perception for the Development of Artificial Sense

of Touch

Abstract

In this thesis, we developed and validated an artificial tactile system by adopting a bio-robotic approach to systematically merge engineering of artificial touch and neurophysiology of the human sense.

This objective was achieved in three steps. First, we proved that a biomimetic fin-gertip implementing neuromorphic representation of tactile stimuli is able to achieve excellent performance in decoding daily life tactile stimuli under varying sensing con-ditions. Second, we developed a neuronal model emulating the learning processes of the Cuneate Nucleus, the second stage of the human somatosensory pathways. Finally, we integrated these two elements to create a functional biomimetic system towards artificial touch.

By means of this approach we achieved three results. We contributed to the devel-opment of more efficient upper limb neuroprostheses by characterizing the properties of the biomimetic fingertip. We increased our knowledge about the biological system under investigation: in neuroscience, the mechanisms of different regions are often studied in isolation, and therefore their functions are not fully captured. Our inte-grated biorobotic system merging low-level tactile sensations with high-level learning, sheds new light on the overall mechanism of sensory processing. Finally, our method-ology indicates a novel mode of assumption-free information representation by the brain, which can be exploited to develop robust and effective autonomous sensing systems able to learn feature extraction in the field of biorobotics.

Contents

Rationale I

Part A: Science of Touch 1

1 Neurophysiology of Tactile sensing 3

1.1 The Somatosensory System: A classic understanding . . . 3

1.2 Skin and its sensors . . . 4

1.2.1 Mechanoreceptors . . . 5

1.3 Dorsal root ganglion neurons . . . 8

1.4 Cuneate Nucleus . . . 9

1.4.1 Excitatory Synapses . . . 10

1.4.2 Inhibitory Synapses . . . 11

1.4.3 Intrinsic properties and responses . . . 11

1.4.4 Thalamus and Somatosensory Cortex . . . 12

2 Tactile Information Processing 13 2.1 Connectivity . . . 14

2.2 Learning . . . 15

2.2.1 Synaptic Learning Mechanism . . . 16

Part B: Engineering of Touch 21 3 Neuromorphic Tactile Receptors 23 3.1 Bio-inspired Tactile Sensor . . . 23

3.3 Neuromorphic Tactile System . . . 26

3.3.1 Classification Based on Spike Responses . . . 27

3.3.2 Effect of Sensing Dynamics . . . 28

3.3.3 Discussion . . . 29

4 Modelling Cuneate Neurons 31 4.1 Cuneate Neurons as Coincidence Detectors . . . 32

4.1.1 Coincidence Detector . . . 32

4.1.2 Modelling Coincidence Detector . . . 33

4.1.3 Encoding Orientation . . . 34

4.1.4 Discussion . . . 35

4.2 Cuneate Neurons as Feature Extractor . . . 36

4.2.1 Cuneate Neuron Model . . . 36

4.2.2 Network Connectivity . . . 40

4.2.3 Effect of Synaptic Learning . . . 40

Conclusion and Perspectives 45

List of Publications 49

Bibliography 51

Article 1 57

Article 2 69

Rationale

Touch is one of the most crucial senses, whose understanding in humans and imple-mentation in robotics remains a challenge [Johansson and Flanagan, 2009, Bartolozzi et al., 2016]. The sense of touch has a broad range of applications ranging from industry (e.g., robotics and factory automation) to health care (e.g., prosthetics and surgical robotics). Tactile perception plays an important role in revealing the non-visual properties of an object where the same task cannot be accomplished with vision or requires heavy machine learning computations [Hoelscher et al., 2015]. Adopting a biorobotics approach [Dario et al., 2008] to systematically merge engineering of ar-tificial touch and neurophysiology of human touch shows great promise and mutual benefits for both robotics and neuroscience [Service, 2014].

In this line, a neuromorphic approach of integrating the hardware sensory system along with spiking neuron models can help us in recreating the natural information coding observed with electrophysiological human recordings (e.g., with microneurog-raphy [Oddo et al., 2011]). This integrative paradigm can be advantageous in various application scenarios such as dexterous manipulation, assistive, human-augmentation, rehabilitation and industrial robotics. Our main interests lie 1) in Neuroprostheses; Integrating the sense of touch in prosthetic devices will enable the perception of the prosthesis as a part of the own body, increasing confidence and dexterity, and decreas-ing the need for constant visual feedback and cognitive effort in control [Tan et al., 2014]. 2) in Neuro-robotics: Even though autonomous robots mainly rely on some form of visual perception to interact with the surrounding environment, there are tasks that would be impossible or too complicated without the sense of touch. Touch helps robots in interactions with unstructured environments, and aids in exploring

and localizing objects with great accuracy [Cramphorn et al., 2017].

Therefore, in the course of understanding the way by which brain processes in-formation, we investigate the existing neurophysiological knowledge for touch and contribute with neurocomputational models that converge into the proposed neuro-robotic implementation.

∗ ∗ ∗

The Brain fascinates with its immense capabilities to process, understand, learn, and memorize an extraordinary amount of information throughout the phases of human life. Recent advances in field of biology & neuroscience shed more light on the functionality of brain and its sub-modalities. Fields of Robotics and Neuro-Engineering are gaining attention, with a vision to render the functional and efficient capabilities of the brain and its sub-systems to create synthetic intelligence in robotics; en route towards self-learning robots.

The human sensory system has undergone a long and complex evolutionary pro-cess, to reach the current robust estimation of sensory experiences. The known five senses - vision, hearing, touch, taste and smell, each linked to specific sense organ in the body: the eyes, ears, skin and tongue. The frequent non-deterministic human interactions with the dynamic world, and the abundance in sensor spread across the human body with varying sensing competence, provide the brain with rich and unique sensory information. In the field of Neuroscience numerous hypotheses are proposed to interpret the role of sensory information on the nervous system; and, coherently, the same limitations apply to the fundamental knowledge of the techniques by which the brain processes these input clues.

The sense of touch is the main focus in this thesis. As all the senses in humans, touch sensations also face the critical challenge of extracting, representing and mem-orizing meaningful information from abundant afferent input. The tactile sensation maps different perceptual dimensions such as force, roughness and slipperiness of the physical stimuli from skin across the body. There are ∼ 17, 000 sensory afferents innervating just the hand skin area, which makes the sensor processing in humans complex and robust.

Thesis Breakdown

In this thesis, we thrive to gain better understanding of the tactile sensing modalities, with an ambition of implementing realistic tactile feedback in robotics and artificial hands. We tackle this objective by combining computational neuroscience and neuro-robotics.

In Chapter 1, We have elaborated the existing neurophysiology knowledge about the afferents and neurons responsible for processing tactile information in humans. In sections 1.2 ∼ 1.3, we discuss the generic knowledge that comes from classical understanding of somatosensory pathways. In section 1.4, we illustrate the state-of-art neuro-physiological understanding of Cuneate Neurons, and demonstrate some original biological results supporting the claims.

In Chapter 2, based on the body of knowledge available in neurophysiological liter-ature, we discuss about the tactile information content and the strategies by which it is processed along the somatosensory pathways.

In Chapter 3, we discuss the approach by which we developed artificial tactile sys-tems, and the methods that are adapted to mimic the functionality of sensory affer-ents that are present in human fingertips. In section 3.3 we illustrate the results from the neuromorphic approach that we adapted and further validated the yield in our neuromorphic sensory information, using offline decoding methods.

