Non-spiking neuron

Last updated

Non-spiking neurons are neurons that are located in the central and peripheral nervous systems and function as intermediary relays for sensory-motor neurons. They do not exhibit the characteristic spiking behavior of action potential generating neurons.

Contents

Characteristic spiking behavior of neuron Relabeled action potential.jpg
Characteristic spiking behavior of neuron
QualitiesSpiking neuronsNonspiking neurons
LocationPeripheral and centralPeripheral and central
BehaviorAction potentialFewer sodium channel proteins

Non-spiking neural networks are integrated with spiking neural networks to have a synergistic effect in being able to stimulate some sensory or motor response while also being able to modulate the response.

Neural synapse Neurons big1.jpg
Neural synapse

Discovery

Animal models

There are an abundance of neurons that propagate signals via action potentials and the mechanics of this particular kind of transmission is well understood [ citation needed ]. Spiking neurons exhibit action potentials as a result of a neuron characteristic known as membrane potential. Through studying these complex spiking networks in animals, a neuron that did not exhibit characteristic spiking behavior was discovered. These neurons use a graded potential to transmit data as they lack the membrane potential that spiking neurons possess. This method of transmission has a huge effect on the fidelity, strength, and lifetime of the signal. Non-spiking neurons were identified as a special kind of interneuron and function as an intermediary point of process for sensory-motor systems. Animals have become substantial models for understanding more about non-spiking neural networks and the role they play in an animal’s ability to process information and its overall function. Animal models indicate that the interneurons modulate directional and posture coordinating behaviors. [1] [2] Crustaceans and arthropods such as the crawfish have created many opportunities to learn about the modulatory role that these neurons have in addition to their potential to be modulated regardless of their lack of exhibiting spiking behavior. Most of the known information about nonspiking neurons is derived from animal models. Studies focus on neuromuscular junctions and modulation of abdominal motor cells. Modulatory interneurons are neurons that are physically situated next to muscle fibers and innervate the nerve fibers which allow for some orienting movement. These modulatory interneurons are usually nonspiking neurons. [3] Advances in studying nonspiking neurons included determining new delineations among the different types of interneurons. These discoveries were due to the usage of methods such as protein receptor silencing. Studies have been done on the non-spiking neuron qualities in animals of specific non-spiking neural networks that have a corollary in humans, e.g. retina amacrine cell of the eye.

Physiology

Definition

A non-spiking neuron is a neuron that transmits a signal via graded potential. The rate of subsequent neurotransmitter release is linearly correlated with the magnitude and sign of summed inputs which allows them to preserve specific features of the eliciting stimulus, such as light quanta information by photoreceptors. [4] They are a fundamental component of visual processing in the retina. [5] They can be more susceptible to noise. Studies show that these neurons may offer a contribution to learning and modulation of motor neuron networks.

Spiking neurons and non-spiking neurons are usually integrated into the same neural network, but they possess specific characteristics. The major difference between these two neuron types is the manner in which encoded information is propagated along a length to the central nervous system or to some locus of interneurons, such as a neuromuscular junction. Non-spiking neurons propagate messages without eliciting an action potential. This is most likely due to the chemical composition of the membranes of the non-spiking neurons. They lack protein channels for sodium and are more sensitive to certain neurotransmitters. They function by propagating graded potentials and serve to modulate some neuromuscular junctions. Spiking neurons are noted as traditional action potential generating neurons. [4]

Identification

"Interneurons" is a name used to indicate neurons that are neither sensory neurons nor motory in nature, but function as an intermediary processing and transmission state for signals that have been received via dorsal root ganglia cells. [3] A large amount of these interneurons seem to exhibit the non-spiking characteristic. To better define non-spiking neuron signal transmission and signal transduction, many experiments have been performed to qualify and quantify the fidelity, speed, and mechanics of signal transmission in non-spiking neurons. There have been classifications based on the larger group "interneurons" where pre-motor nonspiking neurons are referred to as postlateral (PL) or anteriolateral (AL) interneurons, with AL interneurons divided into three types of interneurons based on staining. The initial differentiation between PL and AL interneurons are their responses to GABA, a neurotransmitter for muscle tone. They also have different staining responses permitting quick and qualified classification. [6]

