A Learning Dendritic Neuron-Based Motion Direction Detective System and Its Application to Grayscale Images
Abstract
:1. Introduction
2. Method
2.1. Dendritic Neuron Model
- Synaptic layer: This layer integrates inputs from other neurons and transforms them into a single output using a sigmoid function. For a dendritic neuron model with i inputs and j branches, the synaptic function is given by:This step endows dendritic neurons with learning capabilities. The is the distance parameter of the input, and , are the weight and bias of the synapse. Through a learning algorithm, the model can prune ineffective synapses and branches by the iteration of weights and bias, enhancing detection accuracy.
- Branch layer: This layer distributes inputs to the soma layer through the function [19]:
- Membrane layer: This layer, representing the dendritic neuron’s plasma membrane, regulates ion flow through voltage-gated ion channels. The membrane potential changes in response to synaptic inputs and is integral to the neuron’s ability to generate action potentials. This layer represents the neuron’s ability to process integrated signals and prepare for signal transmission. In real neurons, the membrane potential is influenced by the sum of excitatory and inhibitory inputs, leading to changes in the neuron’s overall activity. The function is expressed as:
- Soma layer: The cell body performs nonlinear processing of the integrated inputs from the branch layer, combining them to produce a single output. This is expressed by:
2.2. Local Motion Direction Detective Dendritic Neuron Model
2.3. Global Motion Direction Detective Dendritic Neuron Model
2.4. Learning Algorithm
3. Simulation Results
3.1. Dataset Description
- (1)
- Black background with solid-colored object (dark background and light object, DL);
- (2)
- Black background with random-colored object (dark background and random object, DR);
- (3)
- Solid background with black object (light background and dark object, LD);
- (4)
- Solid background with solid-colored object (light background and light object, LL);
- (5)
- Solid background with random-colored object (light background and random object, LR);
- (6)
- Random background with black object (random background and dark object, RD);
- (7)
- Random background with solid-colored object (random background and light object, RL);
- (8)
- Random background with random-colored object (random background and random object, RR).
3.2. SOTA Model Description
- Four-Layer CNN: We compare our model with a Four-Layer CNN due to its straightforward structure and ease of understanding. This model retains the core advantages of CNNs, such as automatic feature extraction and hierarchical feature learning, while minimizing computational complexity. These features make it suitable for benchmarking our motion direction detection model.
- LeNet: As one of the earliest deep learning architectures, LeNet offers a balance between simplicity and performance. It extends beyond basic CNNs by introducing a wider application scope and higher accuracy. Its structured approach to convolution and pooling layers makes it a valuable baseline for comparison [42].
- ResNet: ResNet50 is known for its residual connections, which allow for the training of very deep networks by addressing the vanishing gradient problem. The ResNet50 model, with its 50 layers, excels in handling labeled image datasets, making it highly accurate and adaptable to our two-frame grayscale motion detection task. ResNet’s architecture includes residual blocks that facilitate gradient flow, improving convergence and accuracy. The key components include convolutional layers with batch normalization and ReLU activations, along with identity and convolutional shortcuts to form residual blocks, enhancing model performance without significantly increasing complexity [43].
- EfficientNet: Designed for scaling model size efficiently, EfficientNet-B0 provides a balance between computational cost and accuracy. This model is particularly suitable for smaller datasets like ours, as it optimizes learning efficiency while maintaining robust performance. EfficientNet uses a compound scaling method that uniformly scales network depth, width, and resolution, enhancing both efficiency and accuracy. Its architecture includes a series of mobile inverted bottleneck MBConv blocks and squeeze-and-excitation optimization to improve feature representation while being computationally efficient [44].
- ConvNeXt: ConvNeXt represents a modern approach to image classification with deep neural networks. Known for its high accuracy, this model leverages advanced techniques to enhance feature extraction and model training, making it an ideal candidate for validating our proposed model’s performance. ConvNeXt improves upon traditional CNNs by incorporating insights from transformer models, resulting in enhanced feature representation and improved classification accuracy. It employs a hierarchical architecture with stage-wise downsampling and a series of convolutional layers, offering a balance between spatial resolution and computational efficiency [45].
