CN110020715B - Neural network identification method and device using mixed coding of fluctuation and pulse signals - Google Patents
Neural network identification method and device using mixed coding of fluctuation and pulse signals Download PDFInfo
- Publication number
- CN110020715B CN110020715B CN201811201335.4A CN201811201335A CN110020715B CN 110020715 B CN110020715 B CN 110020715B CN 201811201335 A CN201811201335 A CN 201811201335A CN 110020715 B CN110020715 B CN 110020715B
- Authority
- CN
- China
- Prior art keywords
- neural network
- signal
- matrix
- fluctuation
- artificial neural
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 122
- 238000000034 method Methods 0.000 title claims abstract description 40
- 239000011159 matrix material Substances 0.000 claims abstract description 104
- 210000002569 neuron Anatomy 0.000 claims abstract description 51
- 230000015654 memory Effects 0.000 claims abstract description 34
- 230000008569 process Effects 0.000 claims abstract description 14
- 238000012549 training Methods 0.000 claims abstract description 11
- 238000003860 storage Methods 0.000 claims abstract description 7
- 238000013507 mapping Methods 0.000 claims abstract description 4
- 238000013500 data storage Methods 0.000 claims description 26
- 238000004891 communication Methods 0.000 claims description 10
- 230000006870 function Effects 0.000 claims description 6
- 230000004044 response Effects 0.000 claims description 5
- 230000008054 signal transmission Effects 0.000 claims description 3
- 230000002708 enhancing effect Effects 0.000 claims description 2
- 230000000644 propagated effect Effects 0.000 abstract description 5
- 238000013529 biological neural network Methods 0.000 description 6
- 239000000203 mixture Substances 0.000 description 6
- 238000010586 diagram Methods 0.000 description 4
- 238000009826 distribution Methods 0.000 description 4
- 230000005236 sound signal Effects 0.000 description 4
- 230000005540 biological transmission Effects 0.000 description 3
- 238000013473 artificial intelligence Methods 0.000 description 2
- 230000006399 behavior Effects 0.000 description 2
- 238000005094 computer simulation Methods 0.000 description 2
- 238000004519 manufacturing process Methods 0.000 description 2
- 230000001537 neural effect Effects 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 230000002441 reversible effect Effects 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 108010076504 Protein Sorting Signals Proteins 0.000 description 1
- 230000004075 alteration Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000005520 cutting process Methods 0.000 description 1
- 230000009977 dual effect Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 210000005171 mammalian brain Anatomy 0.000 description 1
- 210000005036 nerve Anatomy 0.000 description 1
- 230000010355 oscillation Effects 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 238000007711 solidification Methods 0.000 description 1
- 230000008023 solidification Effects 0.000 description 1
- 238000005728 strengthening Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L15/00—Speech recognition
- G10L15/08—Speech classification or search
- G10L15/16—Speech classification or search using artificial neural networks
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- General Physics & Mathematics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Human Computer Interaction (AREA)
- Acoustics & Sound (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The application discloses an artificial neural network identification method utilizing mixed coding of fluctuation and pulse signals. The method comprises the following steps: generating a distributed artificial neural network; encoding original voice or image into a mixed signal of a fluctuation signal and a pulse signal, and inputting the mixed signal into a neural network; after entering a neural network, the mixed signal is distributed and propagated at a specific speed; sequentially exciting specific neurons by using the mixed signals to generate different characteristic paths, and mapping the neurons on the characteristic paths into a characteristic matrix for storage; repeating the encoding and matrix storage processes for multiple times, taking the average value of the feature matrix, generating a memory matrix, and training a neural network; storing the characteristic path according to the trained characteristic matrix data; inputting the voice or image to be recognized to the trained neural network to generate a feature matrix to be recognized; and (5) performing comparison to complete identification.
Description
Technical Field
The invention relates to computer software and hardware, a neural network, machine learning and artificial intelligence, in particular to a brand new artificial neural network method for voice recognition and image recognition.
Background
At present, the neural network method used in the field of artificial intelligence is an artificial neural network based on a digital computer. Such neural networks essentially binary encode the original information, delivering binary digital signals. When the neural network is used for intelligent voice and image recognition, a large amount of data sample input is needed for training, or the real-time internet connection is needed for calling a large amount of data, so that the dependence on the internet environment of the original data amount is strong.
The real biological neural network, especially the mammalian brain, simultaneously comprises the functions of analog signal and digital signal processing, wherein the analog signal exists in the form of electroencephalogram, the digital signal exists in the form of nerve pulse, and the analog signal and the digital signal are both important for intelligent behaviors such as memory, learning and the like. The learning process of the biological neural network does not require extensive data training and an internet environment. However, the current neural network method cannot integrate the biological neural network.
