Members-Only
Recent Talks & Demos are for members only
You must be an AI Tinkerers active member to view these talks and demos.
Scholé AI
This talk explores AI-Native Lifelong Learning, detailing the concepts and practical implications of continuously evolving artificial intelligence systems.
AI-Native Lifelong Learning
Adaptive AI platform using knowledge tracing for personalized, role-specific LLM fluency.
- RAGRAG (Retrieval-Augmented Generation) is the GenAI framework that grounds LLMs (like GPT-4) on external, verified data, drastically reducing model hallucinations and providing verifiable sources.RAG is a critical GenAI architecture: it solves the LLM 'hallucination' problem by inserting a retrieval step before generation. A user query is vectorized, then used to query an external knowledge base (e.g., a Pinecone vector database) for relevant document chunks (typically 512-token segments). These retrieved facts augment the original prompt, providing the LLM (e.g., Gemini or Llama 3) the specific, current, or proprietary context required. This process ensures the final response is accurate and grounded in domain-specific data, avoiding the high cost and latency of full model retraining.
- MultiModNMultiModN is an interpretable, modular network (NeurIPS 2023) that sequentially fuses any combination of modalities for multi-task prediction, offering inherent robustness against Missing Not-At-Random (MNAR) data bias.MultiModN (Multimodal, Multi-Task, Interpretable Modular Networks) is your next-gen architecture for complex AI tasks. This modular network sequentially fuses latent representations from any number or type of modality, providing granular, real-time predictive feedback for multiple tasks. Unlike parallel fusion baselines, MultiModN is robust to the critical issue of Missing Not-At-Random (MNAR) data: a first-of-its-kind, inherently MNAR-resistant approach. It matches performance benchmarks across 10 real-world tasks (e.g., medical diagnoses, academic performance) while remaining interpretable-by-design and fully composable at inference.
- Bayesian Knowledge TracingA predictive modeling framework that uses Hidden Markov Models to track a student's mastery of specific skills over time.Corbett and Anderson introduced Bayesian Knowledge Tracing (BKT) in 1994 to power Intelligent Tutoring Systems like Cognitive Tutor. The model uses four parameters (p-init, p-learn, p-guess, and p-slip) to estimate the probability that a learner has mastered a latent skill based on their performance history. By continuously updating these probabilities as students interact with problem sets, BKT allows software to personalize instruction: it identifies exactly when a student has reached the 95% mastery threshold and can move to the next concept.
- Deep Knowledge TracingDeep Knowledge Tracing uses Recurrent Neural Networks to model student learning trajectories and predict future performance with high precision.Piech and colleagues introduced Deep Knowledge Tracing (DKT) in 2015 to solve the limitations of traditional Bayesian Knowledge Tracing. By leveraging Long Short-Term Memory (LSTM) networks, DKT maps a student's history of interactions (correct and incorrect responses) onto a high-dimensional latent space. This allows the model to capture complex dependencies between distinct skills without manual tagging. In benchmarks using the Assistments 2009 dataset, DKT achieved an AUC of 0.86, significantly outperforming classical Markov models. It remains a foundational architecture for adaptive learning platforms that require real-time, data-driven mastery estimation.
- Deep Bayesian Knowledge TracingDeep Bayesian Knowledge Tracing integrates the interpretability of Bayesian models with the high-dimensional predictive power of recurrent neural networks to track student mastery.Deep Bayesian Knowledge Tracing (DBKT) solves the black-box problem of standard Deep Knowledge Tracing by injecting structured Bayesian priors into deep learning architectures. This hybrid approach uses Long Short-Term Memory (LSTM) networks to model complex learning curves while maintaining the clear performance parameters found in traditional Bayesian Knowledge Tracing (BKT). By mapping hidden neural states to explicit mastery probabilities, DBKT provides educators with actionable insights and 15 percent higher predictive accuracy on datasets like ASSISTments2009. It effectively balances the need for precise student modeling with the transparency required for pedagogical intervention.
- Performance Factor AnalysisPerformance Factor Analysis (PFA) is a predictive modeling framework that estimates student proficiency by tracking success and failure counts across specific skill sets.PFA improves upon the standard Bayesian Knowledge Tracing (BKT) model by accounting for the specific difficulty of items and the cumulative impact of a student's prior attempts. Developed by researchers like Pavlik, Cen, and Koedinger (2009), the algorithm uses a logistic regression approach to weigh successful practice more heavily than failures. This sensitivity to individual learning curves makes it a staple in Intelligent Tutoring Systems (ITS) like Carnegie Learning's MATHia, where it predicts future performance with higher accuracy than models that ignore the frequency of past mistakes. It is the go-to choice for developers needing a robust, computationally efficient way to personalize educational content in real-time.
Related projects
How AI helps in language learning on the go - personalized podcasts with spaced repetition
Lausanne
Learn how Podcastify AI uses personalized, AI-generated podcast lessons with spaced repetition for efficient language learning on Apple…
AIR3.ai
Lausanne
This talk introduces AIR3.ai, an agentic tool designed to streamline and automate the process of conducting literature reviews.
AI in Education
Pune
The talk explains Sahay, an AI platform using RAG and LLMs to deliver personalized learning paths, career counseling,…
Move Your Bubbles!
Lausanne
Learn how Neural Cellular Automata turn static images into naturally moving textures in 2D and 3D, with interactive…
AI-powered language quiz generator
Waterloo
This talk details building an AI language quiz generator, showing API instantiation, UI flow, and practical lessons learned…
Breaking free from the chat box
Valencia
Explore how AI can move beyond chat to enable structured data input through dynamic interfaces, showcasing intent inference,…