In Chapter 4, we illustrate the methodologies that we adopted to construct a second order neuronal model and network, in order to move from software-based off-line decoding towards an architecture-based on-line implementation. We adopted two methodologies; in section 4.1 we demonstrated the outcome of building a second order layer as coincidence detector based on the precise spike timing hypothesis. In section 4.2, we build a second order layer of neurons, based on the neurophysiology that was observed during in vivo recordings in cats.

In Conclusions we give an overall idea of the results across all the chapters and their integration as a functional system and discuss the future perspective.

Part A

Science of Touch

Chapter 1

Neurophysiology of Tactile sensing

Merely to touch and perceive an object, brain requires an equivalent of humongous computational achievements. These feats are accomplished with the help of its nerve cells (Neurons)1 _{that are wired together in a specific manner, based on early life} experiences. The nerve cells are networked in a hierarchical fashion throughout the brain (Fig. 1), handling information representation across several nuclei and cortical regions. In this chapter, we summarize the elements of the somatosensory system that constitute for tactile signal processing.

1.1

The Somatosensory System: A classic

under-standing

The first step of tactile information processing in humans involves the activation of primary sensory neurons (Primary Afferents (PAs)) present in the finger tip. The cell bodies of these neurons reside within dorsal root ganglia (DRG). These are sensory neurons whose axons on one end associate with the PAs on periphery, and another end of the axons penetrates the spinal cord and forms synapses2 upon the cuneate nucleus (that encompasses second-order neurons) in the dorsal column nuclei of the brainstem [Brown, 1981, Fyffe et al., 1986, Cliffer and Willis, 1994].

1 _{Neuron, is an electrically excitable cell that processes and transmits information through}

elec-trical and chemical signals. (freely excerpt from [Llinas, 2008])

2 _{Synapse, is the site where the axon of a pre-synaptic neuron makes contact with the dendrite}

Figure 1: The major tactile afferent pathway for somatosensory information process-ing (figure from [Kandel et al., 2014]).

Further, these afferent information are fed upon the thalamus and a hierarchy of regions in the cerebellar cortex. These higher cortical regions are extensively inter-connected, allowing heaps of multi-sensory data integration, which in-turn leads to a meaningful representation of the physical world in our brain.

1.2

Skin and its sensors

The Human body is covered by two types of skin: glabrous and hairy. Both convey tactile information of external physical world to the brain. The skin of the human hand is innervated with various tactile afferent3 _{and nerve endings (Fig. 2) that signal} unique characteristics, when this soft tissue of the hand interacts with objects. As per our objective of understanding the tactile sensation from the skin, in this thesis we only focus on the tactile afferents present in the glabrous skin (e.g., in the fingertip). The touch sensation in glabrous skin is meditated by four types of mechanoreceptors4 _{(Fig. 2), that are devised into two groups: Fast Adapting (FA)}

3 _{“Fast-conducting myelinated afferent neurons that convey signals to the brain from}

mechanorecep-tors in body areas that actively contact objects - like, inside of hand, the sole of the foot, the lips the tongue and the oral mucosa” [Johansson and Flanagan, 2009].

4 _{“A mechanoreceptor is a sensory receptor that responds to mechanical pressure or distortion”}

[Johansson and Flanagan, 2009].

Figure 2: The cutaneous mechanoreceptors in the skin. A) In glabrous skin, innocuous touch is mediated by four types of mechanoreceptors and nerve endings. B) In hairy skin, tactile stimuli are transduced through three types of hair follicles, defined in the mouse as guard, awl/auchenne, and zigzag (figure from [Abraira and Ginty, 2013]).

receptors and Slowly Adapting (SA) receptors [Brown, 1981, Johansson and Vallbo, 1983, Vallbo and Johansson, 1984, Johansson and Flanagan, 2009, Abraira and Ginty, 2013].

1.2.1

Mechanoreceptors

The population of mechanoreceptor units responding to a skin deformation serves to the skin sensibility in humans. Mechanoreceptors exhibit varied responses depending on its morphology, innervation pattern and depth in the skin. The glabrous skin of a human hand is innervated approximately by 17000 mechanoreceptor units. About 44% of these units are Slowly Adapting (SA), i.e. they exhibit sustained output throughout the sensing activity reflecting stimuli information. The remaining 56% units are Fast Adapting (FA), that respond with a burst of pulses at the onsets5 of the sensing activity [Johansson and Vallbo, 1983], eventually being responsive to transient events during manipulation.

The SA and FA mechanoreceptor units can be further classified into two different types based on their receptive field properties. The fast-adapting type I (FA-I, Meiss-ner endings) units and the slow-adapting type I (SA-I, Merkel endings) units have

5 _{In the context of sensory information during task manipulation, onset is referred as the instance}

Figure 3: Mechanoreceptors arrangement. A) Positioning: superficial and deep layers of the glabrous (hairless) skin of the hand. B) Receptive field: reflecting the location and distribution of its terminals in the skin. Touch receptors in the superficial layers of the skin have smaller receptive fields than those in the deep layers (figure from [Kandel et al., 2014]).

constrained and well-defined receptive fields, whereas the fast-adapting type II (FA-II, Pacini ending) units and the slow-adapting type II (SA-II, Ruffini-like endings) units have wider receptive fields with obscure borders (Fig. 3B). The receptive field of these receptors can be proportional to their positioning in the skin layers. However, there are a number of additional distinguishing parameters which were in depth studied by various groups [Johansson and Vallbo, 1979, Johansson and Vallbo, 1983, Johansson and Flanagan, 2009, Abraira and Ginty, 2013] for half a century. Some key differ-ences between these afferent types are briefed in Table 1.1 (based on the studies from [Johansson and Vallbo, 1979, Vallbo and Johansson, 1984, L¨ofvenberg and Jo-hansson, 1984,Lundstrom, 1986,Westling and JoJo-hansson, 1987,Knibest¨ol and Vallbo, 1970, Knibest¨ol, 1975, Johansson, 1978, Johansson and Vallbo, 1976]).

Receptor type Adaptation Density on finger tip (units/cm2) Response sensitivity SA-I Slow, static response 70 Low frequency skin deformation (<∼5 Hz) FA-I Fast, no static response 120 High frequency skin deformation (∼5-50 Hz) SA-II Slow,

static response 10 Skin stretches

FA-II Fast,

no static response 20

High frequency vibration (∼40-400 Hz) Table 1.1: Receptor key properties.

Sensory Signals

Figure 4: Illustration of an action potential, and active ion channels that comprehend for it.

Mechanoreceptors, alike the other neurons, inter-act through Action Potentials (APs) also termed Spikes (discrete events of APs, Fig. 5). These are electrical signals generated as a result of chemical reaction within the cell membrane of individual neurons. Briefly, an electrical pulse is generated within a neuron due to the disruption of original balance between the sodium (Na+) and potas-sium (K+) concentration within the cell mem-brane [Hodgkin and Huxley, 1952]. Such unbal-ances are typically caused by an external events, such as sensory stimulus (example, mechanoreceptors).

Sensory information coding

Each class of mechanoreceptor sensory neuron transforms stimulus energy into elec-trical signals, that are encoded as spike trains6 _{(Section 1.2.1, Fig. 4). The rate of} adaptation and frequencies of theses spike trains change between each class of recep-tor, depending on their individual responsive properties [Johansson and Flanagan, 2009] (Table 1.1). Figure 5 illustrates spike responses for each type of nerve when its

receptor is activated by constant pressure against the skin. We can examine that FA receptors adapt rapidly for constant stimulation, while the SA type receptors adapt slowly. Specific behavioural responses of individual mechanoreceptor suggest a coarse information content.

