How do we learn so fast? Towards a biologically plausible model for one-shot learning.
16M Ganesh Kumar shares about his work on modelling one-shot learning in a biologically plausible way, primarily focused on a neuroscience angle. He uses hebbian learning to change synaptic weights, reinforcement learning using Actor and Critic, as well as planning with working memory!
Abstract:
One-shot learning is the ability to learn to solve a problem after a single trial, a feat achievable by algorithms and animals. However, how the brain might perform one-shot learning remains poorly understood as most algorithms are not biologically plausible. After gradually learning multiple cue-location paired associations (PA), rodents learned new PAs after a single trial (Tse et al., 2007), demonstrating one-shot learning. We demonstrate reinforcement learning agents that learn multiple PAs but fail to demonstrate one-shot learning of new PAs. We introduce three biologically plausible knowledge structures or schemas to the agent, 1) the ability to learn a metric representation of the environment 2) the ability to form associations between each cue and its goal location after one trial and 3) the ability to compute the direction to arbitrary goals from current location. After gradual learning, agents learned multiple new PAs after a single trial, replicating the rodent results.
Speaker Information: https://mgkumar138.github.io/
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
References:
Slides: https://github.com/tanchongmin/TensorFlow-Implementations/blob/main/Paper_Reviews/One-shot%20Learning%20(Ganesh).pdf
A nonlinear hidden layer enables actor–critic agents to learn multiple paired association navigation - https://academic.oup.com/cercor/article-abstract/32/18/3917/6509014
One-shot learning of paired association navigation with biologically plausible schemas - https://arxiv.org/abs/2106.03580
My work on Learning, Fast and Slow: https://www.youtube.com/watch?app=desktop&v=Hr9zW7Usb7I
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Other Reading Materials suggested by Ganesh:
Random synaptic feedback weights support error backpropagation for deep learning - https://www.nature.com/articles/ncomms13276
Local online learning in recurrent networks with random feedback - https://elifesciences.org/articles/43299
Prefrontal cortex as a meta-reinforcement learning system - https://www.nature.com/articles/s41593-018-0147-8
Task representations in neural networks trained to perform many cognitive tasks - https://www.nature.com/articles/s41593-018-0310-2
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
0:00 Introduction
2:50 Non-linear representations necessary for one-shot learning
22:43 One-shot learning of new Paired Associations
27:11 Reward modulated Exploratory Hebbian rule
34:27 Hopfield networks do not do one-shot association
37:03 Deleting memories is important for learning
39:12 Goal-directed systems
41:21 Composing three agents into one!
48:52 Adding gates can help minimize distractions
53:00 Question on Forgetting Mechanism
56:24 What if useful signals come after distractors?
1:00:29 Can we learn reward-free?
1:07:16 DetermiNet: Large-Scale Dataset for Complex Visually-Grounded Referencing using Determiners
1:13:51 Summary
1:20:36 Discussion
1:24:20 Towards Biologically-Plausible Neurosymbolic Architectures
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
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
Other Videos By John Tan Chong Min
Other Statistics
Changes Statistics For John Tan Chong Min
There are 366 views in 1 video for Changes. About an hours worth of Changes videos were uploaded to his channel, less than 0.51% of the total video content that John Tan Chong Min has uploaded to YouTube.