JP4209781B2 - Event detection apparatus and measurement method for measuring neural network activity - Google Patents

Description

Detailed Description of the Invention

The present invention relates to a device for measuring the activity of a neural network according to claim 1 and a method for measuring the activity of a neural network according to claim 11 .

  In particular, the main field of application of the apparatus and method of the present invention is to analyze chemical substances with high sensitivity and speed and to evaluate the toxicity of unknown substances quickly and reliably.

  Usually, conventional methods for analyzing the toxicity of chemical substances are based on conventional wet chemical analysis methods, spectroscopic measurement methods, and physicochemical measurement methods. In such an analysis method, the structure of the apparatus is usually relatively complicated, the analysis time is long, and / or the detection sensitivity is usually unsatisfactory. In particular, in the conventional analysis method, the detection threshold (Nachweisschwelle) of the substance or substance mixture to be analyzed is relatively high.

First, neural networks are also used in the prior art, especially for analyzing the toxicity of chemical substances. In this case, the so-called “patch clamping” technique or extra-cellular microelectrodes is usually used to detect neuronal activity of the neural network (neuronalen Aktivitaet). In the patch clamp method, the signal inside the cell is guided by invasiveness, while the signal outside the cell is guided by the noninvasive detection method in the microelectrode outside the cell. In both cases, an external measurement amplifier is connected to the detection electrode and the corresponding voltage over time is recorded. This external measuring amplifier therefore sends out the voltage-time signal transition of the neuronal signal (so-called Transienten), which is later evaluated.

  It is known to use regularly arranged sensors in order to increase the yield (Ausbeute) and the statistical significance (Aussagekraft) from appropriate nerve cells in proper contact. In this publication (Journal of Neuroscience Methods 93 (1999), pp. 61-67, Elsevier), Oka et al., “A new flat multi-electrode array for off-cell recording: the tip of the hippocampus. "A new planar multielectrode array for extracellular recording: Application to hippocampal acute slice".

  Such a sensor structure is capable of signal detection that is resolved over time in individual neurons on the sensor, but in particular to reliably and quickly assess the toxicity of unknown substances. There are limits to use.

  Accordingly, it is an object of the present invention to provide an apparatus and method that can quickly and reliably assess the toxicity of a substance or substance mixture.

This object is achieved by an apparatus for measuring neural network activity according to claim 1 and a method for measuring neural network activity according to claim 11 . Preferred embodiments are described in the dependent claims.

In the present invention, an apparatus for measuring the activity of a neural network is an apparatus for measuring the activity of a neural network using a patterned semiconductor substrate .
The semiconductor substrate is
A plurality of sensor elements each having at least one conductive sensing electrode (E) ;
A plurality of amplifying elements (V; TKV) each having at least one amplifying input and at least one amplifying output (VA) ;
A plurality of event memories (ES) for storing a count reading which is a measure of the number of events in the neuronal cell, and at least one memory reading device for reading the count reading from the event memory (ES); at least one evaluation input unit comprises at least one activity evaluation unit and at least one evaluation output unit,
The conductive detection electrode is disposed on the surface of the semiconductor substrate to detect a nerve cell signal from the neural network, and the sensor element is an electric sensor output signal based on the detected nerve cell signal. Can be output via each sensor output part (SA) of the sensor element,
Each of the sensor elements is associated with one of the amplifying elements (V; TKV) , and the amplifying input section of the amplifying element is electrically connected to the sensor output section (SA) of each sensor element. The output signal of the electric sensor can be output as an amplified output signal via an amplification output unit (VA) ,
The evaluation input unit (digital out; analog out) is electrically connected to at least one of event memory (ES) , and the activity evaluation unit receives an activity signal that is a measure of activity in the neural network as an event. It is generated based on the memory contents read from the memory (ES) and output via the evaluation output unit.
Each of the amplifying elements (V; TKV) is connected to an associated event memory (ES) via a threshold value detecting element (SD) for discretizing the amplified output signal depending on whether or not the amplified output signal is larger than a predetermined threshold value. Connected,
Each of the amplifying elements (V; TKV) is connected to an associated threshold value detecting element (SD) via a rectifying element (GR) for rectifying the amplified output signal,
The event memory (ES) is a digital memory element, specifically a digital counter,
One neuronal event is an amplified output signal rectified by the rectifier element that is greater than a predetermined threshold of the threshold detector element that drives the digital counter to increase the count reading by one. .

  Thus, the device of the present invention for measuring or measuring neural network activity is a sensor structure monolithically integrated on a semiconductor substrate, which is a direct measure of the activity of the neural network to be investigated. Provide an activity signal that is (directes Mass). For this reason, the device comprises a plurality of sensor elements or signal pickups designed to detect nerve cell signals from nerve cells. An amplification element is directly associated with each of these sensor elements. This amplifying element amplifies the sensor output signal (usually a voltage signal) of the sensor element. The sensor output signal amplified in this way is called the amplified output signal and can be read out by the activity evaluation unit. In this case, the activity evaluation unit is designed to be able to generate an activity signal that is a measurement value for the activity of the neural network based on the amplified output signal.