Disclosure of Invention
1. Purpose(s) to
In order to reduce the dependence of the existing neural network on big data training and the Internet environment and manufacture a more practical artificial intelligent recognition device, the invention provides an artificial neural network recognition method utilizing fluctuation and pulse signal mixed coding, which can quickly learn the characteristics of voice or image signals without depending on big data input and accurately recognize the same type of voice or image signals without depending on the Internet environment.
2. Technical scheme
Specifically, the invention provides an artificial neural network identification method using mixed coding of fluctuation and pulse signals, which is characterized by comprising the following steps:
the method comprises the following steps: generating a distributed artificial neural network, wherein the neural network consists of a plurality of single-frequency networks and is initialized to be fully communicated;
step two: encoding original voice or image into a mixed signal of a fluctuation signal and a pulse signal and inputting the mixed signal into a distributed artificial neural network;
step three: after entering a distributed artificial neural network, the mixed signal is distributed and transmitted at a specific speed, wherein the mixed signal is transmitted on any single-frequency line, but only pulse signals are transmitted between different single-frequency networks;
step four: sequentially exciting neurons in the network by using the mixed signal to generate different characteristic paths, and mapping the neurons on the characteristic paths into a characteristic matrix for storage, wherein one dimension is time and the other dimension is frequency;
step five: repeating the encoding and matrix storage processes from the second step to the fourth step for multiple times, taking the average value of the feature matrix, generating a memory matrix, and training the distributed artificial neural network;
step six: storing the characteristic path according to the trained characteristic matrix data;
step seven: inputting the voice or the image to be recognized into the trained distributed artificial neural network, and repeating the processes from the second step to the fourth step to generate a feature matrix to be recognized;
step eight: and comparing the stored characteristic matrix with the characteristic matrix to be identified to obtain the coincidence degree data of the characteristic matrix to be identified and the stored characteristic matrix.
Preferably, the distributed artificial neural network is composed of a plurality of single frequency networks, each single frequency network comprises neurons and connecting channels, the neurons in the single frequency networks are connected with each other, the neurons in the single frequency networks are also connected with each other, one path of each neuron of each neural network is connected to the data storage layer, when the signal is transmitted in the step two, the mixed signal is transmitted in the single frequency networks, and only the pulse signal is transmitted between the single frequency networks.
Preferably, in the distributed artificial neural network constructed in the first step, a single neural network node has multi-directional input and multi-directional output functions, wherein input and output paths are not shared.
Preferably, when encoding the voice, the sound frequency is encoded into the frequency of the fluctuation signal, the sound amplitude is encoded into the number of the pulse signals, and the pulse signals are only positioned near the wave crest position of the fluctuation signal; when the image is coded, the horizontal scanning and the vertical scanning are respectively coded into the frequency of a fluctuation signal, the light intensity is coded into the number of pulse signals, and the pulse signals are only positioned near the wave crest position of the fluctuation signal.
Preferably, the fourth step includes: and recording a signal transmission characteristic path through the response time sequence of the neural network node to the mixed signal.
Preferably, the step five comprises: the voice or image signals are input into the neural network for multiple times, and the feature matrixes generated repeatedly for multiple times are averaged and stored as a memory matrix.
Preferably, the sixth step includes: and feeding back the memory matrix as an identification standard to the distributed artificial neural network to establish a characteristic path of neural network connection.
Preferably, the seventh step includes: and inputting the voice or the image to be recognized into the distributed artificial neural network, repeating the processes from the second step to the fourth step, and generating a characteristic matrix to be recognized in a data storage layer for the voice or the image to be recognized each time.
Preferably, the sixth step includes: and enhancing the communication of each node of the neural network on the characteristic path, and closing the communication of each node of the neural network on the non-characteristic path.
The invention is inspired by the practical biological neural network, takes the connection distance of the node (neuron) as a brand new variable, leads the artificial neural network to have the function of processing fluctuation signals (simulation) and pulse signals (digital) at the same time, is the same as biological individuals when identifying voice and images, does not depend on big data training and internet environment, thereby obtaining a more practical artificial neural network identification method and manufacturing a more practical artificial intelligent device.
3. Advantages and effects
The traditional artificial neural network only comprises a digital neural network based on an existing computer or a pulse neural network (SNN) with a memristor added on a circuit. None of these schemes introduces a fluctuating signal (neural oscillation) and therefore there is no coding behavior of a mixture of fluctuating and pulsed signals.
The invention is different from the traditional digital neural network coding scheme:
(1) the invention introduces the connection distance of the neuron as a brand new variable into the artificial neural network, and increases the freedom degree of information coding, namely the time freedom degree. The timing is physically encoded naturally by the difference in connection distance of the neurons.
(2) The invention does not need to carry out large data volume training on the neural network, better accords with the learning mode of the biological neural network, stores each concept by a characteristic matrix, can realize off-line identification, and is applied to the application scene which can not be connected with the Internet.
(3) Aiming at fixedly recognizing one or a plurality of voice or image modes, the neural connection mode corresponding to the memory matrix is solidified on hardware, namely a characteristic path. The system is not influenced by data loss and is more robust.