3.3. Cross-Validation Accuracy
3.4. Different Training Data Ratio
3.5. Noise-Immunity Simulation
3.6. Simulation on Real-World and Rotated MNIST Datasets
4. Conclusions and Discussion
Author Contributions
Funding
Institutional Review Board Statement
Informed Consent Statement
Data Availability Statement
Conflicts of Interest
References
- Xu, Y.; Lv, C.; Li, S.; Xin, P.; Ma, S.; Zou, H.; Zhang, W. Review of development of visual neural computing. Comput. Eng. Appl. 2017, 53, 30–34. [Google Scholar]
- Todo, Y.; Tang, Z.; Todo, H.; Ji, J.; Yamashita, K. Neurons with multiplicative interactions of nonlinear synapses. Int. J. Neural Syst. 2019, 29, 1950012. [Google Scholar] [CrossRef] [PubMed]
- Medina, J. Brain Rules: 12 Principles for Surviving and Thriving at Work, Home, and School; ReadHowYouWant.com: East Sussex, UK, 2011. [Google Scholar]
- Fiske, S.T.; Taylor, S.E. Social Cognition; Mcgraw-Hill Book Company: New York, NY, USA, 1991. [Google Scholar]
- Vanston, J.E.; Strother, L. Sex differences in the human visual system. J. Neurosci. Res. 2017, 95, 617–625. [Google Scholar] [CrossRef]
- Namboodiri, V.M.K.; Huertas, M.A.; Monk, K.J.; Shouval, H.Z.; Shuler, M.G.H. Visually cued action timing in the primary visual cortex. Neuron 2015, 86, 319–330. [Google Scholar] [CrossRef]
- Smith, A.T.; Snowden, R.J. Visual Detection of Motion; Academic Press: Cambridge, MA, USA, 1994. [Google Scholar]
- Borst, A.; Euler, T. Seeing Things in Motion: Models, Circuits, and Mechanisms. Neuron 2011, 71, 974–994. [Google Scholar] [CrossRef] [PubMed]
- Hubel, D.H.; Wiesel, T.N. Receptive fields, binocular interaction and functional architecture in the cat’s visual cortex. J. Physiol. 1962, 160, 106. [Google Scholar] [CrossRef]
- Hubel, D.H.; Wiesel, T.N. Shape and arrangement of columns in cat’s striate cortex. J. Physiol. 1963, 165, 559. [Google Scholar] [CrossRef]
- Hubel, D.H.; Wiesel, T.N. Receptive fields of single neurones in the cat’s striate cortex. J. Physiol. 1959, 148, 574. [Google Scholar] [CrossRef] [PubMed]
- Hassenstein, B.; Reichardt, W. Systemtheoretische Analyse der Zeit-, Reihenfolgen- und Vorzeichenauswertung bei der Bewegungsperzeption des Rüsselkäfers Chlorophanus. Z. Naturforschung 1956, 11, 513–524. [Google Scholar] [CrossRef]
- Barlow, H.; Levick, W.R. The mechanism of directionally selective units in rabbit’s retina. J. Physiol. 1965, 178, 477. [Google Scholar] [CrossRef]
- Oyster, C.W.; Barlow, H.B. Direction-selective units in rabbit retina: Distribution of preferred directions. Science 1967, 155, 841–842. [Google Scholar] [CrossRef] [PubMed]
- Stone, J.; Freeman, R.B. Neurophysiological Mechanisms in the Visual Discrimination of Form. In Central Processing of Visual Information A: Integrative Functions and Comparative Data; Springer: Berlin/Heidelberg, Germany, 1973; pp. 153–207. [Google Scholar] [CrossRef]
- Marlin, S.G.; Hasan, S.J.; Cynader, M. Direction-selective adaptation in simple and complex cells in cat striate cortex. J. Neurophysiol. 1988, 59, 1314–1330. [Google Scholar] [CrossRef] [PubMed]
- Koch, C.; Poggio, T.; Torre, V. Retinal ganglion cells: A functional interpretation of dendritic morphology. Philos. Trans. R. Soc. London Biol. Sci. 1982, 298, 227–263. [Google Scholar]
- Wang, B.; Zhang, Y. Asymmetric connections with starburst amacrine cells underlie the upward motion selectivity of J-type retinal ganglion cells. Plos Biol. 2023, 21, e3002301. [Google Scholar] [CrossRef]
- Blakemore, C.; Cooper, G.F. Development of the Brain Depends on the Visual Environment. Nature 1970, 228, 477–478. [Google Scholar] [CrossRef]
- ScienceDaily. A Neuromorphic Visual Sensor Can Recognize Moving Objects and Predict Their Path. 2023. Available online: https://www.sciencedaily.com/releases/2023/04/230417142518.htm (accessed on 13 August 2024).
