Ep. 1: Language Learning Models: Basics of Neural Networks

This kickoff episode of the 2024 AI Series breaks down how neural networks learn, covering the architecture of these brain-inspired data structures and the layers of mathematics and data processing that simulate the learning process. The session walks through feedforward computation, in which input is transformed step by step into an output that can approximate complex functions, and unpacks backpropagation as the mechanism by which neural networks adjust their digital synapses to improve over time. This talk is presented as part 1 of a 4-part sub-series on language learning models within the broader AI Series.

Key Highlights

• The webinar demystifies neural network architecture, explaining how layers of mathematics and data processing simulate the brain's learning process.
• It details feedforward computation, the step-by-step transformation of input into an output capable of approximating complex functions.
• It explains backpropagation, the method by which neural networks learn from errors through incremental adjustments to their digital synapses.

‍

Ep. 2: Language Learning Models: Understanding Embeddings

This second installment of the AI Series focuses on embeddings, or word vectors, and their role in refining the architecture of language models. The webinar explores semantic mapping, showing how embedding transforms words into numerical entities so machines can capture the subtle nuances of human language, including the semantic and syntactic relationships that give language models their predictive accuracy. It also covers the techniques used to create and refine embeddings for optimal language output.

Key Highlights

• The webinar explains how word embedding transforms words into numerical vectors, allowing machines to interpret semantic and syntactic relationships within language.
• It covers techniques used to create and refine embeddings to optimize language model output.
• It examines how word relationships impact the performance and predictive capabilities of language models.

‍

Ep. 3: Language Learning Models: Overview of the Attention Mechanism

This third installment of the AI Series examines the attention mechanism, the transformative element behind how machines understand and generate human language. The session explores Transformers, models that use matrix inputs to produce contextually relevant outputs for words within a given vocabulary, along with matrix multiplication and vector similarity as tools for quantifying semantic relationships between words. It also breaks down the types of attention, including self-attention, which enables a model to weigh the importance of different words within the same sentence.

Key Highlights

• The webinar reviews matrix multiplication and vector similarity as the mathematical foundations for quantifying semantic relationships between words in language models.
• It provides an in-depth look at the attention mechanism, the driving force behind a model's ability to focus on relevant parts of input data.
• It breaks down types of attention, including self-attention, which allows models to weigh the importance of different words within the same sentence.

‍

Ep. 4: Language Learning Models: The Transformer

This fourth installment closes out the "Language Learning Models" mini-series within the AI Series with a deep dive into the Transformer architecture, the foundation behind today's most advanced language models. The session covers the underlying structure and components that drive Transformer performance, the training processes used to capture the nuances of language, and how Transformers handle inference to make real-time decisions and generate responses. It also explores high-dimensional meaning spaces, which allow Transformers to navigate the complexity of human language with precision.

Key Highlights

• The webinar breaks down the underlying structure and components that give Transformer models their state-of-the-art performance.
• It covers the training processes used to define and refine Transformers so they accurately capture the nuances of language.
• It examines high-dimensional meaning spaces and how Transformers use them to navigate the complexity of human language during inference.

‍

Ep. 5: Bringing Oncology Intelligence to Early Drug Development With AI-Powered Response Predictions

September 9, 2024

Dr. Michael Boice, PhD, Senior Director of Scientific Engagement and Key Accounts at Certis Oncology Solutions, explores how predictive artificial intelligence and machine learning, paired with in vivo validation, are replacing many traditional high throughput screening studies in oncology drug discovery. Dr. Boice, who holds a PhD in Pharmacology from Weill Cornell Graduate School of Medical Sciences and brings over 20 years of experience in translational oncology, shares preclinical and clinical case studies pairing predictive AI with in vivo validation to generate drug response and synergy data. He also demonstrates a new virtual assistant tool that uses natural language processing to help researchers find tumor models matching desired genetic profiles. Sponsored by Certis Oncology Solutions, Inc.

Key Highlights

• Dr. Michael Boice explains how predictive AI and machine learning are replacing many high throughput screening studies with computer-generated insights in oncology drug discovery.
• He presents new preclinical and clinical case studies pairing predictive AI with in vivo validation for more insightful drug response and synergy data.
• Boice demonstrates a virtual assistant tool that uses natural language processing (NLP) to help researchers quickly find tumor models with specific genetic profiles.

‍

‍

‍

Ep. 6: The Role of Advanced De-Identification in AI-Driven Healthcare

September 24, 2024

Dr. Khaled El Emam, PhD, Canada Research Chair in Medical AI and Professor in the School of Epidemiology and Public Health at the University of Ottawa, and Patricia Thaine, Co-Founder and CEO of Private AI, examine how advanced de-identification techniques can protect patient privacy while preserving the data utility needed to train AI models in healthcare. Dr. El Emam, also Director of the Electronic Health Information Laboratory at the CHEO Research Institute, and Thaine, a University of Toronto Computer Science PhD candidate and Vector Institute alumna, discuss how data redaction can accelerate clinical trials, support compliance with privacy regulations, and enable secure data sharing for healthcare research. Sponsored by Private AI.

Key Highlights

• Dr. Khaled El Emam and Patricia Thaine explore advanced redaction techniques that maintain data utility for AI model development in healthcare.
• They discuss how data redaction can accelerate clinical trials and support compliance with privacy regulations governing patient information.
• The presenters outline strategies for secure, privacy-conscious data sharing in healthcare research, balancing innovation with patient confidentiality.