Molecule-based devices are envisioned to complement silicon devices by providing new functions or already existing functions at a simpler process level and at a lower cost by virtue of their self-organization capabilities, moreover, they are not bound to von Neuman architecture and this may open the way to other architectural paradigms. Here we demonstrate a device made of conjugated molecules and metal nanoparticles (NPs) which behaves as a spiking synapse suitable for integration in neural network architectures. We demonstrate that this device exhibits the main behavior of a biological synapse. These results open the way to rate coding utilization of the NOMFET in perceptron and Hopfield networks. We can also envision the NOMFET as a building block of neuroelectronics for interfacing neurons or neuronal logic devices made from patterned neuronal cultures with solid-state devices and circuits.
Deep Dive into An artificial spiking synapse made of molecules and nanoparticles.
Molecule-based devices are envisioned to complement silicon devices by providing new functions or already existing functions at a simpler process level and at a lower cost by virtue of their self-organization capabilities, moreover, they are not bound to von Neuman architecture and this may open the way to other architectural paradigms. Here we demonstrate a device made of conjugated molecules and metal nanoparticles (NPs) which behaves as a spiking synapse suitable for integration in neural network architectures. We demonstrate that this device exhibits the main behavior of a biological synapse. These results open the way to rate coding utilization of the NOMFET in perceptron and Hopfield networks. We can also envision the NOMFET as a building block of neuroelectronics for interfacing neurons or neuronal logic devices made from patterned neuronal cultures with solid-state devices and circuits.
made of conjugated molecules and metal nanoparticles (NPs) which behaves as a spiking synapse suitable for integration in neural network architectures. We demonstrate that this device exhibits the main behavior of a biological synapse. The device (Fig. 1A) consists of a bottom-gate, bottom source-drain contact organic transistor configuration. The gold NPs (5 nm in diameter) were immobilized into the source-drain channel using surface chemistry (selfassembled monolayers) and they were subsequently covered by a thin film (25-35 nm thick) of pentacene (see supporting online material). This device gathers the behavior of a transistor and a memory (2) and it is referenced to as NOMFET (Nanoparticle Organic Memory Field Effect Transistor).
The most important feature of a synapse is its ability to transmit in a given way an action potentials (APs) from one pre-synapse neuron N1, to a post-synapse neuron N2. When a sequence of APs is send by N1 to N2, the synaptic behavior determines the way the information is treated. The synapse transforms a spike arriving from the presynaptic neurone into a chemical discharge of neurotransmitters that will be detected by the post-synaptic neurone and transformed into a new spike. Markram and Tsodyks (3,4) have proposed a phenomenological model to describe the synapse behavior. The synapse possesses a finite amount of resources: the chemical neurotransmitters. Each spike activates a fraction of these resources and the amplitude of the transmitted spike is proportional to this fraction. The fraction of neurotransmitters spend to transmit the information is then recovered with a characteristic time τrec that is typically in the range of the second. The response of a synapse to a train of pulses with variable frequency can be calculated by an iterative model (5) which describes the biological synapse behavior reasonably well (Fig. 1B). The main feature of a biological synapse is to present a dependence of the amplitude of the output spike with the frequency of the input spike. It also depends on the history of the synapse which determines the amount of available neurotransmitter at a given time. Such a typical behavior is shown in Fig. 1B: at high(low) frequency, the period of the input signal is lower(larger) than τrec and the output signal decreases(increases) at each successive pulse generating a depressing(facilitating) behavior (Fig. 1B). We used the NOMFET as a “pseudo two-terminal device”. The gate receives the same input voltage (a train of pulse at frequency 1/T, amplitude V, and pulse width W) as the source electrode. The output is the drain current (Fig. 1A). We measured the response of the NOMFET to sequences of pulses with different periods, T (Fig. 1C). During such experiments, the NPs are alternatively charged during the pulse duration and discharged between pulses (2).
The value of the current at a certain time depends on the full history of the device that determines the amount of charges presents in the NPs. To illustrate this point let us consider the system at the beginning of a particular sequence with period T (Fig. 1C), where the NPs contain some charges. If T«τd (τd is the NP discharge time constant of about 20 s here), more and more holes are trapped in NPs and the NOMFET presents a depressing behavior. Then, for a larger period T (Fig. 1C), NPs have enough time to be discharged between pulses and the sequence presents a facilitating behavior. This feature exactly reproduces the behavior of a biological synapse. The holes trapped in the NPs play the role of the neurotransmitters and the output signal, ID, is a decreasing function of the number of holes stored in the NPs (2). At each spike, a certain amount of holes are trapped in the NPs. Between pulses the system relaxes: the holes escape with a characteristic time τd. This behavior persists when shrinking the NOMFET to a source-drain channel of 1 µm (Fig. 1D). In that case, we used a constant dc bias on the common gate (due to channel reduction, the lateral field is sufficient to charge and discharge the NP without the need to applied the pulse on the gate electrode). This features would simplify the architecture design of neural networks using NOMFETs since a separate gate is not required for each NOMFET. Note that the depressing/facilitating behaviors are now inverted (with respect of the frequency of the pulses) and the time constant decreased-see supporting online material. Finally, such a synaptic behavior was not observed for the reference pentacene OFET (no NPs) -Fig. S2.
These results open the way to rate coding utilization (6) of the NOMFET in perceptron and Hopfield networks (7). We can also envision the NOMFET as a building block of neuroelectronics for interfacing neurons or neuronal logic devices made from patterned neuronal cultures with solid-state devices and circuits (8,9). The NOMFET electrical characteristics were measured with an Agilent 4155C semiconductor parameter analyzer, the
…(Full text truncated)…
This content is AI-processed based on ArXiv data.
