{"id":{"repo_id":"regina","oai_identifier":"oai:uregina.scholaris.ca:10294/9258"},"canonical_url":"https://search.dev.ndltd.org/etd/regina/oai:uregina.scholaris.ca:10294/9258","repository":{"repo_id":"regina","name":"University of Regina","base_url":"https://uregina.scholaris.ca/server/oai/request"},"display":{"title":"An Empirical Study on Sum-Product Networks Structure Learning and Deep Convolutional Sum-Product Networks","abstract":"Sum-product Networks (SPNs) are generative probabilistic models which obtained the attention of the community for its tractability and easy methods for learning. Our rst main contribution is an empirical comparison of methods for SPN learning and inference. LearnSPN is a popular algorithm for learning SPNs that utilizes chop and slice operations. As g-test is a standard chopping method and Gaussian mixture models (GMM) using expectation-maximization is a common slicing method, it seems to have been assumed in the literature that this is the best pair in LearnSPN. On the contrary, our results show that g-test for chopping and k-means for slicing yields SPNs that are just as accurate. Moreover, it has been shown that implementing SPN leaf nodes as Chow-Liu Trees (CLTs) yields more accurate SPNs for the former pair. Our experiments show the same for the latter pair, and that neither pair dominates the other. Lastly, we report an analysis of SPN topology for unstudied pairs. Deep convolutional sum-product networks (DCSPNs) have very recently been in- troduced and shown to yield competitive results in image completion tasks. A DCSPN consists of a tree structure (a directed acyclic graph) coupled with parameters of the structure. Given that DCSPNs are in their infancy, many open questions remain regarding the properties and topology of their tree structure. As our second main contribution, we undertake three investigations pertaining to the DCSPN structure. The rst two studies revolve around the original structure put forth in the seminal paper. These studies increase the number of pooling layers and vary the hyperparam- eters in attempts to improve accuracy. The third inquiry suggests a new DCSPN tree structure that signi cantly lowers the training time at a modest expense of accuracy.","abstract_html":"Sum-product Networks (SPNs) are generative probabilistic models which obtained the attention of the community for its tractability and easy methods for learning. Our rst main contribution is an empirical comparison of methods for SPN learning and inference. LearnSPN is a popular algorithm for learning SPNs that utilizes chop and slice operations. As g-test is a standard chopping method and Gaussian mixture models (GMM) using expectation-maximization is a common slicing method, it seems to have been assumed in the literature that this is the best pair in LearnSPN. On the contrary, our results show that g-test for chopping and k-means for slicing yields SPNs that are just as accurate. Moreover, it has been shown that implementing SPN leaf nodes as Chow-Liu Trees (CLTs) yields more accurate SPNs for the former pair. Our experiments show the same for the latter pair, and that neither pair dominates the other. Lastly, we report an analysis of SPN topology for unstudied pairs. Deep convolutional sum-product networks (DCSPNs) have very recently been in- troduced and shown to yield competitive results in image completion tasks. A DCSPN consists of a tree structure (a directed acyclic graph) coupled with parameters of the structure. Given that DCSPNs are in their infancy, many open questions remain regarding the properties and topology of their tree structure. As our second main contribution, we undertake three investigations pertaining to the DCSPN structure. The rst two studies revolve around the original structure put forth in the seminal paper. These studies increase the number of pooling layers and vary the hyperparam- eters in attempts to improve accuracy. The third inquiry suggests a new DCSPN tree structure that signi cantly lowers the training time at a modest expense of accuracy.","abstract_has_math":false,"creators":["Lobo Teixeira, Andra"],"institution":"Faculty of Graduate Studies and Research, University of Regina","degree_name":"Master of Science (MSc)","degree_level":"Master's","degree_discipline":"Computer Science","degree_department":null,"school":null,"contributors":[],"advisors":["Butz, Cortney J."],"committee_chairs":[],"committee_members":["Fan, Lisa","Yang, Boting"],"year":2020,"date_issued":"2020-01","date_published":"2020-01","updated_at":"2026-07-24T04:03:41Z","subjects":[],"languages":["en"],"rights":[],"rights_urls":[],"identifier_entries":[{"key":"dc:identifier.doi","label":"DOI","values":["https://doi.org/10.82465/4622"],"render_values":[{"text":"https://doi.org/10.82465/4622","href":"https://doi.org/10.82465/4622","code":true}]}]},"links":{"outbound_url":"https://hdl.handle.net/10294/9258","outbound_label":"Handle","outbound_source":"dc:identifier.uri"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor.advisor","label":"Advisor","values":["Butz, Cortney J."]