AI summary by TwinMind
How Machine Learning is Fighting Cancer
Real Engineering
13m 34s
Introduction to Machine Learning Ubiquity
- Machine learning pervasive in daily life - dozens of interactions daily without awareness
- Netflix examples:
- Curates movies/shows based on viewing history
- Thumbnail artwork selection based on user preferences
- Comedy watchers see Robin Williams thumbnail for Good Will Hunting
- Rom-com bingers see Matt Damon/Minnie Driver thumbnails
- Uber uses ML for pricing (weather, time, traffic, historic data variables)
- Google Maps traffic warnings and alternate routes
- Facebook targeted ads and Amazon recommendations
Machine Learning Applications Beyond Commerce
- Expanding beyond revenue maximization to benefit humanity
- Applications in disaster relief, climate change monitoring, healthcare
- Recent advances in biological and computer sciences driving cancer research
- Microsoft project aiming to solve cancer using AI
- Large-scale computing initiatives for cancer diagnosis and treatment
Machine Learning Fundamentals
- Definition: Science of getting computers to solve problems without explicit programming
- Two main types:
- Unsupervised: No specified target in data, used for pattern exploration
- Supervised: Specify particular features/traits to learn
- Training set contains input variables and output variable
- Algorithm learns input-output relationships
- Predicts outputs for new input data
Neural Networks Architecture
- Inspired by biological neural networks in brains
- Excellent for finding complex variable relationships humans can't recognize
- Components:
- Input layer
- One or more hidden layers
- Output layer
Cancer Prediction Case Study
- Researchers using neural networks to predict breast cancer likelihood
- Training data includes:
- Patient genetics, demographics, nutrition
- Vitamin intake (folate, B2, B6, B12)
- Age, BMI, pregnancies, pre/postmenopausal status
- Data split: 75% training, 25% testing
- Output: 0 (no cancer) or 1 (cancer development)
Neural Network Training Process
- Input data flows through hidden layers to output layer
- Connections have random weights representing strength between layers
- Bias and activation function (sigmoid) added - squashes output between 0-1
- Initial outputs essentially random, almost guaranteed incorrect
- Backpropagation: Compares actual vs intended output, modifies weights
- Iterative process: run network repeatedly, adjust weights, get closer to correct answers
Gradient Descent Algorithm
- Backbone of neural networks, most used learning algorithm
- Cost function: Measures error between predicted and actual output
- Goal: minimize cost function by finding optimal weights
- Process calculates derivative to determine direction toward local minima
- Learning rate determines step size:
- High rate: covers more ground but risks overshooting
- Low rate: more precise but time-consuming
- Usually determined by trial and error
Training Completion and Testing
- Process continues until error sufficiently small
- Actual and intended outputs coincide
- Test with reserved 25% data
- Network predicts cancer development based on input data
- Can then analyze new patients with unknown outcomes
Research Results and Applications
- Cancer prediction neural network achieved 94.2% accuracy
- Much higher than human prediction capabilities
- Similar accuracy rates in other cancer prediction studies
- Identified key factors: folate deficiency and estrogen exposure
- Healthcare providers can focus on ideal treatment courses
- Address nutrient/hormone deficiencies before cancer develops
Microsoft Cancer Research Initiatives
- Multiple research groups using AI and machine learning
- Oncologist support: Help determine individualized cancer treatments
- Intuitive way to sort through massive research data
- Data overwhelm: Patients' entire genomes creating information overload
- Computer vision integration: Help radiologists understand tumor progression
- Cellular programming: Future possibility to program cells to fight cancer
Industry Adoption and Future Applications
- Technology adoption across industries increasing
- Applications span: cancer diagnosis, Amazon recommendations, autonomous vehicles
- Machine learning as elegant, powerful tool for understanding world
- Significant societal impact potential
- Researchers continuously developing new applications