- Su, J.; Wu, C.; Yang, S. Object-Tracking Algorithm Combining Motion Direction and Time Series. Appl. Sci. 2023, 13, 4835. [Google Scholar] [CrossRef]
- Tao, S.; Zhang, X.; Hua, Y.; Tang, Z.; Todo, Y. A Novel Artificial Visual System for Motion Direction Detection with Completely Modeled Retinal Direction-Selective Pathway. Mathematics 2023, 11, 3732. [Google Scholar] [CrossRef]
- Todo, Y.; Mitsui, T. A learning multiple-valued logic network using genetic algorithm. Inf. Control. 2014, 10, 565–574. [Google Scholar]
- Jiang, T.; Wang, D.; Ji, J.; Todo, Y.; Gao, S. Single dendritic neuron with nonlinear computation capacity: A case study on xor problem. In Proceedings of the 2015 IEEE International Conference on Progress in Informatics and Computing (PIC), Nanjing, China, 18–20 December 2015; pp. 20–24. [Google Scholar]
- Chen, T.; Kobayashi, Y.; Todo, Y.; Tang, Z. A learning artificial visual system for motion direction detection. Neural Comput. Appl. 2024, 36, 17181–17197. [Google Scholar] [CrossRef]
- Dacheux, R.F.; Raviola, E. The rod pathway in the rabbit retina: A depolarizing bipolar and amacrine cell. J. Neurosci. 1986, 6, 331–345. [Google Scholar] [CrossRef]
- Werblin, F.S.; Dowling, J.E. Organization of the retina of the mudpuppy, Necturus maculosus. II. Intracellular recording. J. Neurophysiol. 1969, 32, 339–355. [Google Scholar] [CrossRef] [PubMed]
- Nelson, R. AII amacrine cells quicken time course of rod signals in the cat retina. J. Neurophysiol. 1982, 47, 928–947. [Google Scholar] [CrossRef]
- McCulloch, W.S.; Pitts, W. A logical calculus of the ideas immanent in nervous activity. Bull. Math. Biophys. 1943, 5, 115–133. [Google Scholar] [CrossRef]
- Rossi, L.F.; Harris, K.D.; Carandini, M. Spatial connectivity matches direction selectivity in visual cortex. Nature 2020, 588, 648–652. [Google Scholar] [CrossRef]
- Vlasits, A.; Baden, T. Motion Vision: A New Mechanism in the Mammalian Retina. Curr. Biol. 2019, 29, R933–R935. [Google Scholar] [CrossRef] [PubMed]
- Song, S.; Chen, X.; Tang, C.; Song, S.; Tang, Z.; Todo, Y. Training an Approximate Logic Dendritic Neuron Model Using Social Learning Particle Swarm Optimization Algorithm. IEEE Access 2019, 7, 141947–141959. [Google Scholar] [CrossRef]
- Tamura, H.; Tang, Z.; Ishii, M. The neuron model considering difference of time of inputs and its movement direction selection function. Trans. Inst. Electr. Eng. Jpn. Publ. Electron. 2002, 122, 1094–1103. [Google Scholar] [CrossRef]
- Gale, S.D.; Murphy, G.J. Active dendritic properties and local inhibitory input enable selectivity for object motion in mouse superior colliculus neurons. J. Neurosci. 2016, 36, 9111–9123. [Google Scholar] [CrossRef]
- Kuffler, S.W. Discharge patterns and functional organization of mammalian retina. J. Neurophysiol. 1953, 16, 37–68. [Google Scholar] [CrossRef]