Repeating the same manipulation task twice, during which the physical interaction between skin and its surrounding world could be very different from one repetition to another (due to variation in force of touch, precise positioning, external disturbance, etc.), would result in slightly varied mechanoreceptor responses for each individual interaction. This coarse behaviour in individual receptors contribute for a robust information when considered as a population.

Figure 5: Afferent Coding; Illustration of the responses from each receptor against constant pressure.

1.3

Dorsal root ganglion neurons

Dorsal Root Ganglia (DRG) are the first order neurons, that encompass cell bodies of mechanoreceptors. DRG are responsible for transduction and encoding of stimuli (when mechanoreceptors experience physical interaction) into electrical signals and the transmission of those signals to the central nervous system (Fig. 6). These are pseudo-unipolar neurons7_{, where one branch projects on peripheral sensors, and} an-other branch tunnels through the spinal cord and forms synaptic connections on dorsal column nuclei (that encompass Cuneate Neurons) present in the brainstem.

7 _{This neuron contains an axon that has split into two branches.}

Figure 6: Illustration of DRG sensory neuron projections (figure from [Kandel et al., 2014]).

1.4

Cuneate Nucleus

Cuneate nucleus is a part of the dorsal column nuclei of the lower brain-stem. The tac-tile afferent information from mechanoreceptors is processed by these cuneate nucleus neurons before being transmitted to the brain networks (thalamus and somatosensory cortex) [Fyffe et al., 1986,Cliffer and Willis, 1994]. The main cell types of the cuneate nucleus are the projection neurons, which receive direct synaptic input from tactile afferents and from multiple local inhibitory interneurons (Fig. 7). The projections neurons only innervate targets outside the nucleus, implying no recurrent connection occur within the cuneate neurons.

In the previous decades these Cuneate Neurons (CNs) were assumed to be relay neu-rons, that act as a link between individual primary afferents and thalamus [Andersen et al., 1964]; hinting a redundant functionality. However, recent neurophysiology studies suggested that these neurons exhibit strong synaptic plasticity [Nu˜nez and Bu˜no, 2001, Bengtsson et al., 2013], which advocates a major role of these neurons in tactile sensory information processing [Jones, 2000, Bengtsson et al., 2013, J¨orntell et al., 2014]. Based on anatomic studies, it is know that 100s of PAs make synaptic contact with each single CN [Jones, 2000], and on other hand more than 1000 CNs receive input from a single primary afferent [Weinberg et al., 1990]. This conver-gence and diverconver-gence of information between these two neuronal layers, along with domination of few excitatory synapses on CNs suggests their role in maximizing the

Figure 7: Schematic of cuneate nucleus neuronal circuitry. Synaptic weights are indicated by the size of the triangles (green lines) or as a stop ending for near-zero weights. Inhibitory interneurons (red for highlighted neuron, the others are grey) make inhibitory synapses on the projection neuron (black triangles) (figure from Article 3).

information transfer in somatosensory pathways [Bengtsson et al., 2013].

The function of a neuron, depends mainly on the synaptic input it receive and its intrinsic membrane physiology. The in vivo study in cats by Bengtsson et al. [Bengts-son et al., 2013], characterize the excitatory postsynaptic potential (EPSPs) in CNs based on the primary afferent stimulation. To the best of our knowledge, we have characterized inhibitory postsynaptic potentials (IPSPs) with our study presented in Article 3 .

1.4.1

Excitatory Synapses

All the PAs make excitatory synapses with different weights (efficacy) on the CNs (Fig. 7, as indicated by green triangles & lines. Lines with triangles indicating higher weight, whereas other lines indicating zero weight). It is shown from in vivo recordings of adult cats [Bengtsson et al., 2013] that only few (4-8) primary afferents have large excitatory synaptic weight and the rest with zero weight. This phenomenon indicates the dominance of certain PA inputs on an individual CN. In detail characterization of unitary EPSPs can be found in the study by Bengtsson et al. [Bengtsson et al., 2013].

1.4.2

Inhibitory Synapses

The projections neurons receive inhibitory synapses from a large number of interneu-rons, which by themselves are driven by primary afferent synaptic inputs (Fig. 7, red color interneurons). The innervation from inhibitory interneurons to projection neurons are assumed to follow the principle of nearness, i.e. the closer the inhibitory interneuron is to a particular projection neuron, the higher the probability that it will be synaptically connected to the projection neuron. Stimulation of a single skin area will thereby activate many such interneurons and their unitary IPSPs will summate to generate a larger compound IPSP amplitude. The higher the skin stimulating inten-sity, the higher the number of skin sensors and thereby interneurons being activated, and larger the compound IPSP (Fig. 8). Inhibitory synaptic weights are expected to be equal weights. In detail characterization of IPSPs is presented in Article 3 .

Figure 8: Compound IPSP. The gradual recruitment of summated, or compound, IPSPs with increased electrical stimulation intensity to a skin area adjacent to the excitatory receptive skin area of a sample projection neuron (figure from Article 3).

1.4.3

Intrinsic properties and responses

Unlike most of the spiking neurons, CNs exhibit high frequency responses (upto 1000 Hz, averaging ∼200 Hz) even in the absence of any strong excitatory drive (Fig. 9). They often generate triplets, doublets of spike bursts followed by deep afterhyperpo-larization. From Figure 9B, it is observed that high frequency bursts can be generated in a controlled fashion by simply keeping the membrane potential low for a couple of 100 ms (by injecting a hyperpolarizing current in the recorded neuron) and then

suddenly releasing that hyperpolarization.

The presence of voltage-gated calcium channels [Reboreda et al., 2003] and rebound excitation mechanisms [Alvi˜na et al., 2009] could result CNs to generate this kind of high frequency spike responses. Whereas, the deep afterhyperpolarization indicated the presence of calcium-activated potassium channels.

Figure 9: Cuneate neuron responses. A) Intracellular responses for a gentle stroke on paw of cat (data provided by Henrik J¨orntell). B) Simulated responses of projection neuron for a current step input (modelled by Anton Spanne) (figure from Article 3).

1.4.4

Thalamus and Somatosensory Cortex

Tactile information from the dorsal column and trigeminal nuclei enters the lateral and medial ventral posterior nuclei of the thalamus. The thalamus is a structure found between the mid-brain and cerebral cortex. It has a crucial function of relaying sensory any motor signals (from all body parts) around different structures of the brain. Though the earlier studies found thalamus to be a relay station for transport-ing the information to the cortex. Recent evidences suggest that a significant number of fibres come to the thalamus from the cortex, implying bi-directional information flow between these two regions. Therefore suggesting that thalamus could play an important role in cortical computations [Sherman and Guillery, 2002, Barardi et al., 2016]. The primary somatosensory cortex (SI) is a cortical region, that is responsi-ble for meditating the complex tactile information from across the body parts. SI inputs are derived form mechanoreceptor (via thalamus) and proprioceptors. It can be represented as a multimodal processor, that is able to process wide-ranging set of tactile perturbations (e.g., temperature, pain, itch, light touch, and joint position) in parallel-streams [Gottfried, 2011].

Chapter 2

Tactile Information Processing

In humans, the skin tactile sensory system alone contributes a wealth of sensory data to the brain. Given that each tactile afferent in our hand has a unique func-tionality, defends each of them to be considered as a unique dimension [Spanne and J¨orntell, 2013], therefore subsidizing approximately 17000’s input dimensions from the hand alone. Therefore, it is very crucial to realize the paradigm by which the brain could possibly extract rich information of this input data efficiently. Towards creating this understanding, we first studied how the tactile information is being pro-cessed in various parts of the somatosensory pathways. By realizing these strategies, we can model the gaps between generic neurophysiology understanding of low-level mechanisms (receptor level) and high-level mechanisms (learning and representation). Further integrating these models with the available electronic sensory system would result in a functional artificial tactile system.

