{"id":{"repo_id":"bu","oai_identifier":"oai:open.bu.edu:2144/49896"},"canonical_url":"https://search.ndltd.org/etd/bu/oai:open.bu.edu:2144/49896","repository":{"repo_id":"bu","name":"Boston University","base_url":"https://open.bu.edu/oai/request"},"display":{"title":"MixtureGrowth: growing neural networks by recombining learned parameters","abstract":"Most deep neural networks are trained under fixed network architectures and require retraining when the architecture changes. If expanding the network's size is needed, it is necessary to retrain from scratch, which is expensive. To avoid this, one can grow from a small network by adding random weights over time to gradually achieve the target network size. However, this naive approach falls short in practice as it brings too much noise to the growing process. Prior work tackled this issue by leveraging the already learned weights and training data for generating new weights through conducting a computationally expensive analysis step. In this thesis, we introduce MixtureGrowth, a new approach to growing networks that circumvents the initialization overhead in prior work. Before growing, each layer in our model is generated with a linear combination of parameter templates. Newly grown layer weights are generated by using a new linear combination of existing templates for a layer. On one hand, these templates are already trained for the task, providing a strong initialization. On the other, the new coefficients provide flexibility for the added layer weights to learn something new. We show that our approach boosts top-1 accuracy over the state-of-the-art by 2% points on CIFAR-100 and ImageNet datasets while achieving comparable performance with fewer FLOPs to the larger network trained from scratch.","abstract_html":"Most deep neural networks are trained under fixed network architectures and require retraining when the architecture changes. If expanding the network&#x27;s size is needed, it is necessary to retrain from scratch, which is expensive. To avoid this, one can grow from a small network by adding random weights over time to gradually achieve the target network size. However, this naive approach falls short in practice as it brings too much noise to the growing process. Prior work tackled this issue by leveraging the already learned weights and training data for generating new weights through conducting a computationally expensive analysis step. In this thesis, we introduce MixtureGrowth, a new approach to growing networks that circumvents the initialization overhead in prior work. Before growing, each layer in our model is generated with a linear combination of parameter templates. Newly grown layer weights are generated by using a new linear combination of existing templates for a layer. On one hand, these templates are already trained for the task, providing a strong initialization. On the other, the new coefficients provide flexibility for the added layer weights to learn something new. We show that our approach boosts top-1 accuracy over the state-of-the-art by 2% points on CIFAR-100 and ImageNet datasets while achieving comparable performance with fewer FLOPs to the larger network trained from scratch.","abstract_has_math":false,"creators":["Pham, Chau"],"institution":"Boston University","degree_name":"Master of Science","degree_level":"masters","degree_discipline":"Computer Science","degree_department":null,"school":null,"contributors":[],"advisors":["Plummer, Bryan A."],"committee_chairs":[],"committee_members":[],"year":2024,"date_issued":"2024","date_published":"2024","updated_at":"2026-11-18T20:42:50Z","subjects":["Computer science","Artificial intelligence","Efficient deep learning","Network growth","Neural network"],"languages":["en_US"],"rights":[],"rights_urls":[],"identifier_entries":[],"person_identifier_entries":[]},"links":{"outbound_url":"https://hdl.handle.net/2144/49896","outbound_label":"Handle","outbound_source":"dc:identifier.uri","record_url":"https://hdl.handle.net/2144/49896","record_label":"Repository record","record_kind":"handle","record_source":"dc:identifier.uri","download_url":null,"download_label":null,"download_source":null,"access_status":null,"embargo_until":null,"access_note":null},"source_record":{"url":"https://open.bu.edu/oai/request?verb=GetRecord&metadataPrefix=dim&identifier=oai%3Aopen.bu.edu%3A2144%2F49896","prefix":"dim"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor.advisor","label":"Advisor","values":["Plummer, Bryan A."]