(4) The invention only records the characteristic matrix data from the neuron where the signal appears during the identification, the characteristic matrix only depends on the relative relationship between the neurons and does not depend on the absolute frequency of the neurons, so the invention has more advantages for identifying the voice signal with the same content but different absolute frequencies and the image signal with the same content but different absolute positions.
Furthermore, in preferred implementations of the present invention, there are distinct advantages of the present invention with respect to speech and image recognition, respectively.
In speech recognition, the invention uses a mixed mode of a fluctuation signal and a pulse signal to code a one-dimensional sound signal into a two-dimensional characteristic matrix, thereby increasing the recognition freedom. In image recognition, the active lens device is used for dynamically scanning and coding along the characteristics of the image, all pixels of the image do not need to be recorded in the coding process, the dependence of data amount is reduced, and the recognition efficiency is improved.
Drawings
The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the invention and not to limit the invention. In the drawings:
fig. 1 is a flow chart of a signal identification process.
Fig. 2 is a schematic diagram of a mixed form of a wobble signal and a pulse signal. The duration of the pulse signal is fixed and is far less than the period of different fluctuation signals. The pulse signal amplitude is much larger than the wobble signal amplitude. f1 and f2 represent fluctuating signals at two different frequencies.
FIG. 3 is a schematic diagram of the relationship between the neural network layer and the data storage layer. The input signal enters the neural network layer in a mixed coding mode of a fluctuation signal and a pulse signal, and the propagation characteristics are extracted to the data storage layer through the pulse signal. The recognition result is output by the data storage layer.
Fig. 4 is a schematic diagram of a neural network connection. The simpler connection of the two signals (solid lines) is shown, one is the connection of the fluctuation and the pulse mix signal (f1, f2, f3, f4) at the time sequence (t1, t2, t3, t4 … …), and the other is the connection of the fluctuation and the pulse mix signal (f4, f2, f3, f1) at the time sequence (t1, t2, t3, t4 … …). The leftmost column is a sensing neuron and does not record signals. The other neurons are connected to the data storage layer by pulse signals (dotted line)
FIG. 5 is a schematic diagram of data storage layers for the two signals of FIG. 2. The lateral dimension is the temporal order { t1, t2, t3, t4}, and the longitudinal dimension is the frequency order { f1, f2, f3, f4 }. The matrix elements in the figure are neurons, each neuron has a logic threshold value, and the characteristic matrix is represented in a data storage layer that the Boolean logic value of the corresponding neuron is 1, and the Boolean logic values of other neurons are 0. When the neurons of the received pulse signal satisfy the logic value 1 in the arrow direction in sequence, the identification is realized.
Detailed Description
The invention is described in detail below with reference to the drawings and the embodiments thereof, but the scope of the invention is not limited thereto.
The embodiments of the present invention may be implemented using computer simulations.
The equipment conditions required for the implementation of the computer simulation were:
(1) digital computer (including desktop computer, notebook computer, intelligent mobile phone and tablet computer)
(2) Analog/digital converter (ADC)
(3) Active scanning camera (for image acquisition)
(4) Microphone (for voice collection)
(5) Signal propagation simulation software (such as NI series)
Next, a specific process is described. Fig. 1 shows a flow chart of the present embodiment. As shown in the figure, the method specifically comprises the following steps:
the method comprises the following steps: establishing a neural network, and carrying out initialization full connection on the neural network. The neural network is divided into two layers, namely a neural network layer at the bottom layer and a data storage layer above the neural network layer. The neural network layer is composed of a plurality of single-frequency networks, neurons in the single-frequency networks are connected with each other, and neurons in the single-frequency networks are also connected with each other, so that the distributed artificial neural network is finally generated. Each single frequency network can input only a pulse signal and a single frequency of a wobble signal. A single neural network node has both multi-directional input and multi-directional output functions, where the input and output paths are not shared. A mixture of fluctuating and pulsed signals may be transmitted within the single frequency network, but only pulsed signals may be transmitted between the single frequency networks. Each neuron of the neural network layer is connected with a data storage layer through a path
When no data is input, the neural network layer always keeps a full-connected mode. After data is input for the first time, the neural network generates a memory result and does not keep full communication any more. All subsequent data input is directly input to the neural network generating memory, and full communication is not initialized.
Step two: the input signal is input to the neural network through the mixed coding of the fluctuation signal and the pulse signal. For voice, the coding mode is that sound frequency codes are wave signal frequency, sound amplitude codes are pulse signal number, and pulse signals are only near wave signal peak positions; for images, horizontal scanning and vertical scanning are respectively encoded into the frequency of a fluctuation signal, the light intensity is encoded into the number of pulse signals, and the pulse signals are only positioned near the wave crest position of the fluctuation signal.