A retinal amacrine cell Amacrine.png
A retinal amacrine cell

Cell types

Many of the nonspiking neurons are found near neuromuscular junctions and exist as long fibers that help to innervate certain motor nerves such as the thoracic-coxal muscle receptor organ (TCMRO) of a crab. [4] They function in a modulatory role by helping to establish posture and directional behavior. This was intensely modeled in the crustacean and in insects showing how appendages are oriented via these nonspiking neural pathways. [2] Amacrine cells are another major type of non-spiking neuron and their lifetime involves the conversion to a non-spiking neuron from a spiking neuron once the retina obtains maturity. They are one of the first cells to differentiate during prenatal development. Upon the opening of the eyes, these cells begin to shed their sodium ion channels and become non-spiking neurons. It was hypothesized that the reason for its establishment as a spiking neuron was to help with the maturation of the retina by the usage of action potentials themselves, and not necessarily the information the action potential carried. This was supported with the occurrence of synchronous firing by the starburst amacrine cells during the initial stages of development. This study used a rabbit model. [7] However, spiking wide field amacrine cells have been identified in the adult rabbit retina. These cells extend processes spanning >1mm across the retina and actively propagate dendritic spikes to and from the soma [8] Additionally, a spiking GABAergic nitric oxide producing amacrine cell type (nNOS-1 AC) has been identified in mice and is thought to play a role in precise feature extraction from light through a range of noisy background luminance. [9]

Physiological characteristics

Some studies have indicated that even with the volatility of signal transmission with these particular neurons, they still perform very well in maintaining signal strength. Studies show that the ratio of signal to noise in experimental settings of some signals are at least 1000 and upwards to 10000 over 5-7mm of propagation length by nerves. [4]

Effect of noise on non-spiking neuron transmission Nonspiking Transmission.jpg
Effect of noise on non-spiking neuron transmission

These interneurons are connected to one another via synapses and a minority, approximately 15% of the neurons, exhibit bidirectional capacity and were excitatory. About 77% of these neurons indicated a one-way mode of transmitting signals which were inhibitory in nature. These numbers were modeled from an arthropod as pre-motor elements in the motor control system. They were located in the abdominal region. Synapses are known as gaps between neurons which facilitate the spread of a message via neurotransmitters that may excite or depress the subsequent neuron through a complex cascade of electrochemical events. For the interneurons exhibiting one-way signaling, they would receive an excitatory stimulus, experimentally, and the post-synaptic cell was given an inhibitory signal. The interaction between the two cells was modulatory in which the pre-synaptic cell with the initial excitatory signal would mediate the postsynaptic cell even after being inhibited. Signal amplitude was used to determine the effects of the modulation on the signal transmission. [10]

The speed of signal transmission at 200 Hz, the most conserved bandwidth of signal transmission for non-spiking neurons, was approximately 2500 bits/second in which there was a 10-15% decrease in speed as the signal propagated down the axon. A spiking neuron compares at 200bits/ second, but reconstruction is greater and there is less influence by noise. There are other non-spiking neurons that exhibit conserved signal transmission at other bandwidths. [4]

Cell TypeCharacteristics
ArthropodOrienting motor control
Rabbit amacrine cell Eyes, establishment of function
CrustaceanOrienting motor control; 2500 bit/s; bandwidth of 200 Hz

While some non-spiking neurons are specifically involved in neuromuscular modulation, studying amacrine cells has created opportunities to discuss the role of non-spiking neurons in neuroplasticity. Since amacrine cells, which are a type of non-spiking neurons, undergo a transformation from spiking to non-spiking cells, there have been many studies that try to identify the functional reasons for such a transformation. Starburst amacrine cells use action potentials during retinal development, and once the retina is mature, these cells transform into non-spiking neurons. The change from a cell that can generate action potentials to solely functioning off of a graded potential is drastic, and may provide insight into why the two kinds of neural networks exist. The cells lose sodium channels. The loss of the sodium channels is triggered by the opening of the eye correlating to the possibility of the environment playing a crucial role in determination of neural cell types. The rabbit animal model was used to develop this particular study. This transition is not quite understood but heavily concludes that the spiking and non-spiking statuses occupied by the starburst amacrine cells are vital to the maturation of the eyes. [7]