  The present invention takes advantage of the recognition that all rapidly acting toxic substances or substance mixtures used for example in drugs or biological or chemical warfare drugs attack the nervous system. The action of such substances or substance mixtures is particularly pronounced as a change in the electrical activity of the neural network. In this case, an important parameter of the electrical activity of the neural network is the frequency of the electrical pulses generated by the neural cell. The generation of electrical pulses by nerve cells (also called nerve cell “Feuern”) manifests itself as a temporary increase in conductivity, especially due to potassium and sodium currents between and outside the cell . As a result, the potential changes abruptly in the gap between the support structure (aufliegenden Anordnung) of the nerve cell and the detection electrode E. This support structure extends from individual nerve cell cells (one unit), through a stack of nerve cell cells, to a very dense nerve cell tissue layer (units). The frequency spectrum of such a signal usually has a bandwidth of up to 5 kHz, and the amplitude of the nerve cell signal detectable at the detection electrode E is on the order of 1 mV (Groessenordnung). When a neurotoxic substance is applied, the pulse frequency of the nerve cell changes in particular. The fact that the frequency or correlation of the pulse from the nerve cell changes in this way (bereits) can be used to estimate the amount and type of nerve cell active substance.

  Such universally usable sensor structures have received much attention, especially in the fields of safety technology and military fields, as well as in pharmacology and environmental management. For example, sensor structures that allow users to quickly and reliably investigate unknown gases or liquids for their toxic effects will have a variety of potential uses.

In this case, the device of the present invention effectively overcomes the problems associated with identifying neural network activity in the above-described conventional sensor structures (particularly microelectrode arrays). As explained in the introduction, a conventional sensor structure transmits a voltage-time-signal transition (so-called transient) related to a neuronal event from the sensor structure to an external amplifier for amplification, and further evaluation. Yes. However, a number of such transients need to be transmitted and evaluated so that a reliable and statistically meaningful evaluation of neural network activity can be obtained. This results in an “imaging-problem” and a transmission bottleneck (Uebertragungsengpaesse) when transmitting a wide range of transient data to an external instrumentation amplifier or external evaluation operator .

The voltage signal generated by the neuron during "ignition" is typically characterized by a frequency spectrum with a bandwidth of 5 kHz, so a transient must be recorded at an Abtastraten of 10 kHz or higher There is. For example, when it is necessary to recognize a microelectrode structure having 100 × 100 = 104 electrodes with a resolution of 8 bits, a data rate (Datenvolumen) of 800 megabits / second can be obtained because the scanning rate is just 10 kHz. These need to be transmitted to an external instrumentation amplifier or an external computing unit . Therefore, a very complex broadband bus architecture (Busarchitektur) is required, which significantly complicates the sensor structure and increases costs.

  The present invention preferably solves this problem by reducing data using "event characterization" in the device itself. In the present invention, transient data is not transmitted. That is, the voltage-time-signal transition of the nerve cell signal is not transmitted from the sensor element to the activity evaluation unit. Instead, the event is actually evaluated and characterized in the sensor cell. As a result, event information that is significantly compressed compared to the transient data is generated. Therefore, the device of the present invention functions to detect (overall) activity of the neural network, and is a direct and time continuous signal regarding the transients detected by the individual sensor elements. Does not function to track.

  Compared to currently marketed multi-electrode arrays (MEA), the present invention further increases the density of the detection electrodes on the sensor surface, which is manufactured by utilizing manufacturing methods known in microelectronics. Can be made.

  Detection of neuronal signals abgreift on or near the surface of a semiconductor substrate on a capacitive-resistiv basis for signal induction of neuronal signals An electrode is provided. The detection electrode may be covered with a dielectric provided that pure capacitive coupling to the nerve cell is provided. The detection electrode preferably has an electrode area that is approximately the same size or smaller than the normal mounting area (Auflageflaeche) of nerve cells on the sensor surface. When using murine neurons, for example, the preferred electrode diameter is about 10 μm.

  When neurons in the neural network fire, conductivity increases temporarily due to potassium and sodium currents between the cell interior and cell exterior. As a result, the potential changes rapidly in the gap between the support structure of the nerve cell and the detection electrode. This structure extends from individual neuronal cells (one unit) through a stack of multiple neuronal cells to a very dense neuronal tissue layer (multiple units). The frequency spectrum of such a signal usually has a bandwidth up to 5 kHz, and the amplitude of the signal detectable at the detection electrode is on the order of 1 mV.

  An amplifying element connected to the output of the sensor element preferably amplifies this neuronal signal by some order and / or converts it into an output signal.

  In a preferred embodiment of the present invention, a plurality of sensor cells are provided, each sensor cell comprising one sensor element and an amplifying element associated with the sensor element. The sensor cells are preferably arranged in a matrix, in particular an orthogonal matrix, in order to form a sensor cell array or matrix. Therefore, the amplification element is directly associated with each sensor element. This amplifying element is arranged in the same sensor cell, i.e. the physical immediate vicinity of the sensor element. For example, if at least a part of the amplifying element is located below the detection electrode in the normal direction of the semiconductor substrate, a sensor cell structure capable of high integration can be obtained as a result. The matrix structure of such sensor cells relative to the sensor cell array is similar to the arrangement structure known in DRAM.