The mixed coding of the wave signal and the pulse signal is shown in the attached figure 2: the wave signal belongs to transverse waves, and the amplitude and the transmission direction are perpendicular to each other. The transmission speed of the wave signal is consistent with that of the pulse signal, and the amplitudes of the wave signal and the pulse signal can be superposed, namely the pulse signal exists at the position of a fixed phase of the wave signal, and the amplitude of the pulse signal is far larger than that of the wave signal. The exemplary embodiments of the present invention and the description thereof only consider the case where the pulse signal is at the peak position of the fluctuation signal for explaining the present invention, and do not constitute a limitation of the present invention.
The original signal (such as voice or image) can be converted into digital signal by analog/digital converter, and decomposed into single frequency signal by fourier analyzer on software and hardware, or the single frequency signal is superimposed (if the original sound signal contains different frequencies at the same time), and each single frequency signal is mixed and coded.
When the frequency of a sound signal is F1Amplitude of A1The occurrence time is T1. After the signal is mixed and coded by a fluctuation signal and a pulse signal, the corresponding frequency of the fluctuation signal is f1The number of pulse signals in each period of the wobble signal being n1At the time of input to the neural network, t1Wherein n is1And A1Proportional ratio, f1And F1Is in direct proportion. The duration of the encoded wave and pulse mixed signal is consistent with the duration of the original sound signal.
When an image signal is input, the lens is required to actively scan the image characteristic input. When the lens is made to scan the x-axis of the image in the forward direction (from left to right), the fluctuation signal at frequency f1 and the number of pulses n per cycle of the fluctuation signal1Coding; in reverse direction (from right to left), at frequency f2And the number of pulses n per period of the wobble signal2And (5) encoding. When the lens is used to scan the y-axis of the image in the forward direction (from top to bottom), the frequency f is used3And the number of pulses n per period of the wobble signal3Coding; in the reverse direction (from bottom to top), at a frequency f4And the number of pulses n per period of the wobble signal4And (5) encoding. When the lens scans along other directions, the scanning direction is determined by the superposition of two mixed signals and the different ratio of pulse numbers in each wave signal period. In summary, the lens is used to scan the image features in the positive and negative directions of the x and y of the plane where the image is located with different frequencies. For example, 45 degree sweep from top left to bottom right, the fluctuation signal of frequency f1 and frequency f3Is present at the same time, and n1And n3The ratio of (A) to (B) is 1: 1, and so on.
Step three: the coded mixed wave and pulse signals pass through the unique input interfaces (such as four interfaces shown in figure 4, aiming at f from top to bottom) of the single frequency network1，f2，f3，f4A single frequency network of four different fluctuating signal frequencies) into the neural network. After entering the neural network, the signal edges are mixedAll possible paths of the single frequency network are propagated (horizontal direction of fig. 4), but only the pulse signal part can be propagated between the single frequency networks (vertical direction of fig. 4). The propagation speed of the mixed signal into the neural network depends on the system hardware configuration. The propagation direction of the mixed signal depends on the input and output directions of the neuron. Within a single frequency network, the connection of neurons may propagate a mixture of fluctuating and impulsive signals. For single frequency networks, the connection of the neurons only propagates the pulse signal.
The mixed signal that is first propagated to the bottom of the single frequency network, i.e. the first signal in a set of signal sequences, sends a pulse signal to the data storage layer through the neurons at the bottom. As shown in FIG. 4a, the frequency is f1The wave signal of (2) first reaches f1At the bottom of the single frequency network (rightmost side in the figure), a pulse signal is sent by the neurons at the bottom to the data storage layer (dashed line). Therefore, the bottom neuron of each single frequency network has the function of receiving the input of the unidirectional mixed signal and outputting the unidirectional mixed signal to the data storage layer.
Step four: the mixed signal sequentially excites specific neurons, producing different characteristic paths. And mapping the neurons on the characteristic path into a characteristic matrix for storage, wherein one dimension is time and the other dimension is frequency. And recording a signal transmission characteristic path through the response time sequence of the neural network node to the mixed signal. The recording rule is that when the neural network node passes through a complete mixed signal (wave crest + pulse), the response is to output the pulse part of the signal to the data storage layer, and the non-response is not to output the pulse to the data storage layer. After receiving the pulse signal, the data storage layer stores the region where the signal appears as a feature matrix, i.e., a memory matrix. The matrix elements of the characteristic matrix represent nodes of the neural network, the transverse dimension represents time distribution of the nodes, the longitudinal dimension represents frequency distribution of the nodes, and the values of the matrix elements represent the number of pulses.