In this chapter, first, we described the connectivity between a population of primary afferents and Cuneate Neurons (CNs) in section 2.1. second, we discuss the effects of network hierarchy and synaptic learning in CNs (section 2.2). Finally, we illustrate a basic paradigm of synaptic learning mechanism along with the synaptic plasticity approach that we adapted in our learning model (section 2.2.1).

2.1

Connectivity

Connectivity between the PA sensors and the CNs can give us a holistic idea about the information spread when considering a population of sensors. Network hierarchy along with converging (or) diverging connections between neurons (sub-networks and the whole brain network) enable brain to process information efficiently, by therefore supporting complex brain functions [Park and Friston, 2013]. From previous neu-roanatomy studies [Weinberg et al., 1990], we know that ≈ 1000 PA synapses project on each individual CN. This divergence (Fig. 10A) of PAs leads to integration of mul-tisensory inputs, that can be efficient for information exchange, robustness, adaptabil-ity, resilience and divergent functionalists within a fixed structure/network [Sporns and Zwi, 2004]. For sense of touch, this structural connectivity pattern of many-to-one (or) many-to-one-to-many can also lead to redundancy in sensory information (Fig. 10B); Whereas this phenomenon could be crucial in representation of multiple sensors (can also be cross digit representation, as shown in Fig. 10B) along with other advantages. Studies from Shams et. al [Shams and Seitz, 2008] demonstrated that multisensory (w.r.t different sensing modules) training can be more effective than similar unisen-sory training. This same paradigms could be applied in tactile information processing across the sensor population.

Figure 10: Branching between two neuronal layers. A) Divergence of PAs onto CNs through synaptic connections. B) Abstract representation of respective receptive fields of PA on CN.

Without a learning paradigm, every PA response can effect every CN that they project

on, leading the information in pathway to be affected by every perturbation or event that mechanoreceptors experience. This could also result in huge computationally burden on the brain. Therefore, learning within connectivity is crucial in this pathway for information processing. The learning methodologies are discussed in following section 2.2.

2.2

Learning

The effect PA has on CN depends on its synaptic weight (as illustrated in Fig. 11A). Therefore, learning of synaptic weights could help in attaining information specificity along the somatosensory pathway. The synaptic learning can also reduce the compu-tational burden alongside with robust representation of distinct input haptic features from the tactile perception. Indeed such phenomenon of synaptic weight distribution (Fig. 11B) is observed in adult cats, where each CN comprised 4 − 8 high weight excitatory synapses out of 1000s of PA projections [Bengtsson et al., 2013], further resulting in segregation of tactile inputs in low level functional networks [J¨orntell et al., 2014].

Figure 11: Specificity in synaptic learning. A) Sketch of postsynaptic responses (Post-Syn) for a presynaptic input spike train (Pre-(Post-Syn), depending on the size (weight ) of the synapse. B) Sketch illustrating the projection of number of PA synapses onto a single CN, where most of the excitatory synapses are low and only few are high. In this sketch we also illustrate that all inhibitory synapses fall in a certain weight range as per the (see Section 1.4.2).

they exhibit “motor-babbling”1_{, which is believed to play crucial role for the} nor-mal development of central nervous system [Forssberg et al., 1992, Petersson et al., 2003]. During this unstructured movements, synaptic learning could help in creating functional connectivity between the neuronal layers using various Synaptic Learning Mechanism to recognize statistical dependencies. This in turn benefits the network in reducing enormous dimensions of sensory information to few useful/meaningful dimensions.

2.2.1

Synaptic Learning Mechanism

Synaptic weight is a constant parameter that characterizes a synapse by determining the amplitude of post-synaptic response to an incoming action potential (Fig. 11A). In other terms, it can be defined as a gain constant that acts alike an amplifier to the incoming signal.

Synaptic plasticity is the competence of a synapse to strengthen or weaken (change in weight ) over time, depending on the activity that drives its learning paradigm. Based on the recent patterns of activity, if the learning paradigm leads to a persistent increase of synaptic transmission/strength, the effect is called Long-term potentiation (LTP) of synapse. If the result is a decrease of the synaptic transmission then it is long-term depression (LTD) [Gerstner and Kistler, 2002]. The procedure of adjusting the synaptic weights is referred to as a learning rule. There are different synaptic learning rules defined in last decades, in this chapter first we refer to a classic Hebbian learning (Spike Time Dependent Plasticity), in-order to lay a brief understanding of the synaptic weight changes during the development phase of a neural system. Many classes of learning rules fall under Hebb’s principle; one such rule similar to our approach is “synaptic changes are driven by correlated activity of pre- and post-synaptic neurons”. In the later part of this section, we discuss a learning rule that we have adapted (Activity Intensity Dependent Plasticity) in this thesis.

1 _{Random movements of body parts (Ex. random movements of hands and legs).}

Spike Time Dependent Plasticity(STDP, a basic Hebbian paradigm)

Based on the biological study in the hippocampus by Sarvey et al. [Sarvey et al., 1989] and Bliss et al. [Bliss and Collingridge, 1993] we express the basic paradigm of LTP. First, an individual pulse/spike is applied to the presynaptic fiber which results in a small response (Fig. 12A). Second, a series of spikes are presented to evoke a strong postsynaptic potential which results in an action potential (Fig. 12B). Third, again the small pulse is repeated for the same synapse as in Fig. 12A, which results in a significantly increased amplitude of postsynaptic potential (Fig. 12C). This change in synaptic strength persists for many hours indicating the long-term potentiation.

Figure 12: Synaptic Plasticity. A) A single presynaptic spike (left ) evokes the post-synaptic response (right ) B) A train of prepost-synaptic spikes evoking a strong action potential. C) Repeating a single presynaptic spike, to see that it evokes a bigger postsynaptic response compared to before.

When we consider a scenario with multiple neurons and synapses, we need to con-sider the temporal aspects of the presynaptic activity and postsynaptic activity. For this reason, spike-time dependent plasticity (STDP ) is considered; where synaptic weight change(∆w) is defined as function of spike-time difference (tf) between the presynaptic spike (tP) and post-synaptic spike (tO), as shown in Figure 13.

Figure 13: STDP mechanism. A) A single presynaptic spike (left ) on individual synapse, evokes a postsynaptic response (right ). B) A set of neurons receiving presy-naptic spikes withing a time window, evoking a strong postsypresy-naptic action potential. C) Spike time dependent plasticity rule, where weight change (∆w) dependent on the time of pre- and postsynaptic spikes arrival (tf = tP − tO). D) Resulting LTP and LTD for specific synapses based on the STDP rule and the temporal spike input as shown in (B).

As illustrated in Figure 13A, multiple primary afferents project on a single neuron with varying synaptic weights. Presynaptic input across different synapses is consid-ered on a time scale, where consecutive input spikes result an action potential in the neuron (Fig. 13B). From Figure 13B we can observe that input from tP 1, tP 2 and tP 3 result in the postsynaptic spike (tO). Therefore, these synapses that responsible for a postsynaptic spike are potentiated (Fig. 13C). Whereas, tP 4 arrives soon after the postsynaptic spike (during the period of afterspike-hypoerpolarization, also consid-ered as synaptic fatigue) resulting in depression of that particular synapse (Fig. 13C). This paradigm for potentiation helps the neuron to select synapses that signify a cor-relation leading to a postsynaptic spike. Whereas, depression of particular synapse that doesn’t carry significance, helps the neuron to filter noise and spontaneous ac-tivity. The strength of potentiation and depression of a synapse is defined as function of spike-time difference (tf) between the presynaptic spike and post-synaptic spike (Fig. 13D).