},{"key":"dc:contributor.committeemember","label":"Committee Member","values":["Fan, Lisa","Yang, Boting"]},{"key":"dc:creator","label":"Author","values":["Lobo Teixeira, Andra"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date.accessioned","label":"Dc Date Accessioned","values":["2020-08-29T19:15:45Z"]},{"key":"dc:date.available","label":"Dc Date Available","values":["2020-08-29T19:15:45Z"]},{"key":"dc:date.issued","label":"Date","values":["2020-01"]},{"key":"dc:publisher","label":"Institution","values":["Faculty of Graduate Studies and Research, University of Regina"]},{"key":"dc:type","label":"Dc Type","values":["master thesis"]},{"key":"thesis:degree_discipline","label":"Discipline","values":["Computer Science"]},{"key":"thesis:degree_level","label":"Degree Level","values":["Master's"]},{"key":"thesis:degree_name","label":"Degree Name","values":["Master of Science (MSc)"]},{"key":"thesis:institution_name","label":"Thesis Institution Name","values":["Faculty of Graduate Studies and Research, University of Regina"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:language.iso","label":"Language (ISO)","values":["en"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier.doi","label":"DOI","values":["https://doi.org/10.82465/4622"]},{"key":"dc:identifier.uri","label":"Identifier URI","values":["https://hdl.handle.net/10294/9258"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description","label":"Description","values":["A Thesis Submitted to the Faculty of Graduate Studies and Research In Partial Fulfillment of the Requirements for the Degree of Master of Science in Computer Science, University of Regina. viii, 53 p."]},{"key":"dc:description.abstract","label":"Abstract","values":["Sum-product Networks (SPNs) are generative probabilistic models which obtained the attention of the community for its tractability and easy methods for learning. Our rst main contribution is an empirical comparison of methods for SPN learning and inference. LearnSPN is a popular algorithm for learning SPNs that utilizes chop and slice operations. As g-test is a standard chopping method and Gaussian mixture models (GMM) using expectation-maximization is a common slicing method, it seems to have been assumed in the literature that this is the best pair in LearnSPN. On the contrary, our results show that g-test for chopping and k-means for slicing yields SPNs that are just as accurate. Moreover, it has been shown that implementing SPN leaf nodes as Chow-Liu Trees (CLTs) yields more accurate SPNs for the former pair. Our experiments show the same for the latter pair, and that neither pair dominates the other. Lastly, we report an analysis of SPN topology for unstudied pairs. Deep convolutional sum-product networks (DCSPNs) have very recently been in- troduced and shown to yield competitive results in image completion tasks. A DCSPN consists of a tree structure (a directed acyclic graph) coupled with parameters of the structure. Given that DCSPNs are in their infancy, many open questions remain regarding the properties and topology of their tree structure. As our second main contribution, we undertake three investigations pertaining to the DCSPN structure. The rst two studies revolve around the original structure put forth in the seminal paper. These studies increase the number of pooling layers and vary the hyperparam- eters in attempts to improve accuracy. The third inquiry suggests a new DCSPN tree structure that signi cantly lowers the training time at a modest expense of accuracy."]},{"key":"dc:title","label":"Title","values":["An Empirical Study on Sum-Product Networks Structure Learning and Deep Convolutional Sum-Product Networks"]}]}],"canonical_facts":{"dc:contributor.advisor":["Butz, Cortney J."],"dc:contributor.committeemember":["Fan, Lisa","Yang, Boting"],"dc:creator":["Lobo Teixeira, Andra"],"dc:date.accessioned":["2020-08-29T19:15:45Z"],"dc:date.available":["2020-08-29T19:15:45Z"],"dc:date.issued":["2020-01"],"dc:description":["A Thesis Submitted to the Faculty of Graduate Studies and Research In Partial Fulfillment of the Requirements for the Degree of Master of Science in Computer Science, University of Regina. viii, 53 p."],"dc:description.abstract":["Sum-product Networks (SPNs) are generative probabilistic models which obtained the attention of the community for its tractability and easy methods for learning. Our rst main contribution is an empirical comparison of methods for SPN learning and inference. LearnSPN is a popular algorithm for learning SPNs that utilizes chop and slice operations. As g-test is a standard chopping method and Gaussian mixture models (GMM) using expectation-maximization is a common slicing method, it seems to have been assumed in the literature that this is the best pair in LearnSPN. On the contrary, our results show that g-test for chopping and k-means for slicing yields SPNs that are just as accurate. Moreover, it has been shown that implementing SPN leaf nodes as Chow-Liu Trees (CLTs) yields more accurate SPNs for the former pair. Our experiments show the same for the latter pair, and that neither pair dominates the other. Lastly, we report an analysis of SPN topology for unstudied pairs. Deep convolutional sum-product networks (DCSPNs) have very recently been in- troduced and shown to yield competitive results in image completion tasks. A DCSPN consists of a tree structure (a directed acyclic graph) coupled with parameters of the structure. Given that DCSPNs are in their infancy, many open questions remain regarding the properties and topology of their tree structure. As our second main contribution, we undertake three investigations pertaining to the DCSPN structure. The rst two studies revolve around the original structure put forth in the seminal paper. These studies increase the number of pooling layers and vary the hyperparam- eters in attempts to improve accuracy. The third inquiry suggests a new DCSPN tree structure that signi cantly lowers the training time at a modest expense of accuracy."],"dc:identifier.doi":["https://doi.org/10.82465/4622"],"dc:identifier.uri":["https://hdl.handle.net/10294/9258"],"dc:language.iso":["en"],"dc:publisher":["Faculty of Graduate Studies and Research, University of Regina"],"dc:title":["An Empirical Study on Sum-Product Networks Structure Learning and Deep Convolutional Sum-Product Networks"],"dc:type":["master thesis"],"thesis:degree_discipline":["Computer Science"],"thesis:degree_level":["Master's"],"thesis:degree_name":["Master of Science (MSc)"],"thesis:institution_name":["Faculty of Graduate Studies and Research, University of Regina"]},"updated_at":"2026-07-24T04:03:41Z"}