Specifically, in the timing of one set of mixed signals, after the first mixed signal has reached the bottom neuron, the second mixed signal has not yet reached the bottom neuron. But the neuron that arrives when the second mixed signal arrives at this time,when the bottom neuron which receives the first mixed signal transmits a pulse signal, the neuron sends a pulse signal to the data storage layer. As shown in FIG. 4a, the neurons are in the second row (f)2Single frequency network), satisfies the following relationship:
a1+b12＝a2+v(t2-t1)
where v is the transmission speed of the signal, t1And t2Respectively, the frequency f of the wave part1And f2Time in time sequence of the mixed signal of (a)1And a2Are respectively a mixed signal (f)1，t1) And (f)2，t2) The distance from a neuron at the signal input end to a neuron transmitting a pulse signal to the data storage layer in the respective single frequency network, b12Is the distance between the two neurons.
By analogy, the neurons that arrive by the third and fourth mixed signals receive the impulse signal from the neuron of the previous single frequency network, as shown in fig. 4a, each of the following relationships is satisfied:
a2+b23＝a3+v(t3-t2)
a4+b34＝a4+v(t4-t3)
at the data storage layer, the pulse signals received by the neurons are in time sequence (t)1，t2，t3，t4) Is distributed at (f)1，f2，f3，f4). The distribution can be stored as a feature matrix:
wherein the horizontal axis of the matrix is (t)1，t2，t3，t4) The vertical axis is (f)1，f2，f3，f4) Element of matrix (n)1，n2，n3，n4) Are respectively mixedResultant signal (f)1，f2，f3，f4) At a time (t)1，t2，t3，t4) The number of pulse signals owned.
Similarly, for FIG. 4b, the timing sequence (t) is the same1，t2，t3，t4) The distribution of the up-mix signal is (f)4，f2，f3，f1) The equation set satisfied by the neural network connection mode is as follows:
a4+b24＝a2+v(t4-t2)
a2+b23＝a3+v(t3-t2)
a3+b13＝a1+v(t3-t1)
FIG. 4b shows the corresponding feature matrix in the data storage layer as:
fig. 5 is a visual representation of the dual frequency superimposed new signal at the data storage layer when the two mixed signals of fig. 4a and 4b are superimposed at the same timing. The characteristic matrix corresponding to the signal is:
by analogy, the neural network also has unique feature matrix correspondence for any complex multi-frequency signal input.
Specifically, in real devices, the feature matrix generated by each group of signals can be converted into digital signals through voltage signals, and the digital signals are directly stored in digital memories such as flash memories and magnetic disks.
Step five: and (5) training a neural network. For more complex signals, the feature matrixes generated each time are slightly different, and have a certain error rate, so that a plurality of feature matrixes can be generated through multiple inputs, and the feature matrixes are averaged and regressed to form one feature matrix. This feature matrix is called a memory matrix.
For example, for the dual-frequency superimposed signal shown in fig. 5, the resulting memory matrix is:
that is, each matrix element is the value after N times of averaging, and N is the training times of the neural network.
Specifically, in real equipment, the feature matrixes can be deleted after average calculation, and only the calculated memory matrix is reserved. I.e., the feature matrix, may be stored in a volatile memory, such as a computer memory or the like. The memory matrix is eventually stored in a non-volatile memory, such as various types of computer hard disks. The amount of data occupied by the final memory matrix is much smaller than the amount of data required for the training process.
Step six: and (5) solidifying the neural network, namely saving the characteristic path according to the trained characteristic matrix data. Specifically, the communication of each node of the neural network on the characteristic path is enhanced, and the communication of each node of the neural network on the non-characteristic path is closed. The purpose of this step is to consolidate the neural network connections corresponding to the memory matrix for a certain mixed signal onto the neural network, so that the neural network is only suitable for the mixed signal. And feeding back the memory matrix as an identification standard to the neural network connection, namely establishing a characteristic path of the neural network connection. Simulating a biological neural network memory mode, strengthening the communication of each node of the neural network on the characteristic path, increasing the channel capacity in class by hardware, closing the communication of each node of the neural network on the non-characteristic path, and cutting off the channel on hardware.
And (4) for each memory matrix generated in the step five, a unique neural network connection corresponds to the memory matrix. For example, the matrix:
the corresponding neural network connections are the solid line connections in fig. 4 a. If the matrix is used as a memory matrix, the solid line connections in FIG. 4a are detected and all other neural network connections are cut off (not shown in FIG. 4 a). In a similar way, the matrix:
the corresponding neural network connections are the solid line connections in fig. 4 b. If the matrix is used as a memory matrix, the solid line connections in FIG. 4b are detected, and the other neural network connections are cut off (not shown in FIG. 4 b).
Particularly in an application scene, for a neural network only recognizing a single signal, the step can be adopted for solidification, and the memory and recognition efficiency is improved. For neural networks that recognize multiple signals, this step can be bypassed, the neural network is not cured, and the plasticity of the neural network continues to be maintained.
Step seven: and inputting a signal to be identified. And the signal input process to be identified is a process from the second step to the fourth step, a characteristic path appears after the mixed coding of the fluctuation signal and the pulse signal, and then a characteristic matrix is generated.