Activity Intensity Dependent Plasticity (AIDP )

In the synaptic learning mechanism that we adapted, we assumed that potentiation or depression of a synapse was dependent on the coincidence between the individ-ual synaptic activity and total activity in dendrites (summed activity across all the synapses). This paradigm helps in isolating individual synaptic activity and improv-ing the segregation of individual inputs, leadimprov-ing towards a skewed synaptic output (few high weight synapses, Fig. 11B). In our model the synaptic weight learning is driven by the calcium activity in the main compartment of the cuneate neurons with combination to the calcium activity in the individual synapse (synaptic space). Towards this goal we have considered compartmentalization of individual synapses into local synaptic compartment and main dendrite compartment (Fig. 14A, B). Therefore, when the local compartment calcium (Fig. 14C) highly coincides with the main compartment calcium (Fig. 14D), it leads to a chemical process of inserting mode ion channels to that particular synapse leading to LTP of the synapse. Whereas, if the synaptic compartment and main compartment calcium doesn’t highly coincide (this is resulted when only a specific synapse has activity irrespective of any other synapses) lead to removal of ion channels resulting in the LTD of that synapse (Fig. 14E). But because the reaction needed to change the number of ionotropic synaptic channels in the synaptic cleft is a complex process, learning signals take a comparatively long time to take effect. Hence, learning signals are integrated over long time windows, where contradictory signals tend to cancel out (according to linear summation). Therefore we decide a synapse to be potentiated or depressed based on the integral trend over a time window (Fig. 14E). The amount of ion channel insertion and removal depends on the existing number of ion channels for each synapse, which we consider as a sigmoid function in our model (Fig. 14F).

Figure 14: Activity Intensity Dependent Plasticity mechanism. (A) Graphical repre-sentation of the structure of PAs projecting on second order neuron dendrites through synapses. (B) Illustration of synaptic and sub-synaptic compartmentalization. Green and red inset reflect the area of the network structure in (A). (C) The local calcium activity of a synapse (synapse 1 on left and synapse 2 on right ), that defines the eligibility for plasticity for each PA synapse (calcium activity in synaptic space from (B)). (D) The main compartment calcium activity varies over time due to synaptic input and its relationship to the learning threshold defines when the cell is in the positive zone (i.e. potentiation mode) or in the negative zone (depression mode). (E) The net learning drive for the PA synapse varied depending on the temporal correla-tion between the zero-offset main compartment calcium activity and the local calcium activity. The integral of the net learning drive defines if that particular synapse is potentiated or depressed. (F) The synaptic weight compensation constant multiplied with the integral net learning drive to calculate the final weight change of the specific synapse (figure from Article 3).

Part B

Engineering of Touch

Chapter 3

Neuromorphic Tactile Receptors

In the quest towards bio-inspired artificial touch, we integrate the neuronal archi-tectures (Part A of this thesis) with state-of-art electromechanical tactile sensory systems (further presented in Part B). In the first part of this chapter (section 3.1 & 3.2), we present the bio-inspired tactile sensor followed by illustration of our neuro-morphic approach. In second part (section 3.3), we present the results of the valida-tion of our neuromorphic approach, by encoding naturalistic textures under different sensing conditions (25 combinations of sliding velocity and contact force conditions).1

3.1

Bio-inspired Tactile Sensor

The research strategy that we present in this study capitalizes on the availability of a Micro Electro-Mechanical System (MEMS) piezoresistive tactile sensor (1.5mm × 1.5mm × 0.625mm) developed via silicon micro-fabrication technologies [Beccai et al., 2005]. Each sensor had four piezoresistors implanted at the roots of a cross shape, with a mesa-structure in the center acting as a force catalyst towards the transducers (sensory arrangement reflecting a joy-stick like design). This sensor proved to deliver precise and robust performance in coding the normal and shear forces [Oddo et al., 2007]. A 2 x 2 array of these sensors are packed together (with a 2.36mm pitch) to form 16 channel biomimetic fingertip sensor (as illustrated in Fig. 15), with a

1 _{Parts of this chapter are freely excerpt from Article 1 [Rongala et al., 2017] and Article 2 [Rongala} et al., 2015], for which the PhD candidate was the first author.

sampling rate of 380 Hz per channel. The whole electronics is capped with compliant polymeric material (DragonSkin 10, Smooth-on, USA) that allows in re-creating skin like dynamic sensing when the finger comes in contact with the physical world. This gives us an extra possibility to explore the after sensing dynamics when the sensor break its contact with the stimuli.

Figure 15: Bio-inspired tactile sensor rendering.

3.2

Neuromorphic Approach

Neuromorphic systems can emulate spike-based neuronal mechanisms that are observed in natural effectors and senses (as illustrated in Fig. 5), and thus have a potential application scenario in neuro-robotics and neuroprostheses. In our work, we developed a neuromorphic artificial tactile system (Fig. 16A) by injecting the output of the bio-inspired tactile sensor into adaptive spiking neuronal models (Izhikevich models [Izhikevich, 2003]). These neuronal model parameters are tuned in order to recreate Slowly Adapting like responses. We further mimicked Fast Adapting mechanoreceptor like responses by feeding derivative sensor input to neuron model.

The sensor raw output data was preprocessed to provide polarized input to the ar-tificial neuron. We normalized the sensor output based on the difference between individual opponent channels data [Spigler et al., 2012, Oddo et al., 2007], to focus on the shear forces arising at stimulus - finger interface as a result of the tangential sliding motion, exploiting the directional selectivity of coupled sensor outputs. The outcome voltage was then multiplied by a gain factor (GF) and injected as the input current Iinput to an Izhikevich neuron model [Izhikevich, 2003].

The Izhikevich model was chosen in order to reproduce adaptation dynamics, that

is a crucial characteristic observed in mechanoreceptors responses [Johansson and Flanagan, 2009]. According to Izhikevich model, the membrane potential v and the adaptation variable u were updated through the following nonlinear differential equations discretized using Euler method:

˙v = Av2 + Bv + C − u +Iinput∗ GF Cm

˙u = a(bv − u)

When the membrane potential reached the spike threshold of 30mV , one spike was produced, followed by after-spike reset:

v ≥ 30mV, then    v ← c u ← u + d

The A, B, C are the standard optimized parameters of the Izhikevich artificial neuron model, whereas the a, b, c, and d parameters were selected (Table 3.1) so as to allow a regular spiking behavior [Izhikevich, 2004], being able to encode both sustained indentation and dynamic stimulus sliding.

A B C Cm a b c d GF 0.04 s−1V−1 5 s 140 V s−1 1 F 0.02 s−1 0.2 -65 mV 8 mV 10000/Ω

Table 3.1: Izhikevich model parameters

GF is a gain factor that decides the intensity at which our neuron model spikes. Then we have investigated the inter-spike-interval2 _{(ISI) distribution of the neuron model} in response to stimuli as a function of varying GF value. High GF values led to an increase in spike rate, but monotonic decrease in the ISI (increase in firing rate) leading to an increase in the overlapping of the ISI distributions across stimuli. On other hand, a small GF leads low (or) no spiking activity. We have selected a GF value within an optimal range with which we could find a fair ISI distribution across

various stimuli.