  In a preferred embodiment, the activity evaluation unit comprises a plurality of event memories for storing neuronal events and at least one memory reading device for reading from the event memory. Preferably, the activity evaluation unit further comprises a control unit that can reset the event memory or cause a “reset”.

  Each of the sensor cells preferably includes one of the event memories, and the event memory input section is connected to the amplification output section. In this case, the electrical connection between the amplification output and the event memory input is mediated by intermediately connected components, such as other electrical signal connections that are within the scope of the present invention. Also good. As a result, in this embodiment, neuronal events or event information is stored or buffered in each sensor cell of the sensor cell array. Further, neuronal events may be further stored outside the sensor cell.

  Each of the event memories is preferably an activity evaluation unit that can be selected via a select line in the activity evaluation unit for selective reading by the memory reading device. The event memory read operation is performed by a method similar to, for example, a DRAM memory device. After properly addressing the sensor cell (or sensor cell event memory), the contents of the event memory may be requested in the column wiring or evaluation wiring.

  In a preferred embodiment, the activity evaluation unit reads the event memory at a predetermined cycle, and generates an activity signal based on the memory contents read from the event memory every cycle. If the event memory is a digital memory module (digitalen Speicerbaustein), the memory content is preferably interpreted as the number of stored neuronal events (“events”). Preferably, each neuronal event that appears as a complex voltage-time-signal transition (transient) is represented only by certain binary information items. Therefore, the transient data is data compressed to generate binary data. If the event memory is an analog memory element, the memory content may be a transient signal that is integrated over time, for example.

  Each amplifying element is preferably connected to an associated event memory via a threshold detection element (Schwellwertdetektorelemente) for discretizing the amplified output signal. The amplification element of each sensor cell is preferably connected to a threshold detection element on the downstream side. When the amplified output signal exceeds the threshold value of the threshold detection element, the threshold detection element outputs a pulse signal or an activation signal (Triggersignal). It is preferable that the threshold value of the threshold value detection element can be set from the outside by “threshold” wiring using an activity evaluation unit.

  It is preferable that each amplifying element is connected to an associated threshold value detecting element via a rectifying element for rectifying the amplified output signal. Such a rectifying element is advantageous because the polarity of the nerve cell signal does not contain information essential to assessing the activity of the neural network. Therefore, first rectifying the amplified output signal is suitable for the purpose.

In a preferred embodiment, the event memory element is a digital memory element, in particular a digital counter . In this case, the counter is advantageously designed to increase the count reading, for example, by a voltage pulse output by a threshold detection element connected upstream. Thus, the count reading is a measure for the number of neuronal events, ie the number of neuronal “events” in the query cycle.

In other preferred embodiments, the event memory element is an analog memory element, in particular an analog integrator , or an analog minimum or maximum memory.

  In another preferred embodiment, the amplifying element is a transconductance amplifying element (Transkonduktanzverstaerkerelemente) for generating a current signal as an amplified output signal. In this case, the voltage signal applied to the amplification input unit of the amplification element is converted into a current signal and output from the amplification output unit for other processing.

  In another preferred embodiment, the amplifying element is a transconductance amplifying element for generating a current signal as an amplified output signal. It is preferable that at least two amplification output units of the amplification element are connected to the activity evaluation unit through one evaluation wiring or output wiring. As a result, current signals from the two amplifying elements are collected. The activity evaluation unit is preferably adapted to generate an activity signal based on the combined amplitude of the current signal.

Therefore, in this embodiment, each sensor cell includes a transconductance amplifier . This transconductance amplifier amplifies the detected nerve cell signal voltage and converts it (preferably linearly) into a current. This current is combined into a total current on a common evaluation wiring via parallel connection of sensor cells, and is evaluated at the sensor cell array or matrix edge (Rande). The amplitude of the current that varies with time is a direct measure of neural network activity and represents a possible activity signal.

  In this method, neuronal events are not differentiated inside the sensor cell, so that these cells can be easily constructed and therefore have a small area. Such a concept enables high-level integration of sensor cells, and as a result, a sensor cell array having a very high integration degree can be manufactured. However, a disadvantage of this concept is that each connected sensor cell forms a noise contributor (Rauschbeitrag) that aggregates into the evaluation wiring and reduces the signal-noise ratio.

  It is preferable that the amplification output portions of all the amplification elements are connected to the activity evaluation unit via one evaluation wiring.

  Advantageously, each amplifying element can be connected to the activity evaluation unit by one evaluation wiring via a threshold detection element that discretizes the amplified output signal and a downstream reference current source.

  In another particularly preferred embodiment, at least two of the sensor cells are interconnected such that at least one signal of the sensor cells can be exchanged between the sensor cells. This signal may be any analog or digital signal generated in the sensor cell, in particular a sensor output signal, an amplified output signal or other such signal. In this way, the sensor cells can also exchange (event) information directly with each other. Directly adjacent sensor cells preferably exchange analog or digital information about neuronal events. This is to further improve the detection sensitivity and / or to make the sensor cell smaller and / or to pre-record the recorded information.

In another aspect of the invention, a method for measuring neural network activity using the apparatus of the invention is provided. This method
Detecting a nerve cell signal using a plurality of sensor elements;
Generating and outputting a sensor output signal according to the detected nerve cell signal;
Amplifying the sensor output signal by each amplifying element for generating an amplified output signal;
Generating an activity signal that is a rate of activity in the neural network in response to the amplified output signal.