Functions

Modulation

By using known neurotransmitters that affect non-spiking neurons, modeled neural networks may be modified to either ease neuromuscular hyperactivity, or cells themselves may be transformed to be able to provide stronger signals. A calcium transporter study indicates the effect that protein channels have on the overall fidelity and firing capacity of the non-spiking neurons. [11] Since most of the propagated messages are based on a proportionality constant, meaning, there is not a temporal or spatial significance to the presynaptic firing, these signals literally "repeat what they have been told". When it comes down to chemical systems in the body, a non-spiking neural network is definitely an area of exploration. [12] The amacrine cell study poses new and exciting components to the study of altering the chemical and mechanical properties of the non-spiking neural networks. [7]

Memory and learning

Very little is known about the application of these networks to memory and learning. There are indications that spiking and nonspiking networks both play a vital role in memory and learning. [13] [14] Research has been conducted with the use of learning algorithms, microelectrode arrays, and hybrots. By studying how neurons transfer information, it becomes more possible to enhance those model neural networks and better define what clear information streams could be presented. Perhaps, by conjoining this study with the many neurotrophic factors present, neural networks could be manipulated for optimal routing, and consequently optimal learning. [15]

Device production

By studying the nonspiking neuron, the field of neuroscience has benefited by having workable models that indicate how information is propagated through a neural network. This allows for the discussion of the factors that influence how networks work, and how they may be manipulated. Non-spiking neurons seem to be more sensitive to interference given that they exhibit graded potentials. So for non-spiking neurons, any stimulus will elicit a response, whereas spiking neurons exhibit action potentials which function as an "all or none" entity. [4]

In biomedical engineering, it is a priority to understand the biological contributions to an overall system in order to understand how the systems may be optimized. Paul Bach y Rita was a famous believer of neuroplasticity and integrated the principles of device design in order to model what neurons were actually doing in the brain and create a device that simulated functions already prescribed by the biological system itself. Some special advances made in the medical field based on structured models of biological systems include the cochlear implant, practices encouraged by Dr. VS Ramachandran on phantom limbs and other optical applications, and other devices that simulate electrical impulses for sensory signal transduction. By continuing to achieve a workable model of the non-spiking neural network, its applications will become evident. [16]

See also

Related Research Articles

<span class="mw-page-title-main">Neuron</span> Electrically excitable cell found in the nervous system of animals

Within a nervous system, a neuron, neurone, or nerve cell is an electrically excitable cell that fires electric signals called action potentials across a neural network. Neurons communicate with other cells via synapses, which are specialized connections that commonly use minute amounts of chemical neurotransmitters to pass the electric signal from the presynaptic neuron to the target cell through the synaptic gap.

<span class="mw-page-title-main">Nervous system</span> Part of an animal that coordinates actions and senses

In biology, the nervous system is the highly complex part of an animal that coordinates its actions and sensory information by transmitting signals to and from different parts of its body. The nervous system detects environmental changes that impact the body, then works in tandem with the endocrine system to respond to such events. Nervous tissue first arose in wormlike organisms about 550 to 600 million years ago. In vertebrates, it consists of two main parts, the central nervous system (CNS) and the peripheral nervous system (PNS). The CNS consists of the brain and spinal cord. The PNS consists mainly of nerves, which are enclosed bundles of the long fibers, or axons, that connect the CNS to every other part of the body. Nerves that transmit signals from the brain are called motor nerves (efferent), while those nerves that transmit information from the body to the CNS are called sensory nerves (afferent). The PNS is divided into two separate subsystems, the somatic and autonomic, nervous systems. The autonomic nervous system is further subdivided into the sympathetic, parasympathetic and enteric nervous systems. The sympathetic nervous system is activated in cases of emergencies to mobilize energy, while the parasympathetic nervous system is activated when organisms are in a relaxed state. The enteric nervous system functions to control the gastrointestinal system. Nerves that exit from the brain are called cranial nerves while those exiting from the spinal cord are called spinal nerves.

An inhibitory postsynaptic potential (IPSP) is a kind of synaptic potential that makes a postsynaptic neuron less likely to generate an action potential. The opposite of an inhibitory postsynaptic potential is an excitatory postsynaptic potential (EPSP), which is a synaptic potential that makes a postsynaptic neuron more likely to generate an action potential. IPSPs can take place at all chemical synapses, which use the secretion of neurotransmitters to create cell-to-cell signalling. EPSPs and IPSPs compete with each other at numerous synapses of a neuron. This determines whether an action potential occurring at the presynaptic terminal produces an action potential at the postsynaptic membrane. Some common neurotransmitters involved in IPSPs are GABA and glycine.