3.3

Neuromorphic Tactile System

We further investigated and validated the information content in the spike trains elicited by our neuromorphic approach (section 3.2), during a passive touch protocol (Fig. 16B) across 10 naturalistic textures (Fig. 16C). During the passive touch pro-tocol, the tactile sensor (fingertip, Fig. 16) is fixed and the stimuli are slid across the surface of the tactile sensor. The force at the point of contact between stimuli and tactile sensor is further referred as contact force, and the velocity at which the stimuli are slid across the surface of sensor is referred as sliding velocity.

Figure 16: (A) Sensor data processing. From left to right: illustrative representation of the fingertip structure on the top of which the bio-inspired tactile sensor is placed, and a spatial arrangement of the 16 sensor channels. The pre-processed analog data from the sensor was fed into the Izhikevich neuronal model mimicking the dynam-ics of adaptive mechanoreceptors. (B) Passive touch protocol. From left to right: preparation, normal indentation (z-axis) before sliding (x-axis), tangential sliding, indentation after sliding, and retraction. (C) Microscopy images of the tactile stimuli that are presented to the sensors. first five textiles, last five non-textiles. Note that the lines in the glass appear in transparency from the surface below, whereas glass surface is quite smooth.

3.3.1

Classification Based on Spike Responses

Towards this investigation we tried to classify the receptor responses across all the stimuli (Fig. 17) for a fixed contact force (400 mN) and sliding velocity (15 mm/s). The first classification was based on the average spike rate combined with the coeffi-cient of variation (CV) of the inter-spike-interval (ISI). We built a classifier based on these 2-dimensional features using k-nearest neighbors (knn) algorithm with leave-one-out validation and Euclidean distance. This approach yielded a classification accuracy of 68% to 93%, depending on the selected sensor channel.

Figure 17: Adaptive mechanoreceptor response to different tactile stimuli. Each row shows the response of an artificial receptor to a given tactile stimulus, for fixed force (400 mN) and sliding velocity (15 mm/s).

The second classification method was pursued based on precise spike timing [Johans-son and Birznieks, 2004] approach, and to this aim, we measured the distance between each pair of spike trains using Victor Purpura distance metrics (VPd) [Victor and Purpura, 1996]. Briefly, this metrics measures the cumulative cost for transformation of one spike train to another by: 1) adding/removing the spikes (cost = 1) 2) shifting the time of a single spike by an interval ∆t (cost = q∆t, q is a constant parameter). We classified the responses with leave-one-out validation according to the identity of the knn of each response. This method yielded a higher classification accuracy of 93% to 97%, indicating that temporal spike information conveys rich information compared to the spike rate (frequency) or CV of ISI (regularity of spike trains).

3.3.2

Effect of Sensing Dynamics

We also investigated and evaluated the effect of sensing dynamics (varying contact force and sliding velocity) on the output spike responses and decoding performance across varying sensing conditions. We have tested across 25 combinations of force and velocity conditions, that range between 200 mN to 600 mN (steps of 100 mN) and 5mm/s to 25 mm/s (steps of 5 mm/s). The individual combination of force and velocity based decoding showed us a homogeneous activity modulation across the variation of sensing conditions. Most of the combinations had a decoding performance higher than 85%, beside the combination of low velocities and weak forces (as shown in Fig. 18A).

We also performed stimuli decoding irrespective of sensing conditions (contact force or sliding velocity). We found that the information patterns in spike trains were con-sistently repeatable for the same stimulus across different velocities, when expressed in terms of space instead of time (Fig. 18B). Therefore, we devised a spatial VPd (sVPd): after the re-scaling of the spike times according to the sliding velocity, we determined the sVPd replacing the classical time shifting cost with a space-shifting cost. With the help of sVPd, we were able to decode textures across a five-fold vari-ation of velocity with an accuracy of 90% (Fig. 18C). However, this method was less effective when tried across different forces, where the decoding was limited to 48%.

Figure 18: Sensing Dynamics (A) Correct VPd based decoding % across each tested combination of contact force and sliding velocity. Chance level is 10% (B) Spike trains scaled according to the fingertip position relatively to the stimuli, for all sliding velocities and fixed force (400 mN). Responses from 10 repetitions of same stimuli for same configuration is shown. (C) Confusion matrix of stimuli decoding, based on knn clustering applied on the spatial VPd. This classification is done across all the stimuli for all five sliding velocities.

3.3.3

Discussion

The stimuli decoding performance (as high as 97%) based on neuromorphic spike events was robust for a broad range of sensing conditions across multiple stimuli. This shows the efficacy and diversity in the information content of neuromorphic systems. Being able to categorize naturalistic textures under varying conditions shows a promising future of neuromorphic systems in neuro-robotics or neuroprostheses. Whereas, the Victor Purpura decoding algorithm is not real-time, requiring a whole window of data in order to compute the distance between spike trains and classify. We also could not build a VPd based decoding algorithm efficiently across varying forces for the same texture. On the other hand, all the results presented in this chapter were based on the neuronal output associated with a single sensor channel. Therefore, we opted to construct a second layer of neuronal network that paves way towards a bio-inspired architecture, contributing for real-time stimuli decoding. In the following chapter we brief the architectures that we developed towards this objective.

Chapter 4

Modelling Cuneate Neurons

Towards the implementation of on-line tactile learning architecture we modeled second-order neurons (Cuneate Neurons (CNs)) based on two different hypotheses, 1) Con-sidering CNs as coincidence detectors based on the hypothetical model proposed by Johansson et al. [Johansson and Flanagan, 2009] 2) Considering CNs as feature extractors based on the in vivo recordings and observation by Henrik J¨orntell et al. [Bengtsson et al., 2013, J¨orntell et al., 2014].

Following both of these architectures, we had the possibility in integrating a popula-tion of PA sensors that in-turn results in creating a sparse representapopula-tion of tactile stimuli. To the best of our knowledge, previously the only computational model with bio-inspired two layer architecture for tactile stimuli was proposed by Bologna et al. [Bologna et al., 2011, Bologna et al., 2013]. Their implementation of first order neurons is based on leaky integrate-and-fire with spiking adaptation by means of a “threshold fatigue” and the second order neurons are designed as linear probabilistic models, and their activation based on coincidence detection in first order neurons. In our work, we also follow the hypothesis of coincidence detection, but we opted for much more minimalistic approach to understand if precise-spike time dependent plasticity be useful in robotics. Further in our second hypothesis, we designed the cal-cium dynamic model of cuneate neurons and the synaptic learning mechanism which is state-of-art study in the field of neurophysiology and neuro-computation.

coinci-dence detector. In section 4.2, we define a new mechanism that was modeled based on the hypothesis of CN as feature extractor. These new findings could possibly hint a unseen mode by which the brain characterizes the surrounding tactile world.

4.1

Cuneate Neurons as Coincidence Detectors

Johansson et al. claim the design of the somatosensory pathways could enable rapid classification of tactile stimuli by temporal - to - spatial conversion at the level of secondorder neurons [Johansson and Flanagan, 2009]. Based on the argument of rapid classification, they propose that CNs with a function of coincidence detection could be one possibility. Strict implementation of such hypothesis, does not match with many neurophysiology findings [Nu˜nez and Bu˜no, 2001, Bengtsson et al., 2013, Jones, 2000, J¨orntell et al., 2014], therefore this implementation is useful only in understanding the effect of information processing across population of neurons and can also be helpful to verify if such minimalist hypothesis could help creating effective real-time classifier for robotics applications. This hypothetical model of Cuneate Neurons in this thesis will further be address as CNcd (Cuneate Neurons as coincidence detector), in-order to avoid confusion with neurophysiology based studies stated in Part A and their respective model in section 4.2.

