Wisdom O. Ikezogwo

Leading cutting-edge research in multimodal AI for medical imaging and advanced video generation, driving innovation in state-of-the-art models and dataset creation.

• Spearheaded research on Multimodal Large Language Models (LLMs) for medical imaging, developing novel medical multimodal datasets (Quilt-1M, MedNarratives) and state-of-the-art models (QuiltNet, Quilt-LLaVA).
• Engineered a multi-agent AI framework, PathFinder, for clinical diagnosis, achieving performance superior to human experts and establishing improved benchmarks (MedBlink).
• Directing efforts to integrate Newtonian physics into image and video generative models, focusing on generating large-scale, temporally consistent video scene graph datasets and pipelines.

Provided comprehensive instructional and academic support for undergraduate and graduate courses in Computer Science.

• Taught Data Programming (CSE 160) for Fall 2021, Winter 2022, and Fall 2022 semesters, guiding students through fundamental concepts and practical applications.
• Provided instruction and support for Introduction to Artificial Intelligence (CSE 473) during Spring 2023, Fall 2023, and Winter 2024 academic terms.

Led research on efficient multimodal representations for egocentric data, focusing on reducing data capture costs.

• Pioneered research in efficient multimodal representations for egocentric data, including video, text, audio, IMU, and hands.
• Developed 'Perceive-Predict,' a novel framework leveraging predictive coding between co-occurring modalities to reconstruct missing data, significantly reducing capture costs for expensive modalities like video.

Directed research efforts to develop a foundational model for histopathology using large-scale image datasets.

• Led research towards developing a foundational model for histopathology, training on millions of gigapixel-sized histology images.
• Scaled-up compute operations on the Argonne National Lab computing cluster, resulting in clinically evaluated models for improved diagnostic capabilities.

Developed machine learning models for financial information extraction and integration into lending systems.

• Engineered models to extract crucial customer financial information from unstructured banking data.
• Processed key customer earning and spending data to feed into downstream lending pipelines, including predicting income, analyzing spending patterns, and performing reconciliations.

Designed and implemented production-grade recommendation systems for enhanced user experience.

• Developed and deployed production recommendation systems utilizing advanced techniques such as attention mechanisms and epsilon-greedy bandit strategies.

Conducted research in biosignal processing, focusing on EEG signal analysis and neural network applications.

• Integrated disparate multivariate time series data, employing spectral component characterization for dynamical dimensionality reduction.
• Developed and trained neural networks for accurate classification and characterization of brain EEG signals.

Awarded $100,000 for pioneering AI research in population health.

Received $20,000 grant to advance research in foundational AI models.

Awarded travel grant to attend a prestigious computational neuroscience summer school in Cape Town.

Recognized as the top-performing student in the Control Engineering specialization.

Awarded for outstanding academic achievement as the best male graduating student.

Received a cumulative scholarship valued at $1500 for academic excellence.

Awarded a cumulative scholarship valued at $1500 based on national merit.

Received a scholarship valued at $250 for academic merit.

Introduces a novel approach for visual instruction tuning using localized narratives from histopathology videos to enhance multimodal LLMs.

Presents a large-scale dataset of one million image-text pairs specifically curated for histopathology, enabling advanced medical imaging research.

Explores the use of multimodal masked autoencoders to learn compositional representations from histopathological data, improving diagnostic capabilities.

Applies whole slide image analysis for risk stratification of solitary fibrous tumors, contributing to more precise medical diagnostics.

A collaborative reflection on the latest advancements, applications, and challenges in machine learning for health, derived from research roundtables.

Proposes a multi-modal, multi-agent framework to enhance diagnostic decision-making processes in histopathology.

Explores linking medical vision and language through procedural and localized narratives to create a comprehensive framework for medical imaging analysis.

Investigates the foundational medical imaging knowledge embedded within multimodal language models to assess their understanding and capabilities.

Developing efficient multi-modal representations for egocentric data, incorporating modality and time awareness for improved predictive capabilities.

Focuses on generating temporally consistent video frames by integrating physics principles, ensuring realistic motion and interactions.

Research on generating synthetic video scene graphs to improve understanding and manipulation of complex video content.

Introduces MsCAMIL network for automated triage of colorectal polyps, enhancing efficiency in pathological diagnosis.

Compares the diagnostic performance of MsCAMIL against human pathology trainees in colorectal polyp diagnosis.

Investigates domain generalization techniques to integrate diverse EEG datasets for improved automatic epileptic seizure detection.

Provides an empirical characterization of the temporal dynamics of EEG spectral components, contributing to a deeper understanding of brain activity.