For the neural network that has undergone step six, the network allows only a specific signal that matches the cursive neural network (i.e., the signature matrix of the signal is consistent with the memory matrix corresponding to the cursive neural network) to propagate. If the characteristic matrix of the signal to be identified is too far from the memory matrix corresponding to the consolidated neural network, it will not be propagated in the network. E.g. one in time (t)1，t2，t3，t4) Is distributed at (f)2，f3，f4，f1) The feature matrix that should be generated originally is:
however, in the neural network shown in fig. 4a, any connection equation is not satisfied, and therefore no pulse signal is transmitted to the data storage layer, i.e., no feature matrix is generated. This case can be defined as an all 0 matrix, in particular to real devices.
For a neural network that has skipped step six, i.e., is not consolidated, the feature matrix would have the flexibility to be independent of the absolute frequency of the signal. According to the fourth step, the first mixed signal is sent to the data storage layer after reaching the bottom neuron, and the characteristic matrix is established. For example, a signal to be recognized is time sequence (t)1，t2，t3，t4) Is distributed at (f)1’，f2’，f3’，f4') if (f)1’，f2’，f3’，f4') and (f)1，f2，f3，f4) Is different in absolute frequency, but (f)1’，f2’，f3’，f4') relative frequency relationship between and (f)1，f2，f3，f4) The relative relationship of the frequencies is the same, and the amplitude is the same (i.e. the number of pulses is also n)1，n2，n3，n4) Then (f)1’，f2’，f3’，f4') the generated feature matrix is:
that is to say and (f)1，f2，f3，f4) Are consistent.
Specifically, when a speech signal is recognized, if only the absolute frequency of the speech signal has movement, the relative frequency relationship is not changed, and the amplitude is not changed, the same feature matrix is generated through the neural network which is mixed and encoded by the fluctuation signal and the pulse signal.
Specifically, when the image signal is identified, if the identification part only moves in position in the whole scanning area, and the relative relation of various characteristics of the image in the identification area is not changed, the same or very similar characteristic matrix is generated through the neural network which is mixed and coded by the fluctuation signal and the pulse signal. If the relative relation of various characteristics of the images in the identification area is unchanged, but the size of the identification area is changed, the scanning speed can be actively changed through a lens to compensate, and the generated characteristic matrix still keeps similarity.
Step eight: and (4) signal identification. Checking each matrix element value of the characteristic matrix to be identified, comparing the matrix element value with a value corresponding to the stored memory matrix, quantitatively giving comparison result data, setting a total comparison threshold value, and outputting a digital logic signal (namely 1 when being higher than the threshold value and 0 when being lower than the threshold value). This step only occurs in the data storage layer, i.e. only the memory matrix is compared with the feature matrix of the signal to be recognized. The method can be completed by a traditional data comparison method.
During data alignment, one-to-one alignment results for each matrix element can be returned. When the matrix metadata are consistent, the return value is 1; when the matrix metadata are consistent, the return value is 0. A threshold value for the recognition rate, such as 0.9, may then be set. When the one-to-one comparison result of all matrix elements is more than 90% and is 1, namely the result of 0 is less than 10%, the identification result exceeds the threshold value of 0.9, and the whole identification result is true. Otherwise, the whole recognition result is false.
Due to the error rate of the feature matrix of the signal to be recognized, for example, the error of the starting position of the feature matrix. Therefore, during identification, each matrix element of the feature matrix can be integrally translated in a small range, for example, after 1-3 matrix elements are translated in time sequence, identification is performed again and repeatedly for several times, and the result with the highest identification rate is obtained. Like the feature matrix:
after shifting one matrix element forward in time sequence, the memory matrix below can be matched
And matching to obtain 80% of recognition results. If the set threshold is below 80%, the recognition result is true.
While the principles of the invention have been described in detail in connection with the preferred embodiments thereof, it will be understood by those skilled in the art that the foregoing embodiments are merely illustrative of exemplary implementations of the invention and are not limiting of the scope of the invention. The details of the embodiments are not to be interpreted as limiting the scope of the invention, and any obvious changes, such as equivalent alterations, simple substitutions and the like, based on the technical solution of the invention, can be interpreted without departing from the spirit and scope of the invention.