  The present invention is described in detail below with reference to the accompanying drawings of preferred embodiments. It will be understood that individual features described in connection with one embodiment may be used in other embodiments.

FIG. 1 is a schematic block diagram of a sensor cell according to the first embodiment of the present invention. FIG. 2 is a schematic block diagram of a sensor cell according to the second embodiment of the present invention. FIG. 3 is a schematic block diagram of a third embodiment as a reference embodiment of the present invention. FIG. 4a is a schematic block diagram of a fourth embodiment as a reference embodiment of the present invention. FIG. 4b is a schematic block diagram of the fourth embodiment of the present invention. FIG. 5 is a schematic block diagram of the fifth embodiment of the present invention. FIG. 6 is a schematic block diagram of the sixth embodiment of the present invention. FIG. 7 shows an overview of the classification of the preferred embodiment of the device of the present invention for measuring neural network activity. FIG. 8 is a diagram showing an embodiment of a preferred amplifying element of the present invention having a PMOS transistor. FIG. 9a is a diagram illustrating a preferred embodiment of a rectifying element and a threshold detection element. FIG. 9b shows another preferred embodiment of a preferred rectifying element and threshold detection element. FIG. 10 is a diagram showing an embodiment as a preferred reference form of the analog integrator and the analog extreme value memory. FIG. 11 is a diagram showing a preferred structure of a matrix type sensor cell array, in which each sensor cell can be driven and read by various row wirings and column wirings extending over the entire structure. FIG. 12 is a schematic block diagram of a preferred sensor cell array where the analog current output signals are grouped and amplified at the edge of the sensor cell array. FIG. 13 shows another preferred embodiment of a matrix type sensor cell array, where each cell can be driven and read by various row and column wirings extending throughout the structure, Various data can be exchanged between sensor cells.

  FIG. 7 shows an overview of the classification of the preferred embodiment described below with respect to the sensor cell in the apparatus of the present invention. Here, these embodiments are classified into a total of six embodiments based on the form of event storage and event information transmission. It will be appreciated that features described with respect to one embodiment may be used with respect to other embodiments.

  FIG. 1 shows a schematic block diagram of a preferred sensor cell SZ according to the first embodiment of the apparatus of the present invention. This sensor cell SZ is part of a monolithically integrated semiconductor structure, in particular a silicon CMOS structure. A plurality of sensor cells SZ arranged in rows and columns, preferably orthogonally, form a sensor cell array. This sensor cell array is similar to a DRAM memory cell array in that it has a matrix structure.

  The sensor cell SZ includes a sensor element. The sensor element includes a conductive detection electrode E for interacting with a nerve cell signal from a neural network (not shown) or detecting the nerve cell signal. Further, an electrical stimulation signal may be output to the neural network via this detection electrode. As a signal pickup or sensor element for a nerve cell signal, various sensors or transducers that can detect the electrical activity (so-called action potential) of the nerve cell may be used. In the simplest case, the detection electrode E of the sensor element is an “open” metal electrode (for example a gold electrode) that is arranged on the surface of a patterned semiconductor substrate. The detection electrode E may come into contact with a solution or an electrolytic solution (neural network). In that case, electrical connection with nerve cells occurs. However, the detection electrode E may be a conductive electrode covered with a dielectric.

  The electrode area of the detection electrode E is preferably approximately the same size or smaller than the normal mounting area (Auflageflaeche) of neurons on the sensor surface. For example, when using murine neurons, the preferred electrode diameter is about 10 μm. Nerve cell “firing” manifests itself as a temporary increase in electrical conductivity to potassium and sodium currents between and inside the cell. As a result, the potential changes rapidly in the gap between the support structure of the nerve cell and the detection electrode E. This support structure extends from an individual neuronal cell (1 unit), through a stack of multiple neuronal cells, to a very dense neuronal tissue layer (multiple units). The frequency spectrum of such a signal usually has a bandwidth of up to 5 kHz, and the amplitude of the nerve cell signal detectable at the detection electrode E is on the order of 1 mV (Groessenordnung).

The amplifier V connected to the sensor output unit SA of the sensor element normally amplifies the signal by several digits and outputs it as an amplifier output signal for further signal processing via the amplification output unit VA. Both negative and positive voltage pulses can be detected outside the cell during the “fire” period of the neuron. It is advantageous to first rectify the amplifier output signal because the polarity of the pulse does not contain the information necessary to evaluate the activity of the neural network. For this purpose, a rectifying element GR is provided. The rectification input section of the rectifying element GR is electrically connected to the amplification output section VA.

In order to record the activity of the neural network as completely as possible, it is advantageous that the density of the detection electrodes is high. Therefore, it is necessary to reduce the size (Abmessungen) of each sensor cell. Therefore, it is necessary to use a component having a relatively small area (particularly a transistor having a small area) in the sensor cell SZ. As a result, a relatively large noise level (1 / f noise) occurs in the amplifier V. In order to extract an event of a nerve cell from noise, a threshold detection element SD is provided after rectification by the rectification element GR. The threshold value of the threshold value detection element SD is preferably settable from the outside (vorgebbar). For this reason, in particular, a control unit provided with a row wiring “threshold” for connecting the threshold detection element SD to a control unit (not shown). This control unit may be part of an activity evaluation unit. In short (Abstrahiert), the threshold detection element may be regarded as an analog-digital converter having a word length (Wortbreite) of 1 bit in the output section.

