The human brain is comprised of many billions of neurons, which then connect with each other many trillions of times. Modeling how these neurons function provides insight into language, thought, and behavior. However, neurons are not all identical, so an approach that performs well on some neurons will perform poorly on others. I focused on a learning system to detect and model neurons that use pattern-recognition in order to decide when to fire, which would not be expressed well under existing approaches. We researched a strategy for describing these process in a non-Markov state, which means that we do not have enough information to correctly model the neuron, by approximating to Markov models using a subset of the required information. This means that we produce a set of probabilities for the next state of the neuron given the previous state. We approximate the Markov model using Monte-Carlo Tree Search (MCTS) optimizing for smallest confidence interval to select sequences to measure. In order to calculate a confidence interval on a given sequence we apply A*, pronounced A star, which is a targeted pathfinding algorithm, in order to produce execution paths that in turn find confidence intervals. We use two strategies for pruning sequences; a MCTS to find the measured sequence that differs least from chance for long term learning, and exponential decay for short term learning. We assume that after we process information it cannot be recovered, so when we measure new sequences we only measure from the time of consideration onward. If our process successfully locates important patterns the neuron is looking for, it produces an accurate approximation for the neuron that can be rapidly evaluated.

With the spread of large digital painting collections, the demand for automatic theme classification has risen in both academic and commercial fields. Many researchers have investigated various techniques to implement automatic painting classification and developed several approaches to improve classification accuracy. SURF is one of the most commonly used local invariant feature descriptors for painting classification. However, the common method of classification based on traditional SURF local feature description makes the color feature information not comprehensive. Also, the feature space is high dimensional and sparse which will result in low classification accuracy, data redundancy and a time-consuming process. To overcome this drawback, we suggest a new method of classifying painting styles with bag of words (BOW) model based on the RGB-SURF algorithm. First, split RGB channels into separate images for each painting. Then, independently calculate key points and descriptor of each RGB channel using SURF and combine descriptors of each channel to obtain overall RGB-SURF feature for each painting. This approach is to compensate the loss of color feature brought by traditional SURF descriptor which mainly treats colorful images as gray images. Next, construct the visual vocabulary by reducing the number of features through quantization of feature space using K-means clustering to obtain BOW model. The histogram described by appearance frequency of the visual words is used to represent the content of the image. As a result, every image is viewed as a bag full of visual words. Finally, establish the support vector machine (SVM) classifier based on radial basis function (RBF) with the data above for training and testing. Compared to classification method based on the traditional SURF feature descriptor and other painting classification approaches, BOW method based on RGB-SURF algorithm shows the highest average classification accuracy, which means this method produces the most consistent automatic classification results with manual classification.