- Tao, S.; Todo, Y.; Tang, Z.; Li, B.; Zhang, Z.; Inoue, R. A novel artificial visual system for motion direction detection in grayscale images. Mathematics 2022, 10, 2975. [Google Scholar] [CrossRef]
- Rochefort, N.L.; Narushima, M.; Grienberger, C.; Marandi, N.; Hill, D.N.; Konnerth, A. Development of direction selectivity in mouse cortical neurons. Neuron 2011, 71, 425–432. [Google Scholar] [CrossRef] [PubMed]
- Ichinose, T.; Habib, S. ON and OFF Signaling Pathways in the Retina and the Visual System. Front. Ophthalmol. 2022, 2, 989002. [Google Scholar] [CrossRef] [PubMed]
- Kim, U.S.; Mahroo, O.A.; Mollon, J.D.; Yu-Wai-Man, P. Retinal ganglion cells—Diversity of cell types and clinical relevance. Front. Neurol. 2021, 12, 661938. [Google Scholar] [CrossRef] [PubMed]
- Redmon, J.; Farhadi, A. Yolov3: An incremental improvement. arXiv 2018, arXiv:1804.02767. [Google Scholar]
- Liu, W.; Anguelov, D.; Erhan, D.; Szegedy, C.; Reed, S.; Fu, C.Y.; Berg, A.C. Ssd: Single shot multibox detector. In Proceedings of the Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, 11–14 October 2016; Proceedings, Part I 14. Springer: Berlin/Heidelberg, Germany, 2016; pp. 21–37. [Google Scholar]
- LeCun, Y.; Bottou, L.; Bengio, Y.; Haffner, P. Gradient-based learning applied to document recognition. Proc. IEEE 1998, 86, 2278–2324. [Google Scholar] [CrossRef]
- He, K.; Zhang, X.; Ren, S.; Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, Las Vegas, NV, USA, 27–30 June 2016; pp. 770–778. [Google Scholar]
- Tan, M.; Le, Q. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proceedings of the International Conference on Machine Learning, Long Beach, CA, USA, 9–15 June 2019; pp. 6105–6114. [Google Scholar]
- Liu, Z.; Mao, H.; Wu, C.Y.; Feichtenhofer, C.; Darrell, T.; Xie, S. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, New Orleans, LA, USA, 18–24 June 2022; pp. 11976–11986. [Google Scholar]
- Bishop, C.M.; Nasrabadi, N.M. Pattern Recognition and Machine Learning; Springer: Berlin/Heidelberg, Germany, 2006; Volume 4. [Google Scholar]
- Goodfellow, I.; Bengio, Y.; Courville, A. Deep Learning; MIT Press: Cambridge, MA, USA, 2016. [Google Scholar]
- Hebb, D.O. The Organization of Behavior: A Neuropsychological Theory; Psychology Press: East Sussex, UK, 2005. [Google Scholar]
- Zang, Y.; Hong, S.; De Schutter, E. Firing rate-dependent phase responses of Purkinje cells support transient oscillations. eLife 2020, 9, e60692. [Google Scholar] [CrossRef]
- Zang, Y.; De Schutter, E. The cellular electrophysiological properties underlying multiplexed coding in Purkinje cells. J. Neurosci. 2021, 41, 1850–1863. [Google Scholar] [CrossRef]
Model | Train Data | Train Acc | Test Acc | |||||||
---|---|---|---|---|---|---|---|---|---|---|
DL | DR | LD | LL | LR | RD | RL | RR | |||
AVS | DL | 99.71 ± 0.12% | 99.61 ± 0.21% | 100.0 ± 0.00% | 99.57 ± 0.25% | 99.67 ± 0.25% | 100.0 ± 0.00% | 74.08 ± 2.07% | 73.63 ± 1.87% | 80.19 ± 1.01% |