Claims (9)
1. An artificial neural network identification method using hybrid coding of wave and pulse signals, the method comprising the steps of:
the method comprises the following steps: generating a distributed artificial neural network, wherein the neural network consists of a plurality of single-frequency networks and is initialized to be fully communicated;
step two: encoding original voice or image into a mixed signal of a fluctuation signal and a pulse signal and inputting the mixed signal into a distributed artificial neural network;
step three: after entering a distributed artificial neural network, the mixed signal is distributed and transmitted at a specific speed, wherein the mixed signal is transmitted on any single-frequency line, but only pulse signals are transmitted between different single-frequency networks;
step four: sequentially exciting neurons in the network by using the mixed signal to generate different characteristic paths, and mapping the neurons on the characteristic paths into a characteristic matrix for storage, wherein one dimension is time and the other dimension is frequency;
step five: repeating the encoding and matrix storage processes from the second step to the fourth step for multiple times, taking the average value of the feature matrix, generating a memory matrix, and training the distributed artificial neural network;
step six: storing the characteristic path according to the trained characteristic matrix data;
step seven: inputting the voice or the image to be recognized into the trained distributed artificial neural network, and repeating the processes from the second step to the fourth step to generate a feature matrix to be recognized;
step eight: and comparing the stored characteristic matrix with the characteristic matrix to be identified to obtain the coincidence degree data of the characteristic matrix to be identified and the stored characteristic matrix.
2. The method of claim 1, wherein the distributed artificial neural network comprises a plurality of single frequency networks, each single frequency network comprises neurons and connecting channels, the neurons within the single frequency networks are connected to each other, the neurons between the single frequency networks are also connected to each other, one path of each neuron of each neural network is connected to the data storage layer, the mixed signal is transmitted within the single frequency network when the signals are transmitted in the third step, and only the pulse signal is transmitted between the single frequency networks.
3. The artificial neural network identifying method using the hybrid coding of the fluctuation and the pulse signals as claimed in claim 1, wherein in the distributed artificial neural network constructed in the first step, a single neural network node has a multi-directional input and multi-directional output function in which input and output paths are not shared.
4. The artificial neural network recognition method using the hybrid coding of the fluctuation and the pulse signals as claimed in claim 1, wherein when the speech is coded, the sound frequency is coded as the frequency of the fluctuation signal, the sound amplitude is coded as the number of the pulse signals, and the pulse signals are only located near the peak position of the fluctuation signal; when the image is coded, the horizontal scanning and the vertical scanning are respectively coded into the frequency of a fluctuation signal, the light intensity is coded into the number of pulse signals, and the pulse signals are only positioned near the wave crest position of the fluctuation signal.
5. The artificial neural network identification method using the hybrid coding of the fluctuation and the pulse signal as claimed in claim 1, wherein the fourth step comprises: and recording a signal transmission characteristic path through the response time sequence of the neural network node to the mixed signal.
6. The artificial neural network identification method using the hybrid coding of the fluctuation and the pulse signal as claimed in claim 1, wherein the step five comprises: the voice or image signals are input into the neural network for multiple times, and the feature matrixes generated repeatedly for multiple times are averaged and stored as a memory matrix.
7. The artificial neural network identification method using the hybrid coding of the fluctuation and the pulse signal as claimed in claim 1, wherein the sixth step comprises: and feeding back the memory matrix as an identification standard to the distributed artificial neural network to establish a characteristic path of neural network connection.
8. The artificial neural network identification method using the hybrid coding of the fluctuation and the pulse signal as claimed in claim 1, wherein the seventh step comprises: and inputting the voice or the image to be recognized into the distributed artificial neural network, repeating the processes from the second step to the fourth step, and generating a characteristic matrix to be recognized in a data storage layer for the voice or the image to be recognized each time.
9. The artificial neural network identification method using the hybrid coding of the fluctuation and the pulse signal as claimed in claim 1, wherein the sixth step comprises: and enhancing the communication of each node of the neural network on the characteristic path, and closing the communication of each node of the neural network on the non-characteristic path.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811201335.4A CN110020715B (en) | 2018-10-16 | 2018-10-16 | Neural network identification method and device using mixed coding of fluctuation and pulse signals |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811201335.4A CN110020715B (en) | 2018-10-16 | 2018-10-16 | Neural network identification method and device using mixed coding of fluctuation and pulse signals |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110020715A CN110020715A (en) | 2019-07-16 |
| CN110020715B true CN110020715B (en) | 2020-11-27 |
Family
ID=67188484
