Achieving a good measure of model generalization remains a challenge within machine learning. The goal of this project is to investigate alternative techniques for evaluating a machine learning model's expected generalization to unseen data. A core inspiration is observation that activations in the brain's visual system broadly classifies objects during processing, with similar objects causing similar neural activations. Theoretically, a machine learning model that mirrored this observation would indicate a consistent understanding of the stimuli class, indicating an ability to extrapolate to unseen stimuli or stimuli variations. We evaluate models for this property utilizing representational dissimilarity matrices (RDMs), which abstractly represent a model as it's activation similarity between pairs of responses to stimuli. The project experiments with a multitude of techniques to both construct and utilize RDMs in and across model training, validation, and testing.
The goal of this project is to establish the use of accurate benchmarks of machine learning models on real-world data. Machine learning models are often trained and tested on simplified representations of data that cannot be replicated in the real-world. Reported results thus inflate machine learning performance, while obfuscating expected performance "in the wild." In order to combat this trend, we investigate multimodal models which require no human intervention and have real-world, practical usage scenarios. In each domain, the model is implemented to minimize human intervention and to maximize the deployability on at-scale real-world data. A further goal of this project is to monitor how specific feature techniques may generalize across domains, highlighting techniques that generalize across a multitude of domains with reasonable accuracy.
This project addresses the inward-facing and important problem of athlete skill level characterization and tracking. The goal of the project is to measure performance-critical quantities unobtrusively, rapidly, and frequently. The team pursues this goal through the design and deployment of an measurement platform incorporating force plate sensors (that measure the 3D distribution of force produced by a stationary, moving, or jumping athlete) and skeletal tracking based on markerless video analysis.
MICrONS sought to revolutionize machine learning by reverse-engineering the algorithms of the brain. The program was expressly designed as a dialogue between data science and neuroscience. Ultimate computational goals for MICrONS included the ability to perform complex information processing tasks such as one-shot learning, unsupervised clustering, and scene parsing, aiming towards human-like proficiency.
CLASS 5.0 autonomously measured and assessed key indices of classroom discourse. The project's ultimate goal was to provide feedback to teachers on their use of dialogic instruction in classrooms. CLASS 5.0 is a research tool used to non-intrusively and inexpensively build profiles of classroom interaction and to investigate the effects of classroom discourse on student achievement.