DR | 100.0 ± 0.00% | 99.96 ± 0.04% | 100.0 ± 0.00% | 99.99 ± 0.02% | 99.99 ± 0.02% | 100.0 ± 0.00% | 16.99 ± 5.32% | 16.93 ± 5.72% | 41.09 ± 5.94% | |
LD | 99.73 ± 0.09% | 99.73 ± 0.14% | 100.0 ± 0.00% | 99.71 ± 0.15% | 99.69 ± 0.24% | 99.99 ± 0.01% | 74.23 ± 2.54% | 73.58 ± 3.46% | 80.78 ± 1.98% | |
LL | 9.70 ± 0.20% | 99.66 ± 0.17% | 100.0 ± 0.00% | 99.62 ± 0.20% | 99.62 ± 0.19% | 100.0 ± 0.00% | 74.23 ± 3.01% | 73.39 ± 3.62% | 80.69 ± 3.26% | |
LR | 100.0 ± 0.00% | 99.96 ± 0.12% | 100.0 ± 0.00% | 99.98 ± 0.08% | 99.96 ± 0.06% | 100.0 ± 0.00% | 15.52 ± 4.29% | 15.85 ± 3.96% | 39.31 ± 3.80% | |
RD | 99.58 ± 0.08% | 32.09 ± 8.30% | 44.87 ± 5.68% | 31.87 ± 7.30% | 31.31 ± 6.67% | 45.57 ± 5.95% | 99.56 ± 0.14% | 99.26 ± 0.13% | 99.54 ± 0.20% | |
RL | 99.25 ± 0.06% | 34.09 ± 6.92% | 46.48 ± 6.99% | 32.97 ± 8.93% | 33.13 ± 10.42% | 47.10 ± 6.51% | 99.59 ± 0.17% | 99.27 ± 0.19% | 99.53 ± 2.15% | |
RR | 99.63 ± 0.05% | 64.41 ± 29.65% | 86.16 ± 14.29% | 64.69 ± 29.81% | 64.45 ± 30.16% | 86.33 ± 14.34% | 93.68 ± 3.01% | 92.60 ± 3.16% | 99.54 ± 0.18% | |
Four-Layer CNN | DL | 97.67 ± 0.29% | 96.26 ± 0.63% | 96.58 ± 0.47% | 41.13 ± 15.49% | 29.33 ± 5.52% | 36.06 ± 10.57% | 17.63 ± 3.22% | 12.85 ± 1.71% | 12.96 ± 2.41% |
DR | 97.61 ± 0.15% | 95.27 ± 1.00% | 96.06 ± 0.64% | 26.98 ± 13.01% | 24.34 ± 7.02% | 29.52 ± 9.45% | 17.13 ± 3.27% | 12.91 ± 1.24% | 14.16 ± 2.16% | |
LD | 96.46 ± 0.76% | 39.04 ± 11.37% | 38.51 ± 11.74% | 94.08 ± 0.13% | 54.73 ± 5.63% | 66.15 ± 6.39% | 15.18 ± 3.26% | 12.65 ± 2.70% | 14.10 ± 2.54% | |
LL | 93.60 ± 2.23% | 91.69 ± 1.92% | 91.57 ± 1.78% | 92.32 ± 1.70% | 87.43 ± 3.07% | 89.58 ± 2.00% | 19.99 ± 3.11% | 14.51 ± 2.16% | 14.87 ± 2.17% | |
LR | 95.00 ± 1.72% | 88.42 ± 2.63% | 89.31 ± 2.62% | 88.99 ± 3.04% | 86.43 ± 3.47% | 90.74 ± 1.88% | 16.10 ± 2.58% | 13.87 ± 3.41% | 15.42 ± 2.41% | |
RD | 100.0 ± 0.00% | 6.41 ± 3.97% | 4.11 ± 2.61% | 85.85 ± 4.95% | 37.16 ± 4.99% | 37.21 ± 5.03% | 73.61 ± 1.84% | 23.61 ± 2.34% | 26.74 ± 3.27% | |
RL | 79.23 ± 3.15% | 16.94 ± 3.38% | 16.38 ± 3.90% | 17.09 ± 4.96% | 15.01 ± 2.19% | 14.84 ± 3.29% | 16.34 ± 3.42% | 12.90 ± 1.34% | 13.43 ± 1.36% | |
RR | 88.97 ± 3.15% | 33.45 ± 11.09% | 32.99 ± 11.80% | 32.89 ± 12.10% | 23.29 ± 5.65% | 31.10 ± 9.53% | 25.03 ± 5.75% | 17.98 ± 3.39% | 26.56 ± 5.93% | |
LeNet | DL | 97.05 ± 4.65% | 96.27 ± 4.72% | 96.16 ± 5.33% | 26.65 ± 11.20% | 21.55 ± 4.26% | 24.65 ± 7.64% | 18.64 ± 6.84% | 12.76 ± 2.15% | 12.75 ± 1.05% |
DR | 98.21 ± 3.67% | 96.72 ± 3.98% | 97.76 ± 3.31% | 33.37 ± 3.90% | 20.68 ± 2.71% | 23.59 ± 3.35% | 23.07 ± 4.67% | 13.71 ± 1.68% | 13.38 ± 1.14% | |
LD | 92.94 ± 5.17% | 34.54 ± 21.48% | 32.61 ± 19.46% | 89.21 ± 6.35% | 51.01 ± 7.27% | 60.21 ± 8.15% | 17.18 ± 2.42% | 13.00 ± 1.59% | 15.15 ± 2.07% | |
LL | 95.18 ± 3.53% | 88.96 ± 3.73% | 89.11 ± 3.03% | 89.00 ± 3.60% | 88.41 ± 3.31% | 86.80 ± 3.84% | 16.92 ± 2.82% | 12.96 ± 1.99% | 13.92 ± 1.43% | |