When the threshold detector element SD detects a neuronal event, i.e. when the rectified amplified output signal is greater than a predetermined threshold, the threshold detector output signal drives a digital counter , e.g. The degree is increased by 1 depending on the received signal. Therefore, the digital counter is an event memory ES that stores neuronal events. The count reading of the event memory ES may be applied to the digital output wiring (for example, the column wiring “digital out”) by using a plurality of “select” wiring and “reset” wiring. A memory reading device (not shown) in the activity evaluation unit can read the count reading from the event memory ES of each sensor cell SZ in this way. After successful reading, the event memory ES can be reset using the “reset” wiring. In the simplest case, the event memory ES is a latch that stores whether (at least) one neuron event has occurred in two readout intervals.

  The count reading read from the cell is recorded at the edge of the sensor cell array and further processed. In this case, this reading process is performed by addressing the sensor cell SZ using the selection wiring and the signal wiring (word line / bit line) as in the case of the DRAM. As a result, continuous digital reading is performed at periodic time intervals. The activity of the neural network is obtained from the number of events per unit time stored in the event memory ES of the sensor cell SZ.

  Therefore, in order to measure the activity of the neural network, it is not necessary to transmit the transient of the nerve cell voltage, that is, the voltage-time transition of the nerve cell signal. Instead, “event characterization” is performed directly in each sensor cell SZ. As a result, it is only necessary to transmit digital data for each detected neuronal event.

  Another advantage of this embodiment is that the event information is actually read from the sensor cell SZ and available in digital form for further processing. However, this requires relatively complex circuit technology for the sensor cell SZ. As a result, the area of the sensor cell SZ is relatively large, and therefore the cell density of the sensor cell array is relatively low. As a result, only some of the neurons that are statistically distributed on the sensor cell array and sometimes (unter Umstaenden) move freely (frei beweglichen) can be recognized (messtechnisch).

A second embodiment derived from FIG. 1 is shown in FIG. The same or similar members are given the same reference numerals and will not be described again. In the second embodiment, the reading of the counter (event memory ES) is not directly digitally output but is converted into an analog signal, that is, an analog voltage or current signal by the digital-analog converter DAC. The analog signal is applied to the column wiring “analog out” of the sensor cell array for output to the activity evaluation unit. In the following description, the reference symbol “analog out” indicates a correspondence evaluation input unit of the activity evaluation unit.

The advantage in this case is that a plurality of sensor cell arrays, preferably all sensor cells can be read simultaneously. If the output signal of the DAC is a current, the output currents from all connected sensor cells are combined in the column wiring (“analog out”). Therefore, at the edge of the sensor cell array, the sum of the counter or event memory contents in the corresponding sensor cell array column can be directly measured as an analog signal. Further, all the column wirings (“analog out”) may be connected to each other to form one evaluation wiring. As a result, the count readings or memory contents of all event memories ES in the matrix sensor cell array can be summarized in analog form and made available at the edge of the matrix as a result for other processing. In the simplest case, the event memory ES is a latch that stores whether (at least) one event has occurred in the read interval, and the DAC is divided into two different reference currents Ievent and Ino_event depending on the contents of the latch. Is a reference current source.

  One of the two reference currents Ievent or Ino_event is preferably equal to zero. Therefore, the total current Isum = ΣiΣjIij can be directly detected at the edge of the matrix or sensor cell array. The amplitude of the current, which varies with time, is a direct measure of the activity of the neural network and is therefore a real indication of the activity signal. It is also preferred that this activity signal be amplified and / or reshaped for other processing or output via the evaluation output of the activity evaluation unit.

FIG. 3 shows a third embodiment as a reference form of the current sensor cell SZ having the associated readout wiring and control wiring, and a modification thereof. In this figure, event information from nerve cells is stored in an analog format in an analog event memory ES and output in a digital format. In the embodiment shown on the upper side of FIG. 3, the sensor output signal amplified in the transconductance amplifier KTV is stored in an analog integrator forming an analog event memory ES.

In the form without the rectifying element shown here, the noise is gefiltered by an integrator . Conversely, neuronal signals are typically positive or negative pulses, thus significantly changing the contents of the integrator . On the other hand, if the neuronal signal is perfectly symmetric, the integrator will not remember anything. In this case, it is preferable to use a rectifying element. As a result, the above nerve cell signal is also detected. However, in this case, the noise signal rectified in the same way becomes an undesirable offset in the analog event memory ES. The content of the event memory may be read out as a digital value from the sensor cell SZ using the analog-digital converter ADC. The row wiring (“select / reset”) can be used to initialize the reading of the memory contents of the event memory ES and reset the event memory ES.