4.1.1

Coincidence Detector

This hypothesis is mainly based on two major factors: 1) The patterns of divergence and convergence of primary afferents onto second-order neurons. 2) The variation in axonal conduction velocity, resulting in conduction delays (Fig. 19). As it is known from previous biology studies, that 1000’s of PAs project on single CN and a single PA project on 100s of CNs. This random connectivity along with conduction delays, will provide varied spatiotemporal patterns across two different CNscd for the same set of PA activations. The CNscd are based on precise spike time activation, as they conduct when multiple spikes occur at the same time. This property makes CNscd sensitive to specific spatiotemporal properties within the population of primary afferent sensors

Figure 19: Hypothetical model of processing primary afferent information, considering the second order neurons function as coincidence detectors. Figure from [Johansson and Flanagan, 2009].

(Fig. 19). This architecture can provide parallel processing along with “feed-forward rapid classification of information” [Johansson and Flanagan, 2009].

4.1.2

Modelling Coincidence Detector

We have modeled the CNscd as regular spiking Izhikevich neurons (as for artificial mechanosensors). The input to CNscd was given by the sum of the excitatoy inputs from mechnosensory neuron modeled as current-based post-synaptic potentials [Cav-allari et al., 2014], such that the differential equations determining the evolution of their dynamics: dv(t)) dt = Av(t) 2_{+ Bv(t) + C − u(t) + J} X iP RE P SP (t − t+_i )) du(t)) dt = a(bv(t) − u(t)) if (v(t∗) ≥ vth), then a spike occurs at t∗ and resulting,

  

v(t∗+ ∆t) ← c u(t∗+ ∆t) ← u(t∗) + d

Where the values of A, B, C, a, b, c, d are listed in Table 3.1, J is the fixed synaptic strength (8mV/ms), t+_i are the times of spikes fired by the pre-synaptic neuron i,

P RE is the set of the pre-synaptic neurons for the neuron considered, and P SP (t) is a bi-exponential function mimicking the temporal evolution of AMPA synapses Post-Synaptic Potentials, with rise time of 0.5 ms and decay time 2 ms [Mazzoni et al., 2008].

In this simplified model, neurons in CNscd do not differ for the set of presynaptic neurons, but for the conduction delays of their inputs from peripheral neurons (as seen in Fig. 20A). We want to implement the condition in which differential delays between peripheral spike trains play a role in the encoding of information. If response latency and delay difference approximately compensate the postsynaptic delay, the CNcdwill receive two superimposed excitatory stimuli, increasing the probability that the membrane potential will cross spike threshold and fire. In this way, differential delays make possible to have angle-sensitive CNscd, i.e., neurons firing only when the contact with the sensors occur with a given latency, which for constant sliding speed corresponds to a specific presented edge angle (as illustrated in Fig. 20C).

4.1.3

Encoding Orientation

In order to validate the second layer architecture, we devised a stimulus consisting 3 ridges, forming an angle of −450_{, 0}0 _{and 45}0 _{with direction orthogonal to the} stimulus sliding (Fig. 20B). The stimuli was presented in a passive touch protocol (as described in Fig. 16B) across the neuromorphic tactile sensor surface, with a contact force of 400 mN and sliding velocity of 20 mm/s. The relative latency of the neuromorphic response between the two sensor channels (Mechanosensor 1 (MS 1) and Mechanosensor 2 (MS 2)) is shifted during the sliding as a function of stimulus angle. Therefore, the second layer architecture (Fig. 20A) allowed us to reconstruct the angle (based on a predefined labeling strategy, look-up table) of the detected edge based on the input latency in mechanosensors (Fig. 20E) along with the deferential delays (Fig. 20F).

Figure 20: Implementation and validation of two layer architecture. (A) CN process-ing the mechanoreceptor inputs (green and blue spikes). In each row the two channels impinge on a different CNcd with a specific relative delay. Due to the delays, CN1cd mainly responds to the first two stimuli, CN2cd respond for vertical one and CN3cd to the last two. (B) Representation of stimulus and its orientation. (C) Contact times between the two vertically aligned sensors channels (S1, S2), relative to the stimulus motion. (D) Illustration of mechanosensor responses to the three ridges. (E) Raster plot of responses elicited over 10 trails by difference ridges across different mechanosensors (SA-like, FA-like) associated to the two channels (S1, S2) (F) Raster plot of responses elicited over 10 trails in the three CNcd.

4.1.4

Discussion

This implementation helped us understand the efficacy of neuromorphic information across population of PAs. The minimalistic approach followed in this chapter for modelling the first order and second order neurons showed us promising results by decoding geometrical shapes. Implementing this approach with more number of PAs and CNscd, can result in creating a sparse representation of a geometric stimuli in multi-dimensional space. The straightforward architecture design will enable us to implement on-line decoding while gathering the tactile data stream. Neuromorphic implementation of such architectures are advantageous for the leanness, as the burden of data computation and storage is minimal. One major limitations in this hypothesis

is precise spike time consideration, which need the system to be optimally tuned with respect to the stimuli, in-order to create an effective real-time decoder. This also makes the system less robust to external noise. Therefore, we have implemented a CNcdmodelling with dynamics and synaptic weight learning, which makes the system more robust and neurophysiologically plausible.

4.2

Cuneate Neurons as Feature Extractor

In this section, we present our approach towards modelling an overall structure of Cuneate Neuron (CN) network based on the known neurophysiology. We modeled Cuneate Neurons (CNs) and their synaptic learning mechanism in order to create, a representation-free, feature learning mechanism. Based on the present biological observations [Bengtsson et al., 2013] (as briefed in Section 1.4), we modelled the physiology of cuneate nucleus neurons. Further we adapted a network between pop-ulation of primary afferent inputs (from our artificial neuromorphic tactile system (Section 3.3)) and the modelled CNs using synaptic learning mechanism. This net-work was then gradually adapted to input neuromorphic tactile responses, generated from a variety of touch experiences. Integrating all the components in a functional system, we observed that the network model tend to reduce the population of input sensory dimensions to a small utility-oriented dimensions of correlated tactile sensors, suggesting a unseen mode of brain representation of the external world.

4.2.1

Cuneate Neuron Model

The CN model1 _{was implemented as a conductance based Exponential Integrate and} Fire (EIF) model [Ostojic and Brunel, 2011], which can recreate the fast dynamics (∼ 1ms timescale) of spike generation. In addition to the basic EIF model, voltage sensitive calcium channels and calcium dependent potassium channels (Section 1.4.3) were also modelled in order to recreate intermediate cuneate neuron dynamics (∼ 10ms timescale) as observed in in vivo recordings (Fig. 9). The complete dynamics

1 _{The calcium dynamic model of CNs, was work of Anton Spanne}

of the membrane potential are given by,

Cm dVm

dt = IL+ Ispike + Iion+ Iext+ Isyn (4.1)

where, Cm is the membrane capacitance, IL= −¯gL(Vm− EL) is the leak current, Ispike is the spike currents (fast dynamics), Iion is the ion channel currents (interme-diate dynamics), Isyn is the synaptic input currents and Iextdenotes external injected currents that were used to evaluate the intrinsic CN neuron responsiveness to current step commands (Fig. 9). The membrane resistance and time constant were within the range of values recorded in vivo [Bengtsson et al., 2013], whereas other parameter values were chosen through an optimization process.