LR | 93.89 ± 5.72% | 88.08 ± 6.69% | 89.45 ± 6.86% | 89.62 ± 5.54% | 86.93 ± 4.38% | 90.67 ± 4.96% | 14.85 ± 2.80% | 14.44 ± 2.47% | 14.63 ± 3.58% | |
RD | 95.78 ± 2.28% | 45.29 ± 6.73% | 38.23 ± 6.30% | 74.59 ± 6.90% | 26.39 ± 4.72% | 36.29 ± 6.28% | 67.91 ± 4.92% | 21.39 ± 3.05% | 28.09 ± 4.27% | |
RL | 79.34 ± 7.47% | 12.96 ± 1.52% | 12.38 ± 0.83% | 11.84 ± 1.77% | 12.09 ± 1.32% | 11.98 ± 0.71% | 12.69 ± 0.70% | 12.83 ± 0.69% | 12.44 ± 0.65% | |
RR | 77.82 ± 7.39% | 12.52 ± 0.79% | 13.01 ± 0.73% | 12.63 ± 1.09% | 13.33 ± 1.05% | 13.40 ± 1.45% | 12.50 ± 0.51% | 12.66 ± 0.70% | 12.64 ± 0.73% |
Model | Train Data | Train Acc | Test Acc | |||||||
---|---|---|---|---|---|---|---|---|---|---|
DL | DR | LD | LL | LR | RD | RL | RR | |||
Efficien tNet-B0 | DL | 99.79 ± 0.03% | 99.30 ± 0.18% | 99.71 ± 1.77% | 22.79 ± 3.64% | 21.94 ± 2.73% | 25.38 ± 3.47% | 13.11 ± 1.74% | 12.85 ± 0.92% | 12.01 ± 1.76% |
DR | 99.84 ± 0.05% | 98.01 ± 0.40% | 99.49 ± 0.16% | 22.39 ± 2.87% | 20.66 ± 2.26% | 22.82 ± 2.97% | 12.59 ± 3.07% | 12.46 ± 2.91% | 12.94 ± 1.59% | |
LD | 99.62 ± 0.12% | 15.73 ± 8.27% | 14.91 ± 7.37% | 98.99 ± 0.23% | 57.40 ± 5.26% | 60.30 ± 4.53% | 16.16 ± 4.15% | 14.10 ± 2.30% | 12.99 ± 1.82% | |
LL | 99.42 ± 0.08% | 97.84 ± 0.42% | 98.61 ± 0.26% | 97.89 ± 0.61% | 97.85 ± 0.54% | 96.43 ± 1.29% | 14.89 ± 1.95% | 11.74 ± 2.49% | 12.89 ± 2.36% | |
LR | 99.64 ± 0.08% | 95.89 ± 3.30% | 98.44 ± 4.18% | 95.77 ± 6.95% | 95.22 ± 6.73% | 98.11 ± 0.24% | 12.27 ± 2.21% | 12.61 ± 2.05% | 12.99 ± 2.52% | |
RD | 99.97 ± 0.03% | 12.67 ± 2.70% | 13.52 ± 2.46% | 14.83 ± 3.74% | 13.79 ± 2.02% | 14.08 ± 2.37% | 84.15 ± 1.05% | 25.42 ± 1.32% | 23.71 ± 1.71% | |
RL | 99.93 ± 0.03% | 18.24 ± 1.62% | 18.74 ± 3.91% | 15.75 ± 3.20% | 15.64 ± 2.34% | 15.30 ± 2.52% | 37.60 ± 1.74% | 23.59 ± 0.86% | 16.91 ± 1.59% | |
RR | 99.93 ± 0.03% | 13.04 ± 1.83% | 14.00 ± 2.00% | 13.30 ± 1.99% | 11.80 ± 1.51% | 12.87 ± 1.06% | 23.46 ± 3.01% | 19.58 ± 1.24% | 31.05 ± 0.02% | |
ResNet 50 | DL | 99.99 ± 0.01% | 99.96 ± 0.04% | 99.94 ± 0.07% | 12.25 ± 1.92% | 17.06 ± 2.03% | 18.82 ± 1.85% | 12.61 ± 1.49% | 12.43 ± 1.75% | 13.80 ± 2.63% |
DR | 99.94 ± 0.04% | 99.39 ± 0.24% | 99.85 ± 0.08% | 13.07 ± 1.47% | 18.63 ± 2.17% | 21.19 ± 4.12% | 11.68 ± 1.29% | 12.53 ± 1.61% | 11.56 ± 2.17% | |
LD | 99.89 ± 0.06% | 7.67 ± 3.78% | 7.59 ± 4.21% | 99.96 ± 0.04% | 46.91 ± 2.18% | 50.87 ± 2.74% | 17.17 ± 4.63% | 13.53 ± 1.92% | 12.81 ± 2.09% | |
LL | 99.81 ± 0.08% | 98.32 ± 0.89% | 99.31 ± 0.44% | 98.03 ± 1.37% | 97.35 ± 1.05% | 98.21 ± 1.48% | 14.47 ± 2.15% | 14.06 ± 2.82% | 13.36 ± 1.76% | |
LR | 99.85 ± 0.07% | 96.79 ± 0.83% | 99.32 ± 0.25% | 97.18 ± 0.59% | 96.56 ± 0.62% | 99.10 ± 0.18% | 16.57 ± 4.22% | 13.49 ± 1.48% | 15.36 ± 4.16% | |
RD | 100.0 ± 0.00% | 13.15 ± 2.64% | 12.54 ± 2.05% | 12.98 ± 0.69% | 13.40 ± 1.92% | 12.66 ± 1.56% | 13.88 ± 1.12% | 12.77 ± 0.82% | 12.69 ± 1.08% | |
