Moving Beyond Probabilities: Memory as World Modelling
42How do humans think and reason, is it by probability or by matching to memory?
How can we use memory to do world modelling?
I posit that memory is the fast mechanism of learning, as backpropagation via updating weights (aka Long Term Potentiation / Depression) may be a slower process than encoding and retrieving from memory.
Future systems which incorporate memory, reflections of basic chunks of memory to form meta-memory, adaptation of existing memory to new situations, will likely be faster and adaptable than most of deep learning right now.
~~~~~~~~~~~~
Slides: https://github.com/tanchongmin/TensorFlow-Implementations/blob/main/Paper_Reviews/Moving%20Beyond%20Probabilities%2C%20Memory%20as%20World%20Modelling.pdf
Learning, Fast and Slow: https://www.youtube.com/watch?v=Hr9zW7Usb7I
Learning, Fast and Slow (arxiv) - Shortened version of this paper accepted at International Conference on Development and Learning (ICDL): https://arxiv.org/abs/2301.13758
My related past videos on memory:
LLMs and Knowledge Graphs: https://www.youtube.com/watch?v=1RZ5yIyz31c
Memory Encoding and Retrieval: https://www.youtube.com/watch?v=Lpl1zleA4ws
New Framework of Memory for Learning: https://www.youtube.com/watch?v=q9uMEAcB3lM
Related videos based on this video:
Voyager (125 episodes to diamond pickaxe): https://www.youtube.com/watch?v=Y-pgbjTlYgk
Generative Agents Video: https://www.youtube.com/watch?v=_pkktFIcZRo
Related papers based on this video:
Uncertainty Planning for Self-Driving Cars: https://arxiv.org/pdf/2207.00788.pdf
Daniel Kahneman's work:
https://pages.ucsd.edu/~mckenzie/TverskyKahneman1983PsychRev.pdf
https://apps.dtic.mil/sti/tr/pdf/ADA099502.pdf
Attention is all you need (Transformer paper): https://arxiv.org/abs/1706.03762
Memorizing Transformers: https://arxiv.org/abs/2203.08913
Generative Agents Paper: https://arxiv.org/abs/2304.03442
World Model by Probabilities (Dreamer v3): https://arxiv.org/abs/2301.04104
~~~~~~~~~~~~
0:00 Introduction
2:40 Self-Driving Cars
12:06 What’s wrong with probabilities
20:23 Live Psychological Experiments using Daniel Kahneman’s work
38:14 How to interpret Daniel Kahneman’s Results in a Modern Lens
44:40 Transformers: Context as memory
50:07 Retrieval Augmented Generation to boost context
52:22 Imbuing Memory into Transformers
55:58 Hierarchical Memory Referencing
1:00:40 Light Bulb Experiment
1:07:11 Knowledge Graphs and Causal Linkages
1:11:20 Reflections to Consolidate Memory in Generative Agents
1:16:40 Learning, Fast and Slow: Memory as a way to approximate probabilities
1:27:21 Discussion
~~~~~~~~~~~~~~
AI and ML enthusiast. Likes to think about the essences behind breakthroughs of AI and explain it in a simple and relatable way. Also, I am an avid game creator.
Discord: https://discord.gg/bzp87AHJy5
LinkedIn: https://www.linkedin.com/in/chong-min-tan-94652288/
Online AI blog: https://delvingintotech.wordpress.com/
Twitter: https://twitter.com/johntanchongmin
Try out my games here: https://simmer.io/@chongmin