Abstract:There have been many attempts to implement neural networks in the analog circuit. Most of them had a lot of input terms, and most studies implemented neural networks in the analog circuit through a circuit simulation program called Spice to avoid the need to design chips at a high cost and implement circuits directly to input them. In this study, we will implement neural networks using a capacitor and diode and use microcontrollers (Arduino Mega 2560 R3 boards) to drive real-world models and analyze the results.
What problem does this paper attempt to address?
The main problem that this paper attempts to solve is to implement neural networks by using analog circuits in order to reduce the time and cost required for training deep - learning models. Specifically, the author proposes a new structure, aiming to eliminate the instability in traditional analog neural network circuits and build circuits that are superior to using GPU or TPU for deep - learning model training in terms of speed, power consumption and price. In addition, the author also uses this structure to learn the MNIST data set and analyzes its performance.
### Core problems of the paper:
1. **Reducing training time and cost**: Deep - learning techniques are increasingly widely used in various fields, but training neural networks requires a large amount of electricity and time, especially when using high - parameter - quantity models such as ResNet. The GPU capabilities on personal computers are limited, causing tasks to be blocked. Therefore, finding a more efficient method to train models has become the key to research.
2. **Solving the instability of analog circuits**: Traditional analog neural network circuits are easily affected by noise because they use resistance values as weights, and the resistance values are difficult to adjust and control, resulting in these circuits not being effectively used for learning. In addition, analog circuits also have problems of high power consumption and large occupied area.
3. **Improving circuit performance**: By designing a new circuit structure, the author hopes to improve the operation speed and accuracy of the circuit while maintaining low power consumption, making it more competitive in practical applications.
### Solutions:
- **New circuit structure**: The author designs a new circuit structure based on capacitors and diodes, uses the Arduino Mega 2560 R3 board to drive the real - world model, and analyzes the results.
- **Non - linear proof**: It is proved through mathematical derivation that the designed multi - layer analog circuit has non - linear characteristics, thus ensuring its effectiveness as a multi - layer analog circuit.
- **Performance comparison**: By comparing with the Python implementation based on Numpy, the performance of the new circuit in terms of accuracy, running time and power consumption is verified.
### Experimental process:
1. **Data pre - processing**: Pre - process the MNIST data set, including matrix transformation, zero - padding and size adjustment.
2. **Circuit assembly**: Build the circuit using four sub - modules (Module 1, Module 2, Module 3 and Module 4).
3. **Learning process**: Use the completed circuit and microcontroller (Arduino Mega 2560 board) to learn.
4. **Performance analysis**: Compare the learning results of the MNIST data set with the Numpy implementation in Python, and analyze the accuracy, running time and power consumption.
### Results and discussion:
- **Accuracy**: Although the new circuit is not as accurate as the Numpy - based implementation (0.9 vs 0.3), it performs better in Top 3 accuracy (0.7).
- **Running time**: The running time of the new circuit is longer (37 minutes and 39 seconds vs 14 seconds), mainly due to the lower clock frequency (16MHz vs 5GHz) and memory limitations of the Arduino 2560 R3 board.
- **Power consumption**: The new circuit performs excellently in terms of power consumption, which is lower than that of traditional analog deep - learning circuits.
- **Future improvements**: The author proposes future improvement directions, including using more accurate current sensors, optimizing data storage and access methods, and improving the efficiency of the output structure.
Through the above methods, the author shows the potential of implementing deep - learning in analog circuits, especially the advantages when dealing with large - scale input data.