Sigmoid function generator
By generating a sigmoid IV curve function using parallel resistors composed of metal phosphorus sulfides (selenides) and transition metal chalcogenides, the complexity of the sigmoid function generator and the low efficiency of the von Neumann architecture in existing technologies are solved, and a high-precision neural network activation function is realized, which is suitable for image recognition tasks.
Patent Information
- Application Number
- CN202411069433.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-06
- Publication Date
- 2025-11-14
AI Technical Summary
Existing sigmoid function generators are complex and difficult to generate through hardware simulation, and traditional von Neumann architectures are inefficient in large datasets.
A parallel resistor is constructed using metal phosphorus sulfide (selenide) and transition metal chalcogenides. Its ferroelectricity and ionic conductivity are used to generate an S-type current-voltage curve function, which is then used as an activation function for neural networks.
The generated S-shaped IV curve function training model achieves an accuracy of 97.1%. The device has a simple structure, low power consumption, and is suitable for image recognition tasks.
Smart Images
Figure CN120952066A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to function generator technology in the field of electronic components, and more particularly to a Sigmoid function generator. Background Technology
[0002] Neuromorphic computing and machine learning have powerful applications in image and speech recognition, language processing, predictive analytics, and other fields. However, the traditional von Neumann architecture, limited by frequent data transfers between memory and processor, has gradually become unable to handle workloads involving large datasets. Researchers have begun to optimize hardware architectures to accelerate computation and reduce power consumption, such as Google's Tensor Processing Unit (TPU), Neural Processing Unit (NPU), and Data Processing Unit (DPU).
[0003] The activation function of a neuron is a crucial parameter in neural networks, preventing linearization during model training. It is typically a nonlinear function, such as the ReLU, Sigmoid, or tanh function, which can be implemented using LUTs (Linguistic Undertouch Tests) and analog circuits. However, due to its nonlinear characteristics, it is difficult to generate via hardware simulation. Researchers have generated approximate Sigmoid functions using complex electronic circuits such as CMOS transistor circuits, low-gain comparator circuits, and inverter circuits; however, the structures of these Sigmoid function generators are extremely complex. Summary of the Invention
[0004] In view of this, the present invention proposes a sigmoid function generator. The generator is a resistor composed of a transition metal chalcogenide (TMDC) and a metal phosphorus sulfide (selenide) (MPSC), which utilizes the coexistence of ferroelectricity and ionic conductivity of the metal phosphorus sulfide (selenide) to generate an S-shaped current-voltage (IV) curve function.
[0005] To achieve the above objectives, the present invention provides a sigmoid function generator, comprising a substrate 1, a metal phosphide (selenide) 2, a transition metal chalcogenide 3, and electrodes 4. The metal phosphide (selenide) 2 is located on the substrate 1, the transition metal chalcogenide 3 is located on the metal phosphide (selenide) 2, and the two electrodes 4 are located at both ends of the transition metal chalcogenide 3 and connected to the metal phosphide (selenide) 2. The metal phosphide (selenide) 2 and the transition metal chalcogenide 3 form a parallel resistor.
[0006] Furthermore, the substrate 1 is a highly insulating dielectric material.
[0007] Furthermore, the metal phosphorus sulfide (selenide) 2 is a material having the chemical formula MM'X2Y6, wherein M is one of Cu, Ag, Fe, Co, Ni, and Zn, M' is one of In, Sn, Sc, Bi, and Pb, X is P, and Y is one of S and Se.
[0008] Furthermore, the thickness of the metal phosphorus sulfide (selenium) oxide layer 2 ranges from 10 to 200 nanometers.
[0009] Furthermore, the transition metal chalcogenide 3 is a material having the chemical formula MX2, wherein M is one of Mo, W, Pt, Hf, Ti, Bi, Ga, and Sn, and X is one of O, S, Se, and Te.
[0010] Furthermore, the thickness of the transition metal chalcogenide layer 3 ranges from 0.5 to 50 nanometers.
[0011] Furthermore, the material of the electrode 4 is one of Ti, Au, Ni, and Pt, or a combination of two in any proportion.
[0012] In summary, this invention provides a Sigmoid function generator. This invention uses a parallel resistor composed of metal phosphorus sulfide (selenide) and transition metal chalcogenides, utilizing the coexistence of ferroelectricity and ionic conductivity of the metal phosphorus sulfide (selenide) to generate an S-shaped IV curve function, which is used as an activation function for a neural network. The trained model achieves an accuracy of 97.1% and can perform image recognition tasks.
[0013] The above-described technical solution of the present invention has the following beneficial technical effects:
[0014] (1) The present invention utilizes a parallel resistor composed of metal phosphorus sulfide (selenide) and transition metal chalcogenide to generate an S-type function, and the device structure and preparation process are extremely simple.
[0015] (2) The present invention utilizes two-dimensional semiconductor materials to prepare a Sigmoid function generator. The device size can reach the micrometer level, and it has the characteristics of extremely small size and extremely low power consumption.
[0016] (3) The S-shaped IV curve function generated by this invention is used as an activation function, and the accuracy of the trained model can reach 97.1%, which is higher than the commonly used ReLU function and tanh function.