RL | 100.0 ± 0.00% | 13.71 ± 1.50% | 12.59 ± 1.87% | 12.51 ± 2.13% | 13.30 ± 1.42% | 12.04 ± 1.49% | 12.07 ± 1.06% | 12.56 ± 0.91% | 12.53 ± 0.83% | |
RR | 100.0 ± 0.00% | 12.43 ± 1.53% | 12.68 ± 2.30% | 11.84 ± 1.39% | 12.66 ± 1.80% | 13.21 ± 1.40% | 12.47 ± 0.81% | 12.44 ± 0.97% | 12.59 ± 0.69% | |
Conv NeXt | DL | 99.95 ± 0.04% | 99.86 ± 0.10% | 98.35 ± 0.26% | 14.03 ± 4.70% | 12.44 ± 1.64% | 13.10 ± 1.54% | 14.26 ± 2.60% | 12.73 ± 1.68% | 11.84 ± 0.94% |
DR | 99.95 ± 0.05% | 99.91 ± 0.08% | 99.00 ± 0.33% | 12.29 ± 2.12% | 12.49 ± 1.97% | 13.25 ± 1.40% | 11.81 ± 2.82% | 11.59 ± 1.84% | 11.89 ± 2.06% | |
LD | 99.96 ± 0.02% | 35.46 ± 14.22% | 30.46 ± 13.44% | 99.86 ± 0.11% | 47.23 ± 3.67% | 47.43 ± 1.29% | 41.88 ± 8.59% | 19.80 ± 2.69% | 13.40 ± 0.98% | |
LL | 13.81 ± 0.53% | 13.05 ± 2.51% | 13.25 ± 2.59% | 13.31 ± 3.42% | 12.25 ± 1.83% | 11.15 ± 1.63% | 12.40 ± 1.39% | 12.25 ± 2.54% | 11.50 ± 2.54% | |
LR | 97.53 ± 1.90% | 83.85 ± 22.15% | 82.34 ± 20.66% | 79.20 ± 23.54% | 69.79 ± 23.98% | 57.23 ± 15.17% | 33.21 ± 9.58% | 19.29 ± 2.87% | 18.08 ± 2.62% | |
RD | 96.32 ± 0.88% | 14.17 ± 2.03% | 13.04 ± 2.85% | 81.69 ± 9.32% | 33.64 ± 5.72% | 32.88 ± 5.95% | 65.11 ± 6.69% | 22.78 ± 2.19% | 15.23 ± 1.80% | |
RL | 60.50 ± 25.68% | 13.24 ± 1.90% | 13.23 ± 1.90% | 12.29 ± 1.57% | 13.18 ± 1.48% | 13.53 ± 1.83% | 12.57 ± 0.79% | 12.81 ± 0.71% | 12.46 ± 0.65% | |
RR | 98.30 ± 0.97% | 12.84 ± 1.10% | 13.34 ± 1.67% | 12.59 ± 1.43% | 12.60 ± 0.97% | 12.79 ± 1.67% | 13.49 ± 0.94% | 12.71 ± 1.02% | 12.31 ± 0.47% |
Model | Train Acc | Test Acc | Train Acc | Test Acc |
---|---|---|---|---|
Train–Test | 75:25 | 50:50 | ||
AVS | 99.77 ± 0.18% | 99.73 ± 0.21% | 99.83 ± 0.04% | 99.88 ± 0.06% |
Four-Layer CNN | 97.65 ± 0.45% | 64.58 ± 2.55% | 86.78 ± 6.76% | 55.42 ± 4.11% |
LeNet-5 | 88.69 ± 1.80% | 51.91 ± 5.11% | 91.09 ± 1.79% | 47.23 ± 4.14% |
EfficientNet-B0 | 99.85 ± 0.61% | 76.49 ± 2.85% | 98.74 ± 1.08% | 72.73 ± 3.07% |
ResNet50 | 99.93 ± 0.06% | 70.10 ± 2.64% | 99.57 ± 0.29% | 66.33 ± 6.26% |
ConvNeXt | 98.23 ± 4.44% | 46.36 ± 17.2% | 97.17 ± 4.33% | 41.46 ± 19.6% |
Train–Test | 10:90 | 5:95 | ||
AVS | 99.79 ± 0.16% | 99.70 ± 0.31% | 99.52 ± 0.51% | 99.30 ± 0.72% |
Four-Layer CNN | 92.35 ± 3.05% | 20.06 ± 2.47% | 93.89 ± 3.02% | 17.10 ± 2.17% |
LeNet-5 | 92.45 ± 1.35% | 14.71 ± 5.98% | 93.64 ± 1.70% | 13.50 ± 0.69% |
EfficientNet-B0 | 98.67 ± 0.58% | 53.84 ± 2.75% | 98.16 ± 1.14% | 32.38 ± 3.90% |
ResNet50 | 100.0 ± 0.00% | 12.63 ± 1.00% | 98.09 ± 2.28% | 12.17 ± 0.48% |
ConvNeXt | 99.12 ± 1.42% | 13.88 ± 0.66% | 99.90 ± 0.25% | 12.69 ± 0.95% |
Noises | 1% | 2% | 5% | 10% |
---|---|---|---|---|
Train–Test | 75:25 | |||
AVS | 99.51 ± 0.12% | 98.89 ± 0.27% | 95.43 ± 2.65% | 89.03 ± 4.54% |
Four-Layer CNN | 45.62 ± 4.05% | 39.92 ± 3.82% | 33.58 ± 3.85% | 28.53 ± 3.69% |
LeNet-5 | 35.19 ± 2.68% | 30.06 ± 2.96% | 23.81 ± 2.59% | 20.39 ± 2.74% |
EfficientNet-B0 | 67.14 ± 2.08% | 61.84 ± 2.66% | 53.36 ± 3.29% | 46.54 ± 2.31% |