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811201335.4A Active CN110020715B (en) | 2018-10-16 | 2018-10-16 | Neural network identification method and device using mixed coding of fluctuation and pulse signals |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110020715B (en) |
Families Citing this family (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110728974A (en) * | 2019-10-29 | 2020-01-24 | 甘肃龙科智能科技有限公司 | Speech recognition method, device, equipment and computer readable storage medium |
| CN112070063B (en) * | 2020-09-28 | 2023-11-17 | 贵州省山地资源研究所 | Neural network visual recognition method for unmanned aerial vehicle |
| CN112511273B (en) * | 2020-12-01 | 2021-12-14 | 江苏博沃汽车电子系统有限公司 | Information processing method and device for reducing error rate of digital circuit |
| CN112686381A (en) | 2020-12-30 | 2021-04-20 | 北京灵汐科技有限公司 | Neural network model, method, electronic device, and readable medium |
| CN113257282B (en) * | 2021-07-15 | 2021-10-08 | 成都时识科技有限公司 | Speech emotion recognition method and device, electronic equipment and storage medium |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2007115116A (en) * | 2005-10-21 | 2007-05-10 | Hiroshima Industrial Promotion Organization | Neuron element and information processing method using it |
| CN107092959A (en) * | 2017-04-07 | 2017-08-25 | 武汉大学 | Hardware friendly impulsive neural networks model based on STDP unsupervised-learning algorithms |
| CN107194426A (en) * | 2017-05-23 | 2017-09-22 | 电子科技大学 | A kind of image-recognizing method based on Spiking neutral nets |
| CN108470190A (en) * | 2018-03-09 | 2018-08-31 | 北京大学 | The image-recognizing method of impulsive neural networks is customized based on FPGA |
| CN108615533A (en) * | 2018-03-28 | 2018-10-02 | 天津大学 | A kind of high-performance sound enhancement method based on deep learning |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11003984B2 (en) * | 2016-05-31 | 2021-05-11 | Samsung Electronics Co., Ltd. | Timing sequence for digital STDP synapse and LIF neuron-based neuromorphic system |
| US10846590B2 (en) * | 2016-12-20 | 2020-11-24 | Intel Corporation | Autonomous navigation using spiking neuromorphic computers |
-
2018
- 2018-10-16 CN CN201811201335.4A patent/CN110020715B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2007115116A (en) * | 2005-10-21 | 2007-05-10 | Hiroshima Industrial Promotion Organization | Neuron element and information processing method using it |
| CN107092959A (en) * | 2017-04-07 | 2017-08-25 | 武汉大学 | Hardware friendly impulsive neural networks model based on STDP unsupervised-learning algorithms |
| CN107194426A (en) * | 2017-05-23 | 2017-09-22 | 电子科技大学 | A kind of image-recognizing method based on Spiking neutral nets |
| CN108470190A (en) * | 2018-03-09 | 2018-08-31 | 北京大学 | The image-recognizing method of impulsive neural networks is customized based on FPGA |
| CN108615533A (en) * | 2018-03-28 | 2018-10-02 | 天津大学 | A kind of high-performance sound enhancement method based on deep learning |
Non-Patent Citations (2)
| Title |
|---|
| "Temporal precision of spike response to fluctuating input in pulse-coupled networks of oscillating neurons";Jun-nosuke Teramae等;《arXiv》;20090622;1-10页 * |
| "基于忆阻器的组合电路及神经网络研究";董哲康;《中国优秀硕士学位论文全文数据库 信息科技辑》;20151215;正文7-13页,37-45页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110020715A (en) | 2019-07-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110020715B (en) | Neural network identification method and device using mixed coding of fluctuation and pulse signals | |
| Agrawal et al. | C-vqa: A compositional split of the visual question answering (vqa) v1. 0 dataset | |
| Alemi et al. | Fixing a broken ELBO | |
| CN112801209B (en) | Image classification method based on dual-length teacher model knowledge fusion and storage medium | |
| Dodge et al. | Quality robust mixtures of deep neural networks | |
| CN109858486B (en) | Deep learning-based data center cloud target identification method | |
| CN110287770B (en) | Water individual target matching identification method based on convolutional neural network | |
| Fu et al. | Model compression for IoT applications in industry 4.0 via multiscale knowledge transfer | |
| CN109259764B (en) | Method for determining dynamic brain function network threshold | |
| CN115941112B (en) | Portable hidden communication method, computer equipment and storage medium | |
| Li et al. | Residual quantization for low bit-width neural networks | |
| CN116797796A (en) | Signal identification method based on time-frequency analysis and deep learning under DRFM intermittent sampling | |
| Redfern et al. | BCNN: a binary CNN with all matrix ops quantized to 1 bit precision | |
| Nida et al. | Video augmentation technique for human action recognition using genetic algorithm | |
| CN109448039B (en) | Monocular vision depth estimation method based on deep convolutional neural network | |
| CN113743317A (en) | License plate picture generation method, medium and equipment based on generation countermeasure network | |
| Wan et al. | 3D-Mask-GAN: Unsupervised Single-View 3D Object Reconstruction | |
| Revi et al. | Gan-generated fake face image detection using opponent color local binary pattern and deep learning technique | |
| CN116106880B (en) | Underwater sound source ranging method and device based on attention mechanism and multi-scale fusion | |
| Cai et al. | DCASE2023 task1 submission: Device simulation and time-frequency separable convolution for acoustic scene classification | |
| CN106680800A (en) | Dual-frequency identification sonar data processing method | |
| CN114462707B (en) | Web service multidimensional QoS joint prediction method based on feature depth fusion | |
| CN115861779A (en) | Unbiased scene graph generation method based on effective feature representation | |
| CN116402995A (en) | Lightweight neural network-based ancient architecture point cloud semantic segmentation method and system | |
| US20210158153A1 (en) | Method and system for processing fmcw radar signal using lightweight deep learning network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |