Find Research That Solves Problems

To develop a rigorous theoretical framework for emergent intelligence and scaling laws in foundation models using limit theory.

Thispaperexploresthephenomenaoftheemergenceoftheuseofartificialintelligence inteachingandlearninginhighereducation.Itinvestigateseducationalimplicationsof emergingtechnologiesonthewaystudentslearnandhowinstitutionsteachandevolve. Recenttechnologicaladvancementsandtheincreasingspeedofadoptingnew technologiesinhighereducationareexploredinordertopredictthefuturenatureof highereducationinaworldwhereartificialintelligenceispartofthefabricofour universities.Wepinpointsomechallengesforinstitutionsofhighereducationand studentlearningintheadoptionofthesetechnologiesforteaching,learning,student support,andadministrationandexplorefurtherdirectionsforresearch.

Knowledge Graphs (KGs) store structured factual information as RDF triples, but their util- ity depends on preservation through NLP transformations. This thesis presents QUASAR (Query-oriented System for Multi-hop Knowledge Graph Data Summarisation), a systematic investigation into information loss when KG triples are verbalized into natural language text, optionally summarized, and then reconstructed back into a KG. We implement a complete six-stage round-trip evaluation pipeline on the WebQSP dataset over Freebase, testing 6 pipeline variants formed by combining 2 verbalization methods (rule-based templatesvs.neuralFlan-T5-XL)with3summarizationlevels(extractiveTextRank, abstractive Flan-T5-XL,andnosummarization). EachvariantproducesareconstructedKGviatheREBEL relation extraction model, which is then evaluated by executing the original 1,639 test SPARQL queries and comparing predicted answers against ground truth. Our key (cid:28)ndings are: (1) the reconstruction stage is the dominant information bottleneck, with REBELrecoveringonly∼12%oftriples(ℓ 5 ≈ 0.88);(2)rule-basedverbalizationconsistentlyout- performs neural across all downstream metrics (Answer F1: 0.0295 vs. 0.0275 for best variants); (3) extractive summarization preserves 93% of recoverable information while abstractive loses 44%; (4) 2-hop queries involving CVT intermediate nodes are essentially unrecoverable across all approaches; and (5) even the identity baseline (original 810K-triple KG) can only answer 28.8% of cross-split test queries, establishing a hard ceiling. The cumulative preservation through the best pipeline variant (nosumm_rb) is 7.77%, demon- strating that text remains a highly lossy compression format for structured knowledge. These (cid:28)ndingshaveimplicationsforKGstorage, transfer, andaugmentationvialargelanguagemodels.

To track progress along the WHO Neglected Tropical Diseases Road Map to 2030 using a Gap Assessment Tool (GAT) and provide a guide to its methodology and results.

To bridge the modality gap between forensic sketches and mugshots for improved face recognition accuracy.

Cryptography, optimization, simulation, and machine learning are just a few of the industries that might be completely transformed by quantum computing. This abstract gives a thorough introduction to quantum computing with an emphasis on its algorithms, architectures, and applications. In conclusion, this abstract offers an in-depth analysis of quantum computing, including its algorithms, structures, and applications. It highlights the revolutionary potential of quantum computing in tackling difficult issues that are beyond the scope of conventional computers, laying the groundwork for further research and understanding of this quickly developing topic.

Thispaperexploresthephenomenaoftheemergenceoftheuseofartificialintelligence inteachingandlearninginhighereducation.Itinvestigateseducationalimplicationsof emergingtechnologiesonthewaystudentslearnandhowinstitutionsteachandevolve. Recenttechnologicaladvancementsandtheincreasingspeedofadoptingnew technologiesinhighereducationareexploredinordertopredictthefuturenatureof highereducationinaworldwhereartificialintelligenceispartofthefabricofour universities.Wepinpointsomechallengesforinstitutionsofhighereducationand studentlearningintheadoptionofthesetechnologiesforteaching,learning,student support,andadministrationandexplorefurtherdirectionsforresearch.

—Automated detection of bacterial flagellar motors in cryo-electron tomography (cryo-ET) is challenged by extremely low signal-to-noise ratios, limited annotations, and complex macromolecular structures. We propose DA-SSL, a domain- adaptive self-supervised learning framework that pretrains a YOLOv8s backbone on unlabeled cryo-ET data to learn domain- specific noise characteristics and structural priors. Pretraining is guided by a masked reconstruction objective with a structural consistency regularizer, followed by systematic weight transfer andsupervisedfine-tuningwithaninitialbackbone-freezestrategy. The proposed approach achieves mAP@0.5 = 84.47% and precision = 85.05%, outperforming supervised baselines while maintaining real-time inference capability. Ablation studies show that both pretraining and freezing contribute non-redundant performance gains, and evaluation on unseen data confirms improved generalization.

This report presents an end-to-end implementation of implicit-feedback collaborative fil- tering, centered on Alternating Least Squares (ALS) with confidence weighting inspired by the classical work of Hu, Koren, and Volinsky [? ]. The pipeline processes MovieLens 100K interactions, binarizes explicit ratings into implicit preference signals, performs leave-one-out evaluation, and reports ranking metrics (HR@10, NDCG@10, and F1@10). In addition to ALS, we implement and evaluate two practical baselines: a popularity- based recommender and an ItemKNN model. The project contributes three outcomes. First, it provides a reproducible training and evaluation pipeline that can be adapted to arbitrary collaborative filtering datasets. Sec- ond, it integrates hyperparameter search and multi-seed evaluation to reduce variance and improve confidence in reported results. Third, it produces a comprehensive visual analytics package to support interpretation and communication of outcomes. Under the current high-score sampled-negative protocol (Top-K=10 and one sampled negativeperuser), allthreemodelsreachHR@10=1.0000andNDCG@10near0.93, with ItemKNNslightlyleadingNDCG.Wealsodiscusswhythissettingcaninflatemetricsand how stricter protocols influence conclusions. The report concludes with reproducibility notes, limitations, and practical recommendations for fair benchmarking.