Fast Dynamics

The spike current (Ispike) is generated using a basic EIF model (Eq. 4.2) to achieve the fast dynamics and recreate the initiation of the action potential [Ostojic and Brunel, 2011] Ispike = −¯gL∆Texp  V_m− Vt ∆T  (4.2)

As the depolarization reaches a threshold speed, an action potential is said to be generated, and the membrane potential is immediately reset to a value close to the steady state potential following a real action potential. In the above equation 4.2, gL is the leak conductance and ∆t gives width of the spike.

Intermediate Dynamics

The intermediate dynamics include currents from additional ion channels (Iion, Eq. 4.1) that are not directly involved in forming action potentials, but have more of a mod-ulating role in episodes leading up to the generation of action potential and the episodes between action potentials when the input activity is high. The intermedi-ate dynamics of the model were optimized to mimic the reactive conductances that could generate the types of responses to current injections we recorded in the cuneate neurons in vivo (Fig. 9). Such responses, i.e. post-inhibitory rebound responses and

a tendency to generate bursts of action potential firing have been observed in other neuron types [Huguenard, 1996, Llinas and Jahnsen, 1982] and such responses can be explained by low-threshold voltage gated calcium channels (LVA) and calcium-activated potassium channels (CAP). Hence, Iion can be divided according to

Iion = ICa+ IK (4.3)

where ICa is the current through LVA channels or equivalent channels and IK the current through the CAP channels. The modelling of these currents in explained in Article 3 (page 85).

Synaptic Input

The synaptic current (Isyn) through the cell membrane is the summated synaptic currents of the activated synapses. Each individual synapse (i) is activated by a primary afferent spike and once activated this spike gives rise to a stereotyped time course of conductance injection at the synapse (Isyn) which is described by

Isyn = {gmax X

i

wexc,iexp(−τ t∗)(Erev,exc− Vm)

+gmaxwinh X

i

exp(−τ t∗)(Erev,inh − Vm)}

(4.4)

where Erev is the reversal potential of the type of synapse (Erev,exc or Erev,inh depending on whether the synapse is excitatory or inhibitory, see Table 4), Vm is the membrane potential and t∗ is the time of activation of the synapse. Each spike in each sensory afferent was converted into a synaptic conductance in the simulated neuron. For each synapse, the peak amplitude of the synaptic response was determined by the product of their individual weight (wexc or winh) and the overall maximum synaptic conductance constant (gmax, see Article 3). The time constants of the decay of the synaptic membrane potential responses were 6.4 ms for both excitatory and inhibitory synapses, in accordance with the time courses recorded in the cuneate nucleus neurons in vivo [Bengtsson et al., 2013]. Note that as all the PA synapses of our system stayed

well below 200 Hz of firing activity, we did not simulate any rate adaptation of the PA synapses as such adaptation in vivo primarily occurs at intervals shorter than 5 ms [Jorntell and Ekerot, 2006].

Subsynaptic Activity

In the learning process, excitatory synaptic weight learning was driven by the cal-cium activity in the main compartment of the cuneate neuron (i.e. as calculated by the calcium dynamic model, shown in Article 3) in combination with the calcium activity in the individual synapses. An essential component of this combination is the intensity of activation of the individual synapses. According to the learning rule (the synaptic learning rule of both excitatory and inhibitory synapses is briefed in Article 3.) that we applied, a synapse that fires at high frequency with respect to the main compartment total calcium activity (ACa2+

tot = k[Ca2+], where k is an arbitrary constant that is here assumed to be 1) will be “rewarded”, as it has strong correlation with the learning signal ACa2+

tot (Fig. 21C). Conversely, strong firing in a synapse in relation to low or zero ACa_tot2+will be “punished” (i.e. similar to the classical BCM rule for Hebbian plasticity) [Bienenstock et al., 1982]. Therefore, the local calcium time constants (τ[Ca]_loc2+), defined as the temporal properties of the calcium signal in the local space underneath the each individual synapse, play a major role in the learning process (the local postsynaptic calcium activity can be considered an analogy with the calcium activity in a local dendritic spine [Koester and Sakmann, 1998, Tigaret et al., 2016]). The learning rule critically depends on this time constant. For instance, if τ[Ca]_loc2+ time constants are too high, the rewarding effects on synapses that have a high degree of correlation with the ACa_tot2+ will be lost. However, as there is no data on the relevant time constants in the cuneate neurons in vivo, we had to make assumptions of the values of this time constant. In order to avoid pitfalls in relation to this assumption, we studied a range of time constants for ACa2+

loc during the CN learning process (effect of this parameter variation can be seen in Fig. 23).

For each synapse, each input spike at time t* contributes to the subsynaptic spine calcium concentration, the time course of which is given by the kernel [Mazzoni et al., 2008] of equation 4.5

α = τ1 τd− τr ∆ACa_loc2+(t) = α ∗  exp  −t − τl− t ∗ τr  − exp  −t − τl− t ∗ τd  (4.5)

The parameters describing the relative local calcium concentration (or activity), are the decay time τd = 12.5 ms and the rise time τr = 4 ms multiplied with a constant τ1 = 21 ms (which is a constant to calculate the ratio). The initial values chosen were derived from our assumption that the time course of the slow afterhyperpolarization of the cuneate neuron spike (Fig. 9), which is known to reflect the activation of calcium-dependent potassium channels, matched the time course of the calcium concentration induced in the synapse.

4.2.2

Network Connectivity

Our model system mimicked the overall structure of the cuneate network (Fig. 7), but replaced the biologically sensorized skin with neuromorphic inputs from the tactile sensor (Fig. 21A). We have created 80 different primary afferent responses, by multi-plexing the existing sensory channels data of our bio-inspired tactile sensor (Fig. 16). Further we designed a functional structure of CNs (Fig. 21B), where the model CN was composed of a central calcium dynamic model, driven by excitatory and in-hibitory synaptic inputs from the sensors (Fig. 21B). In order to attain the specificity in learning and to avoid influence between neighboring synapses, we have opted for compartmentalization of our calcium dynamics, based on spine local calcium and main compartment total calcium (Fig. 21C).

4.2.3

Effect of Synaptic Learning

Learning Protocol

The CN learning process was simulated using 1500 presentations of a varied set of stimuli presented in a pseudorandom order that was the same for all simulations. Each of the 1500 presentations, comprise spatiotemporal spike input patterns, that

Figure 21: Functional structure of CN model. Colored lines indicate PA connections between the physical sensors of the artificial fingertip and the neurons of the model, blue triangles indicate variable weight synapses. Only projection neurons (CN1 -CNn) were simulated explicitly. PA inputs were also provided as a lump inhibitory synaptic input via an interneuron (red). (B) Graphic representation of the functional structure of the CN model. (C) Subcellular factors in the synaptic plasticity of the CN, with variable weight PA synapses (blue) and inhibitory synapses (red). The CN neuron is divided into a main compartment, with reactive LVA and CAP con-ductances, and synaptic spaces containing VGCCs [Higley and Sabatini, 2012] and a variable number of AMPAr:s. AMPAr, excitatory glutamate receptors; VGCCs, volt-age gated calcium channels; GABAa, inhibitory synaptic receptors; LVA, lowthreshold voltage gated calcium channels; CAP, calcium-dependent potassium channels.

are derived from five different stimuli (3-naturalistic textures, 2-shapes), repeated 300 times each with a Gaussian noise with zero mean and standard deviation σ. (σ = 5ms).