<span class="mw-page-title-main">Excitatory postsynaptic potential</span> Process causing temporary increase in postsynaptic potential

In neuroscience, an excitatory postsynaptic potential (EPSP) is a postsynaptic potential that makes the postsynaptic neuron more likely to fire an action potential. This temporary depolarization of postsynaptic membrane potential, caused by the flow of positively charged ions into the postsynaptic cell, is a result of opening ligand-gated ion channels. These are the opposite of inhibitory postsynaptic potentials (IPSPs), which usually result from the flow of negative ions into the cell or positive ions out of the cell. EPSPs can also result from a decrease in outgoing positive charges, while IPSPs are sometimes caused by an increase in positive charge outflow. The flow of ions that causes an EPSP is an excitatory postsynaptic current (EPSC).

<span class="mw-page-title-main">Neuroeffector junction</span> Site where a motor neuron releases a neurotransmitter to affect a target cell

A neuroeffector junction is a site where a motor neuron releases a neurotransmitter to affect a target—non-neuronal—cell. This junction functions like a synapse. However, unlike most neurons, somatic efferent motor neurons innervate skeletal muscle, and are always excitatory. Visceral efferent neurons innervate smooth muscle, cardiac muscle, and glands, and have the ability to be either excitatory or inhibitory in function. Neuroeffector junctions are known as neuromuscular junctions when the target cell is a muscle fiber.

<span class="mw-page-title-main">Neural circuit</span> Network or circuit of neurons

A neural circuit is a population of neurons interconnected by synapses to carry out a specific function when activated. Multiple neural circuits interconnect with one another to form large scale brain networks.

<span class="mw-page-title-main">Amacrine cell</span> Interneuron cells in the retina of the eye

In the anatomy of the eye, amacrine cells are interneurons in the retina. They are named from Greek a– 'non', makr– 'long', and in– 'fiber', because of their short neuronal processes. Amacrine cells are inhibitory neurons, and they project their dendritic arbors onto the inner plexiform layer (IPL), they interact with retinal ganglion cells, and bipolar cells or both of these.

An apical dendrite is a dendrite that emerges from the apex of a pyramidal cell. Apical dendrites are one of two primary categories of dendrites, and they distinguish the pyramidal cells from spiny stellate cells in the cortices. Pyramidal cells are found in the prefrontal cortex, the hippocampus, the entorhinal cortex, the olfactory cortex, and other areas. Dendrite arbors formed by apical dendrites are the means by which synaptic inputs into a cell are integrated. The apical dendrites in these regions contribute significantly to memory, learning, and sensory associations by modulating the excitatory and inhibitory signals received by the pyramidal cells.

<span class="mw-page-title-main">Neurotransmission</span> Impulse transmission between neurons

Neurotransmission is the process by which signaling molecules called neurotransmitters are released by the axon terminal of a neuron, and bind to and react with the receptors on the dendrites of another neuron a short distance away. A similar process occurs in retrograde neurotransmission, where the dendrites of the postsynaptic neuron release retrograde neurotransmitters that signal through receptors that are located on the axon terminal of the presynaptic neuron, mainly at GABAergic and glutamatergic synapses.

<span class="mw-page-title-main">Synapse</span> Structure connecting neurons in the nervous system

In the nervous system, a synapse is a structure that permits a neuron to pass an electrical or chemical signal to another neuron or to the target effector cell.

<span class="mw-page-title-main">Synaptic gating</span>

Synaptic gating is the ability of neural circuits to gate inputs by either suppressing or facilitating specific synaptic activity. Selective inhibition of certain synapses has been studied thoroughly, and recent studies have supported the existence of permissively gated synaptic transmission. In general, synaptic gating involves a mechanism of central control over neuronal output. It includes a sort of gatekeeper neuron, which has the ability to influence transmission of information to selected targets independently of the parts of the synapse upon which it exerts its action.

Neural backpropagation is the phenomenon in which, after the action potential of a neuron creates a voltage spike down the axon, another impulse is generated from the soma and propagates towards the apical portions of the dendritic arbor or dendrites. In addition to active backpropagation of the action potential, there is also passive electrotonic spread. While there is ample evidence to prove the existence of backpropagating action potentials, the function of such action potentials and the extent to which they invade the most distal dendrites remain highly controversial.

