In this project, we checked the possibility of using "biological' features, which bears resemblance to the visual preprocessing pathways and Biological Neural Networks to analyze the complex multi-dimensional features for purposes of data recognition and classification. This project includes complex data analysis, adaptation of visual features to the neural microcircuit and comparison of the results to baseline algorithms.
In this project, we checked the possibility of using “biological’ features, which bears resemblance to the visual preprocessing pathways and Biological Neural Networks to analyze the complex multi-dimensional features for purposes of data recognition and classification. This project includes complex data analysis, adaptation of visual features to the neural microcircuit and comparison of the results to baseline algorithms.
Although there has been remarkable progress in pattern recognition algorithms in recent years, these algorithms still do not compare with the human ability for object recognition. Hence, in this project we used biologically motivated techniques for object recognition which in some cases engendered improved recognition results.
The steps of the algorithm are:
- Applying different filters on the images for extraction of “biological” features
- Transforming the features into spike trains
- Feeding the spike trains to neural network
- Running an linear Perceprton algorithm on the output for the network of the
The algorithm was created in Matlab-6.5 enviorment. In addition, the CSIM toolbox was used.
Results and Conclusions
Optimal parameters for the neural network
The neural network model is very complex and includes a large number of controllable parameters. Those parameters’ values greatly influence on the performance of the system.
1. The statistics of the spike trains – The spike trains in the human vision system are usually modeled using Poisson processes. 3 possibilities for the statistics were examined: Poisson, Gaussian and deterministic. The tests show that when using deterministic statistics, better results are achieved.
2. Maximal average frequency of the spike trains – the maximal frequency in the human vision system is close to 80Hz.The tests that were held show that it is preferable to use the maximal allowed frequency.
3. Spatial arrangement of the network’s inputs – Two possible arrangements were tested and the tests show the it is slightly better to use the arrangement that preserves the closeness between neighboring pixels, as opposed to column stack arrangement.
4. The network’s connectivity parameter – A too small connectivity might cause isolated neurons while too large connectivity might cause saturation.
5. Size of the neural network – The network is of high complexity, therefore it is essential to choose the minimal size while maintaining satisfactory classification results.
6. Using the neural network – It has been shown that using the network provides better performance in comparison to using a linear classifier on the network’s input.
Performance comparison to standard classification algorithms
1. Dimension reduction using PCA with a KNN classifier – The results achieved using this classifier have high correlation with those of the biological system. (correlation = 91%)
2. Dimension reduction using PCA with a linear classifier – the results are very similar to those of the previous classifier.
3. Linear SVM classifier – The mean performance is slightly higher than that of the PCA, but in this case, the correlation is very low.
Conclusions from the classification results using PCA
In this part the optimal features were selected for use of a biological classification system.
1. Features’ sizes – It was shown that using larger features’ size improves the classification performance conclusively.
2. Features’ types – It was shown that the best performance is achieved using the directional filters. Light intensity features provided average performance, while the color features performed in the worst way.
3. Features’ fusion – It was shown that a very high performance can be achieved by using pairs of features’ fields.
We would like to thank our supervisor Erez Berkovich for his support and guidance throughout this project.
Also we would like to thank Johanan Erez and Ina Krinski that helped us in every technical aspect and hardware issues.
Many thanks to the Ollendorff Minerva Center which supported this project.