In the embodiment shown in the lower part of FIG. 3, a circuit for storing the maximum value and the minimum value of the amplified output signal is provided instead of the analog integrator as the event memory ES. This circuit stores in analog form the maximum and minimum values of the amplified output signal from the amplifier TKV that occur within the range of the read cycle. The connected analog-digital- converter ADC converts the peak-to-pwak-Wert of the sensor signal stored as described above into digital information (digitale information), and this digital information. Are made available on the output wiring ("digital out"). When reading is performed, the analog memory ES is reset. In the simplest form, only one maximum or minimum value memory can be used. The analog-to-digital converter ADC may be designed as a simple threshold detector .

4a and 4b show a fourth embodiment as a reference form for storing event information and outputting it in an analog format. The embodiment of FIG. 4a is derived from the embodiment of FIG. In this case, when the sensor cell SZ is read, the analog value stored in the integrator or the extreme value memory ES is directly input to the analog output wiring (“analog out”). Again, the output signal is advantageously in the form of a current. This is because in this case, a plurality of sensor cells SZ on the column wiring can be read simultaneously.

In the embodiment of FIG. 4b, the amplified and rectified signal is first sent via the threshold detection element SD. Thereafter, the information of the value discretized in this way (wertdiskretisierte Information) is stored in the integrator ES. Thereby, the influence of noise on the event memory contents of the event memory ES is greatly suppressed because the event memory ES is in the form of an integrator .

  FIG. 5 shows a fifth embodiment in which no event is stored in the sensor cell SZ. The amplified output signal is input directly to the ADC, which, after being activated by the “select” wiring, outputs its measured value to the digital output wiring or evaluation wiring.

FIG. 6 shows a sixth embodiment in which neither nerve cell events are stored nor sensor cells SZ to be read are selected. In the embodiment shown on the upper side of FIG. 6, the current output part of the transconductance amplifier TKV is directly connected to the corresponding column wiring that functions as the evaluation wiring. The current output signal from the amplifier TKV is combined with the current output signal from all of the other connected sensor cells on the column wiring (“analog out”). Therefore, the total current signal from the sensor cell can be ablated at the edge of the sensor cell array. The total current signal is a signal obtained by overlapping all individual signals or amplified output signals from the sensor cell (Ueberlagerung).

  The noise power from the n sensor cells connected to the column wiring is collected, and as a result, the current noise for all the columns is based on I (noise, column) = n1 / 2 × I (noise, cell). Obtained. In contrast, the nerve cell signals from the active network are strongly correlated. Therefore, the signal or evaluation current on the column wiring in the extreme case (when the signals are completely overlapped) is considered to be I (signal, column) = n × I (signal, cell).

In the embodiment shown on the lower side of FIG. 6, a threshold detection element SD is connected downstream of the transconductance amplifier TKV. It is preferable that the threshold value of the threshold detection element SD can be set from the outside via a “threshold” wiring. As long as the signal-to-noise ratio is sufficient and the threshold value is set appropriately, this threshold value detection element SD can be used to distinguish a noise signal from a neuronal event. The output signal from the threshold detection element drives the reference current source IREF. The reference current source IREF is connected to an output wiring or an evaluation wiring (“analog out”). Very generally, a circuit block is provided between the amplified output VA and the output wiring “analog out” to improve the signal-to-noise ratio or to separate the effective signal from the noise signal. .

FIG. 8 shows a voltage amplifier composed only of PMOS transistors. This is advantageous with regard to making the sensor cell SZ as small as possible. This is because if an NMOS transistor and a PMOS transistor are simultaneously used in a cell, well burying is required, and the required area increases. The amplifier circuit shown here consists of three stages, each stage having a constant voltage gain set for the (width / length) W / L ratio of the transistor. For example, if transistor M1 has a W / L ratio of 10: 1 and transistor M2 has a W / L ratio of 1:10, the first stage gain is approximately 10. If the other two substages in the amplifier have this W / L ratio, the gain factor for the entire circuit is approximately 1000. The switching transistor M7 connected to the capacitor C functions to set the operating point of the amplifier . When transistor M7 is turned on, a voltage is generated at the input node of the amplifier (between C and the gate of M1). This voltage ensures that the amplifier at the output does not uebersteuert and is critical to the operating voltage. In operation, if M7 is switched off, the voltage stored at the capacitive input node during the calibration period is maintained, and therefore the amplifier is operated at the optimum operating point.

  9a and 9b show a preferred embodiment of a current signal rectifier and threshold detector.

In FIG. 9a, a current mirror (Stromspiegeln) is used to generate copies and inverted copies of the current input signal (M1-M10). The signal processed in this way is passed alternately by the transistors M11 and M12 according to the priority order (durchleitet), and the connected threshold detector is evaluated. M13 functions as a constant current source (Konstantstromquelle) regarding the threshold value, while the circuit constituted by M14 to M17 operates as a zero current detector (Nullstrom-Detektor).

In FIG. 9b, the nonlinear characteristic curve of the MOS transistor is used to extract the rectification effect. The feedback input inverters (M1 and M2) function as a current sink (Stromsenke) for the input current and convert it into a voltage. The transconductance of the input stage can be set using the value of the resistor R. When the input power is Iin = 0, the operating point of the circuit is set so that the same current flows through M1 and M2. It is preferable that M1 and M3, and M2 and M4 each have the same electrical characteristics, that is, channels having the same length and the same width. When an alternating current (Wechselstrom) is applied to the input of the circuit, these are non-linearly rectified by the transistors M3 to M6.