[0017] (4) The S-shaped IV curve function generated by this invention can be used for training on different datasets, including but not limited to MNIST, KMNIST and FMNIST datasets. Attached Figure Description
[0018] Figure 1This is a schematic diagram of the structure of the Sigmoid function generator provided by the present invention.
[0019] Figure 2 The sigmoid function generator provided in this invention generates the sigmoid IV curve.
[0020] In the figure, 1 is the substrate, 2 is a metal phosphorus sulfide (selenide), 3 is a transition metal chalcogenide, and 4 is an electrode. Detailed Implementation
[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.
[0022] The sigmoid function generator disclosed in this invention, such as Figure 1 As shown, it includes a substrate 1, a metal phosphide (selenide) 2, a transition metal chalcogenide 3, and an electrode 4. The metal phosphide (selenide) 2 is located on the substrate 1, the transition metal chalcogenide 3 is located on the metal phosphide (selenide) 2, and the two electrodes 4 are located at both ends of the transition metal chalcogenide 3 and connected to the metal phosphide (selenide) 2. The metal phosphide (selenide) 2 and the transition metal chalcogenide 3 form a parallel resistor.
[0023] The sigmoid function generator in this invention is a parallel resistor composed of metal phosphide (selenide) 2 and transition metal chalcogenide 3. It utilizes the coexistence of ferroelectricity and ionic conductivity of metal phosphide (selenide) 2 to generate an S-shaped current-voltage (IV) curve function.
[0024] In this design, substrate 1 can be any highly insulating dielectric material. Metal phosphorus sulfide (selenide) 2 is a two-dimensional metal phosphorus sulfide (selenide) film with a thickness of approximately 10-200 nm, such as CuInP2S6, CuInP2Se6, CuBiP2Se6, or AgBiP2Se6, AgScP2Se6, AgInP2Se6, etc. Transition metal chalcogenide 2 is a two-dimensional transition metal chalcogenide film with a thickness of approximately 0.5-50 nm, such as tungsten selenide, molybdenum selenide, platinum selenide, molybdenum sulfide, tungsten sulfide, platinum sulfide, or tungsten telluride, molybdenum telluride, platinum telluride. Electrode 4 is made of one of Ti, Au, Ni, and Pt, or a composite of any two of these materials.
[0025] Here are more specific examples:
[0026] Example 1
[0027] The device structure is as described above, wherein substrate 1 is silicon dioxide, metal phosphorus sulfide (selenide) 2 is CuInP2S6, transition metal chalcogenide 3 is molybdenum disulfide, and electrode 4 is a Ti electrode. A DC voltage scan from −3 V to 3 V is applied across the generator, and the current signal is acquired to obtain the following... Figure 2 The S-shaped IV curve is shown.
[0028] The curve data was imported into a neural network training model and used as the activation function, replacing the commonly used ReLU or tanh functions. The model trained with this curve achieved an accuracy of 97.1%, which is higher than the accuracy of models trained using ReLU and tanh functions. Furthermore, we used three different datasets for training to verify the universality of this activation function, and the model accuracy comparison is as follows:
[0029] Table 1. Comparison of Model Accuracy
[0030] Activation function dataset IV curve ReLU tanh MNIST 97.1% 96.1% 95.4% KMNIST 83.2% 76.3% 79.4% FMNIST 77.9% 75.2% 76.7%
[0031] In summary, this invention provides a Sigmoid function generator. This invention uses a parallel resistor composed of metal phosphorus sulfide (selenide) and transition metal chalcogenides, utilizing the coexistence of ferroelectricity and ionic conductivity of the metal phosphorus sulfide (selenide) to generate an S-shaped IV curve function, which is used as an activation function for a neural network. The trained model achieves an accuracy of 97.1% and can perform image recognition tasks.
Claims
1. A sigmoid function generator, characterized in that, It includes a substrate (1), a metal phosphide (selenide) (2), a transition metal chalcogenide (3), and an electrode (4); the metal phosphide (selenide) (2) is located on the substrate (1), the transition metal chalcogenide (3) is located on the metal phosphide (selenide) (2), and the two electrodes (4) are located at both ends of the transition metal chalcogenide (3) and connected to the metal phosphide (selenide) (2). The metal phosphide (selenide) layer 2 and the transition metal chalcogenide layer 3 form a parallel resistor.
2. A Sigmoid function generator according to claim 1, characterized in that, The metal phosphorus sulfide (selenide) (2) is a material with the chemical formula MM'X2Y6, where M is one of Cu, Ag, Fe, Co, Ni, Zn, M' is one of In, Sn, Sc, Bi, Pb, X is P, and Y is one of S and Se; the thickness of the metal phosphorus sulfide (selenide) layer (2) ranges from 10 to 200 nanometers.
3. A Sigmoid function generator according to claim 1, characterized in that, The transition metal chalcogenide layer (3) is a material with the chemical formula MX2, wherein M is one of Mo, W, Pt, Hf, Ti, Bi, Ga, Sn, and X is one of O, S, Se, Te; the thickness of the transition metal chalcogenide layer (3) ranges from 0.5 to 50 nanometers.