<span class="mw-page-title-main">Dendritic spike</span> Action potential generated in the dendrite of a neuron

In neurophysiology, a dendritic spike refers to an action potential generated in the dendrite of a neuron. Dendrites are branched extensions of a neuron. They receive electrical signals emitted from projecting neurons and transfer these signals to the cell body, or soma. Dendritic signaling has traditionally been viewed as a passive mode of electrical signaling. Unlike its axon counterpart which can generate signals through action potentials, dendrites were believed to only have the ability to propagate electrical signals by physical means: changes in conductance, length, cross sectional area, etc. However, the existence of dendritic spikes was proposed and demonstrated by W. Alden Spencer, Eric Kandel, Rodolfo Llinás and coworkers in the 1960s and a large body of evidence now makes it clear that dendrites are active neuronal structures. Dendrites contain voltage-gated ion channels giving them the ability to generate action potentials. Dendritic spikes have been recorded in numerous types of neurons in the brain and are thought to have great implications in neuronal communication, memory, and learning. They are one of the major factors in long-term potentiation.

<span class="mw-page-title-main">Summation (neurophysiology)</span>

Summation, which includes both spatial summation and temporal summation, is the process that determines whether or not an action potential will be generated by the combined effects of excitatory and inhibitory signals, both from multiple simultaneous inputs, and from repeated inputs. Depending on the sum total of many individual inputs, summation may or may not reach the threshold voltage to trigger an action potential.

Cellular neuroscience is a branch of neuroscience concerned with the study of neurons at a cellular level. This includes morphology and physiological properties of single neurons. Several techniques such as intracellular recording, patch-clamp, and voltage-clamp technique, pharmacology, confocal imaging, molecular biology, two photon laser scanning microscopy and Ca2+ imaging have been used to study activity at the cellular level. Cellular neuroscience examines the various types of neurons, the functions of different neurons, the influence of neurons upon each other, and how neurons work together.

Retinal waves are spontaneous bursts of action potentials that propagate in a wave-like fashion across the developing retina. These waves occur before rod and cone maturation and before vision can occur. The signals from retinal waves drive the activity in the dorsal lateral geniculate nucleus (dLGN) and the primary visual cortex. The waves are thought to propagate across neighboring cells in random directions determined by periods of refractoriness that follow the initial depolarization. Retinal waves are thought to have properties that define early connectivity of circuits and synapses between cells in the retina. There is still much debate about the exact role of retinal waves. Some contend that the waves are instructional in the formation of retinogeniculate pathways, while others argue that the activity is necessary but not instructional in the formation of retinogeniculate pathways.

An autapse is a chemical or electrical synapse from a neuron onto itself. It can also be described as a synapse formed by the axon of a neuron on its own dendrites, in vivo or in vitro.

<span class="mw-page-title-main">Heterosynaptic plasticity</span>

Synaptic plasticity refers to a chemical synapse's ability to undergo changes in strength. Synaptic plasticity is typically input-specific, meaning that the activity in a particular neuron alters the efficacy of a synaptic connection between that neuron and its target. However, in the case of heterosynaptic plasticity, the activity of a particular neuron leads to input unspecific changes in the strength of synaptic connections from other unactivated neurons. A number of distinct forms of heterosynaptic plasticity have been found in a variety of brain regions and organisms. These different forms of heterosynaptic plasticity contribute to a variety of neural processes including associative learning, the development of neural circuits, and homeostasis of synaptic input.

Frank Werblin is Professor of the Graduate School, Division of Neurobiology at the University of California, Berkeley.

<span class="mw-page-title-main">Brain cell</span> Functional tissue of the brain

Brain cells make up the functional tissue of the brain. The rest of the brain tissue is structural or connective called the stroma which includes blood vessels. The two main types of cells in the brain are neurons, also known as nerve cells, and glial cells, also known as neuroglia.