},{"key":"dc:creator","label":"Author","values":["Pham, Chau"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date.accessioned","label":"Dc Date Accessioned","values":["2025-03-06T15:42:49Z"]},{"key":"dc:date.available","label":"Dc Date Available","values":["2025-03-06T15:42:49Z"]},{"key":"dc:date.issued","label":"Date","values":["2024"]},{"key":"dc:date.updated","label":"Dc Date Updated","values":["2025-03-05T23:01:29Z"]},{"key":"dc:type","label":"Dc Type","values":["Thesis/Dissertation"]},{"key":"thesis:degree_discipline","label":"Discipline","values":["Computer Science"]},{"key":"thesis:degree_level","label":"Degree Level","values":["masters"]},{"key":"thesis:degree_name","label":"Degree Name","values":["Master of Science"]},{"key":"thesis:institution_name","label":"Thesis Institution Name","values":["Boston University"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["Computer science","Artificial intelligence","Efficient deep learning","Network growth","Neural network"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:language.iso","label":"Language (ISO)","values":["en_US"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier.uri","label":"Identifier URI","values":["https://hdl.handle.net/2144/49896"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description","label":"Description","values":["2024"]},{"key":"dc:description.abstract","label":"Abstract","values":["Most deep neural networks are trained under fixed network architectures and require retraining when the architecture changes. If expanding the network's size is needed, it is necessary to retrain from scratch, which is expensive. To avoid this, one can grow from a small network by adding random weights over time to gradually achieve the target network size. However, this naive approach falls short in practice as it brings too much noise to the growing process. Prior work tackled this issue by leveraging the already learned weights and training data for generating new weights through conducting a computationally expensive analysis step. In this thesis, we introduce MixtureGrowth, a new approach to growing networks that circumvents the initialization overhead in prior work. Before growing, each layer in our model is generated with a linear combination of parameter templates. Newly grown layer weights are generated by using a new linear combination of existing templates for a layer. On one hand, these templates are already trained for the task, providing a strong initialization. On the other, the new coefficients provide flexibility for the added layer weights to learn something new. We show that our approach boosts top-1 accuracy over the state-of-the-art by 2% points on CIFAR-100 and ImageNet datasets while achieving comparable performance with fewer FLOPs to the larger network trained from scratch."]},{"key":"dc:title","label":"Title","values":["MixtureGrowth: growing neural networks by recombining learned parameters"]}]}],"canonical_facts":{"dc:contributor.advisor":["Plummer, Bryan A."],"dc:creator":["Pham, Chau"],"dc:date.accessioned":["2025-03-06T15:42:49Z"],"dc:date.available":["2025-03-06T15:42:49Z"],"dc:date.issued":["2024"],"dc:date.updated":["2025-03-05T23:01:29Z"],"dc:description":["2024"],"dc:description.abstract":["Most deep neural networks are trained under fixed network architectures and require retraining when the architecture changes. If expanding the network's size is needed, it is necessary to retrain from scratch, which is expensive. To avoid this, one can grow from a small network by adding random weights over time to gradually achieve the target network size. However, this naive approach falls short in practice as it brings too much noise to the growing process. Prior work tackled this issue by leveraging the already learned weights and training data for generating new weights through conducting a computationally expensive analysis step. In this thesis, we introduce MixtureGrowth, a new approach to growing networks that circumvents the initialization overhead in prior work. Before growing, each layer in our model is generated with a linear combination of parameter templates. Newly grown layer weights are generated by using a new linear combination of existing templates for a layer. On one hand, these templates are already trained for the task, providing a strong initialization. On the other, the new coefficients provide flexibility for the added layer weights to learn something new. We show that our approach boosts top-1 accuracy over the state-of-the-art by 2% points on CIFAR-100 and ImageNet datasets while achieving comparable performance with fewer FLOPs to the larger network trained from scratch."],"dc:identifier.uri":["https://hdl.handle.net/2144/49896"],"dc:language.iso":["en_US"],"dc:subject":["Computer science","Artificial intelligence","Efficient deep learning","Network growth","Neural network"],"dc:title":["MixtureGrowth: growing neural networks by recombining learned parameters"],"dc:type":["Thesis/Dissertation"],"thesis:degree_discipline":["Computer Science"],"thesis:degree_level":["masters"],"thesis:degree_name":["Master of Science"],"thesis:institution_name":["Boston University"]},"updated_at":"2026-11-18T20:42:50Z"}