  A reference circuit comprised of transistors M7-M10 provides a comparison current for the total current of M4 and M6. As a result, the comparison results are actually available in digital form at the output (out) of the circuit. If a capacitor is connected in series upstream of the current inputs of the two circuits, this circuit is also suitable for voltage input signals. As the voltage signal changes, current displacement (Verschiebestrom) occurs in the capacitor and can be processed by the proposed circuit.

FIG. 10 shows an embodiment as a preferred reference form of the analog event memory ES. The embodiment shown in the upper part of FIG. 10 shows a simple analog integrator with a reset function (Ruecksetzmoeglichkeit). Charge carriers of the input current Iin are stored in the capacitor C. As a result, the output voltage Uout = C−1 × ∫Iindt is set. After the memory content is successfully measured or read, the reset transistor (Ruecksetztransistor) may be activated to empty the content as before.

  The two circuits shown at the bottom of FIG. 10 are designed to store the maximum input voltage (left side) and the maximum input current (right side) in an analog manner. In the memory circuit with the maximum input voltage (left side), a capacitor is used as a memory element. The capacitor tracks the input voltage using a source follower circuit (Sourcefolger-Schaltung) that essentially includes an offset derived from the transistor threshold voltage (Schwellenspannung). After successfully measuring or reading the memory contents, the contents may be emptied as before by activating the reset transistor.

  In the memory circuit regarding the maximum input power, the transistor M4 functions as a diode. Through this diode, the gate voltage of the transistor M1 for measuring the current is increased until the input current is compensated. Even when the input current decreases and returns to the original value, the gate voltage of M1 remains at the value that has been reached almost before. The accumulated maximum value of the input current can be taken from the drain terminal of transistor M2, which forms a current mirror with transistor M1.

  If the maximum value memory presented here is constructed in a complementary manner, a minimum value memory, i.e. a memory for the maximum negative current, is obtained.

  FIG. 11 shows a preferred matrix-like structure including sensor cells forming a sensor cell array, control wiring, supply wiring and readout wiring, and evaluation circuits, control circuits and operation circuits (Betriebsschaltungen) at the edge of the sensor cell array (matrix). Show. An orthogonal arrangement of cells and wiring (row wiring and column wiring orthogonal) is a preferred embodiment. However, diagonal wiring and, for example, hexagonal arrangement of sensor cells are also possible. In FIG. 11, the reference number STE represents the control unit for the control signal and the auxiliary voltage, and the reference number AWS represents the digital / analog evaluation circuit. The control unit STE and the evaluation circuit AWS are part of the activity evaluation unit.

FIG. 12 shows a preferred embodiment including a sensor cell array together with a control unit STE when purely analog signal processing of nerve cell signals as described in connection with the sixth embodiment. Analog output or evaluation wiring (column wiring) is coupled at the edge of the sensor cell array and adjusted for other processing using an appropriate amplifier V with a current input.

  FIG. 13 shows a preferred embodiment including a sensor cell array together with a control unit STE and an evaluation circuit AWA. In the present embodiment, the sensor cells SZ can exchange information directly with each other. It is preferred that cells directly adjacent exchange information in analog or digital form about neuronal events. This is to allow further improvement in detection sensitivity and / or further downsizing of the sensor cell SZ and / or preprocessing of recorded information. In this case, the exchanged information may be an analog or digital signal described in the above-described embodiment of the sensor cell SZ.

  In particular, in the first embodiment (digital storage / digital output), a shift register may be provided inside the sensor cell SZ. As a result, the information stored in the cells can be read continuously from the columns or rows. This is particularly true since there is no need to read out digital data from the sensor cell SZ via wiring with high parasitic capacitance across the entire column or row, rather from one sensor cell SZ to the next. Has the advantage that it only needs to be weitergereicht. Therefore, after multiple clock cycles, all data is available for other processing at the edge of the sensor cell array.

  For all the previously described embodiments, each sensor cell SZ is preferably designed to be activated or deactivated via a control wiring. Deactivation refers to the fact that the current neuron event does not generate a signal at the sensor cell output, especially when sensor information is transmitted in analog form. Do not supply to wiring or evaluation wiring.

1 is a schematic block diagram of a sensor cell according to a first embodiment of the present invention. It is a schematic block diagram of the sensor cell which concerns on 2nd Embodiment of this invention. It is a schematic block diagram of 3rd Embodiment as a reference form of this invention. It is a schematic block diagram of 4th Embodiment as a reference form of this invention. It is a schematic block diagram of 4th Embodiment as a reference form of this invention. It is a schematic block diagram of 5th Embodiment of this invention. It is a schematic block diagram of 6th Embodiment of this invention. It is a figure which shows the outline | summary of the classification | category of preferable embodiment in the apparatus of this invention which measures the activity of a neural network. 1 is a diagram showing an embodiment of a preferred amplifying element of the present invention having a PMOS transistor. It is a figure which shows one preferable embodiment of a rectification | straightening element and a threshold value detection element. It is a figure which shows other preferable embodiment of a preferable rectifier and a threshold value detection element. It is a figure which shows embodiment as a preferable reference form of an analog integrator and an analog extreme value memory. FIG. 2 is a diagram illustrating a preferred structure of a matrix type sensor cell array (each sensor cell can be driven and read out by various row and column wirings extending throughout the structure). FIG. 2 is a schematic block diagram of a preferred sensor cell array (analog current output signals are combined and amplified at the edge of the sensor cell array); Other preferred embodiments of a matrix-type sensor cell array (each cell can be driven and read by various row and column wirings extending throughout the structure, and various data can be exchanged between sensor cells. FIG.