ResNet50 | 47.73 ± 4.55% | 43.39 ± 3.17% | 37.42 ± 2.08% | 30.74 ± 1.83% |
ConvNeXt | 40.98 ± 13.44% | 37.89 ± 10.00% | 31.87 ± 9.48% | 27.40 ± 8.11% |
Train–Test | 50:50 | |||
AVS | 99.54 ± 0.19% | 99.08 ± 0.28% | 97.79 ± 0.95% | 94.24 ± 1.74% |
Four-Layer CNN | 39.06 ± 3.62% | 34.66 ± 2.84% | 29.18 ± 2.70% | 24.29 ± 2.78% |
LeNet-5 | 31.87 ± 3.03% | 27.34 ± 2.88% | 22.63 ± 2.14% | 18.70 ± 1.08% |
EfficientNet-B0 | 58.29 ± 4.57% | 53.72 ± 3.25% | 47.09 ± 3.02% | 40.34 ± 2.18% |
ResNet50 | 46.46 ± 5.25% | 41.39 ± 4.27% | 35.07 ± 4.87% | 27.82 ± 3.37% |
ConvNeXt | 35.09 ± 16.79% | 31.86 ± 15.46% | 26.84 ± 10.89% | 23.09 ± 9.68% |
Train–Test | 10:90 | |||
AVS | 99.24 ± 0.47% | 98.03 ± 1.65% | 93.90 ± 3.60% | 86.86 ± 4.38% |
Four-Layer CNN | 16.92 ± 1.48% | 16.23 ± 0.96% | 14.89 ± 1.26% | 13.98 ± 1.12% |
LeNet-5 | 13.90 ± 0.92% | 13.03 ± 0.72% | 13.09 ± 0.91% | 12.67 ± 0.85% |
EfficientNet-B0 | 33.83 ± 3.92% | 29.39 ± 4.10% | 24.69 ± 3.14% | 21.06 ± 1.89% |
ResNet50 | 13.04 ± 0.75% | 12.79 ± 1.01% | 12.57 ± 0.88% | 12.75 ± 0.47% |
ConvNeXt | 13.36 ± 1.18% | 13.28 ± 0.40% | 13.24 ± 0.80% | 12.71 ± 0.94% |
Train–Test | 5:95 | |||
AVS | 99.04 ± 0.55% | 95.08 ± 6.64% | 93.26 ± 3.92% | 87.46 ± 6.39% |
Four-Layer CNN | 14.83 ± 2.27% | 14.79 ± 1.49% | 14.13 ± 1.54% | 13.88 ± 1.15% |
LeNet-5 | 12.48 ± 1.35% | 12.44 ± 0.84% | 12.37 ± 0.74% | 12.59 ± 0.69% |
EfficientNet-B0 | 23.40 ± 2.39% | 20.81 ± 2.28% | 17.66 ± 1.41% | 16.04 ± 1.39% |
ResNet50 | 12.64 ± 0.51% | 12.60 ± 0.95% | 12.23 ± 0.85% | 12.78 ± 0.71% |
ConvNeXt | 13.11 ± 0.93% | 12.84 ± 1.21% | 13.09 ± 0.82% | 12.48 ± 0.86% |
Dataset | Real Object | Narrowed Real Object | Rotated MNIST |
---|---|---|---|
AVS | 99.90 ± 0.21% | 100.00 ± 0.00% | 100.00 ± 0.00% |
Four-Layer CNN | None | 44.80 ± 5.91% | 68.65 ± 7.16% |
LeNet-5 | None | 35.65 ± 7.75% | 66.45 ± 9.24% |
EfficientNet-B0 | 57.30 ± 12.52% | 78.80 ± 12.04% | 78.70 ± 8.62% |
ResNet50 | 62.85 ± 6.25% | 67.45 ± 8.27% | 89.19 ± 10.46% |
ConvNeXt | 36.05 ± 12.31% | 58.05 ± 21.41% | 67.45 ± 16.32% |
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. |
© 2024 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).
Share and Cite
Chen, T.; Todo, Y.; Takano, R.; Qiu, Z.; Hua, Y.; Tang, Z. A Learning Dendritic Neuron-Based Motion Direction Detective System and Its Application to Grayscale Images. Brain Sci. 2024, 14, 864. https://doi.org/10.3390/brainsci14090864
Chen T, Todo Y, Takano R, Qiu Z, Hua Y, Tang Z. A Learning Dendritic Neuron-Based Motion Direction Detective System and Its Application to Grayscale Images. Brain Sciences. 2024; 14(9):864. https://doi.org/10.3390/brainsci14090864
Chicago/Turabian StyleChen, Tianqi, Yuki Todo, Ryusei Takano, Zhiyu Qiu, Yuxiao Hua, and Zheng Tang. 2024. "A Learning Dendritic Neuron-Based Motion Direction Detective System and Its Application to Grayscale Images" Brain Sciences 14, no. 9: 864. https://doi.org/10.3390/brainsci14090864