AI - Deep Learning Flashcards
Deep learning
- A form of machine learning loosely based on brain principles
-The computer is trained by learning from examples
-Superior in seeing patterns in large data sets - Currently, by far the most successful machine learning technology
Brain inspired computing - DL
2006 revolution in artificial neural networks:
- More data for training
-Faster computers which can handle larger neural networks
-Currently dominating AI technology
Geoffrey Hinton being an AI doomer
Left google to express concerns
- Alarmed after googles LLM could explain why a joke was funny
-Surprised by how quick LLMs increased in intelligence without requiring more complex brain like concepts
Brain inspired computing and visual processing
Discusses how the human brain processes visual information.
Explains sensory input from the retina, feature detection in the thalamus, and visual cortex processing.
Highlights findings from Quiroga et al. (2005), such as “Jennifer Aniston neurons,” demonstrating invariant visual representations by single neurons.
ANNs
Details the connection between sensory input, hidden layers, and outputs in ANNs.
Hidden layers act as feature detectors, recognizing increasingly complex patterns across layers.
Introduces backpropagation for training neural networks.
Deep learning requirements
Emphasizes the need for large datasets, powerful computing, and multi-layered ANNs for successful deep learning.
Mentions examples of deep learning applications like generative AI and links to examples of creative outputs.
Generative AI
Highlights AI’s capability to create unique and visually appealing images.
Discusses how AI starts with random noise and refines it based on prompts using neural network models.
Provides examples such as AI-generated artwork and deep fakes.
Neuromorphic chips
Machine learning tasks require special chip designs
-Neuromorphic chips: IBM wants to take the next step and build chips with neural network architecture
-Real time object classification and tracking
IBM brain chip
Hardware implementation of neural networks will make it much more power efficient and scalable to small devices e.g wearables
Google deepmind
Machine learning and systems neuroscience company with breakthroughs in AI.
- their goal is to create general purpose AI, not handcrafted for a specific task but applicable to a large range of different problems
- Motto = first solve intelligence, then everything else
Breakthroughs by google deepmind
- 2015 General AI for playing Atari games, beating human players
- 2016 AlphaGO beats the human champion in the game of GO. GO is such a complex board game that it can not be solved by brute force and is thought to require intuition to win the game.
- 2017 AlphaZero beats Stockfish (strongest chess program) in Chess after training from scratch for four hours, with only the chess rules as input.
- 2020 AlphaFold cracked the protein folding problem, predicting 3D structure of proteins based on their amino acid sequence
Generative AI with deep learning
-Input pixels on the screen and score
-output joystick commands
-Task to optimize the score
LLMs
Large neural networks trained to predict the next word in a sentence based on the previous input
>100 Billion parameters