Explanation of symbols

ADC Analog-to-digital converter AWS Evaluation circuit DAC Digital-to-analog converter E Detection electrode ES Event memory, especially digital counter or analog integrator GR Rectifier element IREF Reference current source SA Sensor output part SD Threshold detection element STE Control unit SZ Sensor Cell TKV Transconductance amplifier V Amplifying element VA Amplified output section

Claims (11)

In an apparatus for measuring neural network activity using a patterned semiconductor substrate,
The semiconductor substrate is
A plurality of sensor elements each having at least one conductive sensing electrode (E);
A plurality of amplifying elements (V; TKV) each having at least one amplifying input and at least one amplifying output (VA);
A plurality of event memories (ES) for storing a count reading which is a measure of the number of events in the neuronal cell, and at least one memory reading device for reading the count reading from the event memory (ES); at least one evaluation input unit comprises at least one activity evaluation unit and at least one evaluation output unit,
The conductive detection electrode is disposed on a surface of a semiconductor substrate to detect a nerve cell signal from a neural network, and the sensor element is an electric sensor output signal based on the detected nerve cell signal. Can be output via each sensor output part (SA) of the sensor element,
Each of the sensor elements is associated with one of the amplifying elements (V; TKV), and the amplifying input section of the amplifying element is electrically connected to the sensor output section (SA) of each sensor element. The output signal of the electric sensor can be output as an amplified output signal via an amplification output unit (VA),
The evaluation input unit (digital out; analog out) is electrically connected to at least one of event memory (ES) , and the activity evaluation unit receives an activity signal that is a measure of activity in the neural network as an event. It is generated based on the memory contents read from the memory (ES) and output via the evaluation output unit.
Each of the amplifying elements (V; TKV) is connected to an associated event memory (ES) via a threshold value detecting element (SD) for discretizing the amplified output signal depending on whether or not the amplified output signal is larger than a predetermined threshold value. Connected,
Each of the amplifying elements (V; TKV) is connected to an associated threshold value detecting element (SD) via a rectifying element (GR) for rectifying the amplified output signal,
The event memory (ES) is a digital memory element, specifically a digital counter,
One neuronal event is an amplified output signal rectified by the rectifier element that is greater than a predetermined threshold value of the threshold detector element that drives the digital counter to increase the count reading by one. apparatus.   A plurality of sensor cells (SZ) each including one of the sensor elements and an amplifying element (V; TKV) associated with the sensor element are provided, and the sensor cells (SZ) are arranged in a matrix shape. The apparatus of claim 1, wherein the apparatus is disposed to form a sensor cell array.   Each of the sensor cells (SZ) includes one event memory (ES), and an event memory input section of the event memory (ES) is connected to an amplification output section (VA). The device described in 1.   4. The apparatus according to claim 3, wherein each of the event memories (ES) is selected by a selection wiring (select) of an activity evaluation unit and can be selectively read out. The activity evaluation unit reads the event memory (ES) in a predetermined cycle, claims 1 and adapted to generate an activity signal for each cycle apparatus according to one of 4. The amplifying element (V; TKV) A device according to one of claims 1 transconductance amplifier element (TKV) for generating a current signal as an amplified output signal 5. The amplifying element (V; TKV) is a transconductance amplifying element (TKV) for generating a current signal as an amplified output signal.
At least two amplification outputs (VA) of the amplifying element (TKV) are connected to the activity evaluation unit via one evaluation wiring (analog out), so that a current signal from the amplifying element (TKV) is transmitted. Are summarized,
Apparatus according to claim 1 or 2, wherein the activity evaluation unit is adapted to generate an activity signal based on the amplitude of the combined current signal. 8. The apparatus according to claim 7 , wherein the amplification outputs (VA) of all the amplification elements (TKV) are connected to the activity evaluation unit via one evaluation wiring (analog out). Each of the amplification elements (TKV) is an activity evaluation unit by one output wiring (analog out) through a threshold detection element (SD) for discretizing the amplified output signal and a downstream reference power source (IREF). 9. The device according to claim 7 or 8 , which is connected to the device. At least 2 of the sensor cell (SZ), the claims 1 are connected to each other so as to be exchanged between at least one signal sensor cells of the sensor cell (SZ) (SZ) according to one of 9 Equipment. Using the apparatus described in one of claims 1 to 10, a method of measuring the activity of the neural network,
Detecting a nerve cell signal using a plurality of sensor elements;
Generating and outputting a sensor output signal based on the detected nerve cell signal;
Amplifying the sensor output signal by each of the amplifying elements (V; TKV) for generating an amplified output signal;
Generating an activity signal that is a measure of activity in the neural network based on the amplified output signal.