References

  1. Hikosaka, R; Takahashi M; Takahata M (1996). "Variability and invariability in the structure of an identified nonspiking interneuron of crayfish as revealed by three-dimensional morphometry". Zoological Science. 13 (1): 69–78. doi: 10.2108/zsj.13.69 .
  2. 1 2 Brunn, DE (1998). "Cooperative mechanisms between leg joints of Carausius morosus - I. Nonspiking interneurons that contribute to interjoint coordination". Journal of Neurophysiology. 79 (6): 2964–2976. doi:10.1152/jn.1998.79.6.2964.
  3. 1 2 Namba, H (1994). "Descending control of nonspiking local interneurons in the terminal abdominal ganglion of the crayfish". Journal of Neurophysiology. 72 (1): 235–247. doi:10.1152/jn.1994.72.1.235. PMID   7965008.
  4. 1 2 3 4 5 6 DiCaprio, Ralph (2004). "Information Transfer Rate of Nonspiking Afferent Neurons in the Crab". Journal of Neurophysiology. 92 (1): 302–310. doi:10.1152/jn.01176.2003. PMID   14973322.
  5. Purves, Dale; Augustine, George J.; Fitzpatrick, David; Katz, Lawrence C.; LaMantia, Anthony-Samuel; McNamara, James O.; Williams, S. Mark (2001). Phototransduction.
  6. Nagayama, T; Namba H; Aonuma H (1997). "Distribution of GABAergic premotor nonspiking local interneurones in the terminal abdominal ganglion of the crayfish". Journal of Comparative Neurology. 389 (1): 139–148. doi:10.1002/(sici)1096-9861(19971208)389:1<139::aid-cne10>3.0.co;2-g.
  7. 1 2 3 Zhou, ZJ; Cheney M; Fain GL (1996). "Starburst amacrine cells change from spiking to non-spiking neurons during visual development". Investigative Ophthalmology & Visual Science. 37 (3): 5263.
  8. Bloomfield, Stewart A.; Völgyi, Béla (July 2007). "Response properties of a unique subtype of wide-field amacrine cell in the rabbit retina". Visual Neuroscience. 24 (4): 459–469. doi:10.1017/S0952523807070071.
  9. Park, Silvia JH; Lieberman, Evan E; Ke, Jiang-Bin; Rho, Nao; Ghorbani, Padideh; Rahmani, Pouyan; Jun, Na Young; Lee, Hae-Lim; Kim, In-Jung; Briggman, Kevin L; Demb, Jonathan B; Singer, Joshua H (15 May 2020). "Connectomic analysis reveals an interneuron with an integral role in the retinal circuit for night vision". eLife. 9: e56077. doi: 10.7554/eLife.56077 . ISSN   2050-084X. PMC   7228767 . PMID   32412412.
  10. Namba, H; Nagayama T (2004). "Synaptic interactions between nonspiking local interneurones in the terminal abdominal ganglion of the crayfish". Journal of Comparative Physiology A. 190 (8): 615–622. doi:10.1007/s00359-004-0516-5.
  11. Rela, Lorena; Yang, Sung Min; Szczupak, Lidia (26 November 2008). "Calcium spikes in a leech nonspiking neuron". Journal of Comparative Physiology A. 195 (2): 139–150. doi:10.1007/s00359-008-0393-4. PMID   19034463.
  12. Hayashida, Y; Yagi T (2002). "Contribution of Ca2+ transporters to electrical response of a non-spiking retinal neuron". Neurocomputing. 44: 7–12. doi:10.1016/s0925-2312(02)00354-5.
  13. Zipser, D; Kehoe B; Littlewort G; Fuster J (1993). "A spiking network model of short-term active memory". Journal of Neuroscience. 13 (8): 3406–3420. doi: 10.1523/jneurosci.13-08-03406.1993 . PMC   6576542 . PMID   8340815.
  14. Christodoulou, C; Banfield G; Cleanthous A (2010). "Self-control with spiking and non-spiking neural networks playing games". Journal of Physiology-Paris. 104 (3–4): 108–117. doi:10.1016/j.jphysparis.2009.11.013. PMID   19944157.
  15. Vassiliades, Vassilis; Cleanthous, Christodoulou (2011). "Multiagent Reinforcement Learning: Spiking and Nonspiking Agents In the Iterated Prisoner's Dilemma". IEEE Transactions on Neural Networks. 22 (4): 639–653. doi:10.1109/tnn.2011.2111384.
  16. Lo, JT-H (2011). "A low-order model of biological neural networks". Neural Computation. 23 (10): 2626–2682. arXiv: 2012.06720 . doi:10.1162/neco_a_00166.