Please enable JavaScript.
Coggle requires JavaScript to display documents.
AI (Deep Learning, Limitations, Future of LLM, Language, Learning, Models,…
AI
Deep Learning
Inspiration from brain
Artificial neural networks
Computational models
Interconnected nodes
Process
Key components
Neurons
Connections (synapses)
Layers
Activation functions
Learning Process
Backpropogation
Training
Calculate errors
Adjust connection strength
Advantages
High capacity
Automatic feature extraction
Improved performance
LLM
Deep Learning model
Architecture
Convolutional Neural Networks for image recognition
Recurrent Neural Networks for language processing
Limitations
Factual accuracy
Reasoning and common sense
Bias and fairness
Nonsensical outputs
Limited understanding of context
Future of LLM
Moving Beyond Statistics
Reasoning and Logic Integration
Transparent Decision-Making
Human-like Explanation Capabilities
Context-Aware Learning
Language
How
Communication
Thoughts and reasoning
Social interaction
Creativity and expression
Generation
Planning and ideation
Forming sentences
Considering context
Foundation
Symbolic system
Arbitrary relationships
Key features
Grammar
Syntax
Semantics
Learning
Core
Acquisition of skills
Experience and Practice
Changes in the brain
Theories
Behaviorism
Cognitive psychology
Contructivism
Human Language learning
Early exposure
Statistical learning
Imitation and Practice
Feedback and correction
Inherent ability vs Learned environment
Innate Ability
Environmental Factors
Models
Types of models
Physical
Conceptual
Mathematical
Computational
Purpose
Explanation
Prediction
Design and optimization
Communication
How models represent reality
Selection of Variables
Relationships between variables
Trade off between complexity and accuracy
Algorithm
Back propagation
Forward pass
Output generation
Error Calculation
Error distribution
Weight adjustments
Gradient Descent
Visualization error landscape
Calculating the gradient
Taking steps downward
The training loop
The LLM processes a piece of training data.
Backpropagation calculates the error between the output and the desired output.
Gradient descent uses the error to adjust the weights of the network.
Steps 1-3 are repeated for numerous data points over many iterations.
Black Box vs. Interpretability
Complex LLMs with billions of parameters
Concerns
Trustworthiness
Debugging
Explainability
Approaches to interpretability
Local Interpretable Model-Agnostic Explanations (LIME)
SHapley Additive exPlanations (SHAP)
Attention Mechanisms
Training Data
Role
Analyze vast amount of text
Identify patterns
Impact
Volume
Quality
Bias
Inclusive Training Data
Fairness Metrics and Benchmarks
Variety
Statistical Learning
Statistical Patterns and relationships
Probabilities
Probabilistic Output
Multiple ways to interpret
multiple potential continuations