Vector quantization, a problem rooted in Shannon’s source coding theory, aims to quantize high-dimensionalEuclideanvectorswhileminimizingdistortionintheirgeometricstructure. We propose TurboQuant to address both mean-squared error (MSE) and inner product distor- tion, overcoming limitations of existing methods that fail to achieve optimal distortion rates. Our data-oblivious algorithms, suitable for online applications, achieve near-optimal distortion rates (within a small constant factor) across all bit-widths and dimensions. TurboQuant achieves this by randomly rotating input vectors, inducing a concentrated Beta distribution on coordinates, and leveraging the near-independence property of distinct coordinates in high dimensions to simply apply optimal scalar quantizers per each coordinate. Recognizing that MSE-optimal quantizers introduce bias in inner product estimation, we propose a two-stage ap- proach: applying an MSE quantizer followed by a 1-bit Quantized JL (QJL) transform on the residual, resulting in an unbiased inner product quantizer. We also provide a formal proof of the information-theoretic lower bounds on best achievable distortion rate by any vector quan- tizer, demonstratingthatTurboQuantcloselymatchesthesebounds, differingonlybyasmall constant ( ≈ 2.7) factor. Experimental results validate our theoretical findings, showing that for KV cache quantization, we achieve absolute quality neutrality with 3.5 bits per channel and marginalqualitydegradationwith2.5bitsperchannel. Furthermore, innearestneighborsearch tasks, ourmethodoutperformsexistingproductquantizationtechniquesinrecallwhilereducing indexing time to virtually zero.

Knowledge Graphs (KGs) store structured factual information as RDF triples, but their util- ity depends on preservation through NLP transformations. This thesis presents QUASAR (Query-oriented System for Multi-hop Knowledge Graph Data Summarisation), a systematic investigation into information loss when KG triples are verbalized into natural language text, optionally summarized, and then reconstructed back into a KG. We implement a complete six-stage round-trip evaluation pipeline on the WebQSP dataset over Freebase, testing 6 pipeline variants formed by combining 2 verbalization methods (rule-based templatesvs.neuralFlan-T5-XL)with3summarizationlevels(extractiveTextRank, abstractive Flan-T5-XL,andnosummarization). EachvariantproducesareconstructedKGviatheREBEL relation extraction model, which is then evaluated by executing the original 1,639 test SPARQL queries and comparing predicted answers against ground truth. Our key (cid:28)ndings are: (1) the reconstruction stage is the dominant information bottleneck, with REBELrecoveringonly∼12%oftriples(ℓ 5 ≈ 0.88);(2)rule-basedverbalizationconsistentlyout- performs neural across all downstream metrics (Answer F1: 0.0295 vs. 0.0275 for best variants); (3) extractive summarization preserves 93% of recoverable information while abstractive loses 44%; (4) 2-hop queries involving CVT intermediate nodes are essentially unrecoverable across all approaches; and (5) even the identity baseline (original 810K-triple KG) can only answer 28.8% of cross-split test queries, establishing a hard ceiling. The cumulative preservation through the best pipeline variant (nosumm_rb) is 7.77%, demon- strating that text remains a highly lossy compression format for structured knowledge. These (cid:28)ndingshaveimplicationsforKGstorage, transfer, andaugmentationvialargelanguagemodels.

Vector quantization, a problem rooted in Shannon’s source coding theory, aims to quantize high-dimensionalEuclideanvectorswhileminimizingdistortionintheirgeometricstructure. We propose TurboQuant to address both mean-squared error (MSE) and inner product distor- tion, overcoming limitations of existing methods that fail to achieve optimal distortion rates. Our data-oblivious algorithms, suitable for online applications, achieve near-optimal distortion rates (within a small constant factor) across all bit-widths and dimensions. TurboQuant achieves this by randomly rotating input vectors, inducing a concentrated Beta distribution on coordinates, and leveraging the near-independence property of distinct coordinates in high dimensions to simply apply optimal scalar quantizers per each coordinate. Recognizing that MSE-optimal quantizers introduce bias in inner product estimation, we propose a two-stage ap- proach: applying an MSE quantizer followed by a 1-bit Quantized JL (QJL) transform on the residual, resulting in an unbiased inner product quantizer. We also provide a formal proof of the information-theoretic lower bounds on best achievable distortion rate by any vector quan- tizer, demonstratingthatTurboQuantcloselymatchesthesebounds, differingonlybyasmall constant ( ≈ 2.7) factor. Experimental results validate our theoretical findings, showing that for KV cache quantization, we achieve absolute quality neutrality with 3.5 bits per channel and marginalqualitydegradationwith2.5bitsperchannel. Furthermore, innearestneighborsearch tasks, ourmethodoutperformsexistingproductquantizationtechniquesinrecallwhilereducing indexing time to virtually zero.