neuroscience · slides · talks & lectures

The Brain and the Modern AI: Drastic Differences and Curious Similarities

In this presentation we use Marr’s three levels of analysis to approach the question of how similar or different the brain and the machines are. By approaching this discussion in a more structured manner we find that the answer is different depending on the level of abstraction we are operating at.

Download PDF: Ilya-Kuzovkin-Brain-and-Modern-AI-Differences-Similarities.pdf

Slide 1 of 61: The Brain and the Modern AI Drastic differences and curious similarities Ilya KuzovkinSlide 2 of 61Slide 3 of 61Slide 4 of 61Slide 5 of 61: We need to structure this conversation Suggestion David Marr's Three levels of analysisSlide 6 of 61: We need to structure this conversation Suggestion David Marr's Three levels of analysis Goal of the computation What is the purpose of computation?Slide 7 of 61: We need to structure this conversation Suggestion David Marr's Three levels of analysis Goal of the computation What is the purpose of computation? Algorithm and representation What representations does the system use? What processes are in use to manipulate representations?Slide 8 of 61: We need to structure this conversation Suggestion David Marr's Three levels of analysis Goal of the computation What is the purpose of computation? Algorithm and representation What representations does the system use? What processes are in use to manipulate representations?…Slide 9 of 61: ImplementationSlide 10 of 61: https://medium.com/@jayeshbahire/the-artificial-neural-networks-handbook-part-4-d2087d1f583e Output Activation function Sum Bias Weights Inputs https://www.dreamstime.com/pyramidal-cell-cerebral-cortex-neurons-stained-golgiâ €…Slide 11 of 61: https://medium.com/@jayeshbahire/the-artificial-neural-networks-handbook-part-4-d2087d1f583e Output Activation function Sum Bias Weights Inputs https://www.dreamstime.com/pyramidal-cell-cerebral-cortex-neurons-stained-golgiâ €…Slide 12 of 61: https://medium.com/@jayeshbahire/the-artificial-neural-networks-handbook-part-4-d2087d1f583e Output Activation function Sum Bias Weights Inputs https://www.dreamstime.com/pyramidal-cell-cerebral-cortex-neurons-stained-golgiâ €…Slide 13 of 61: https://medium.com/@jayeshbahire/the-artificial-neural-networks-handbook-part-4-d2087d1f583e Output Activation function Sum Bias Weights Inputs https://www.dreamstime.com/pyramidal-cell-cerebral-cortex-neurons-stained-golgiâ €…Slide 14 of 61: https://medium.com/@jayeshbahire/the-artificial-neural-networks-handbook-part-4-d2087d1f583e Output Activation function Sum Bias Weights Inputs https://www.dreamstime.com/pyramidal-cell-cerebral-cortex-neurons-stained-golgiâ €…Slide 15 of 61: https://medium.com/@jayeshbahire/the-artificial-neural-networks-handbook-part-4-d2087d1f583e Output Activation function Sum Bias Weights Inputs https://www.dreamstime.com/pyramidal-cell-cerebral-cortex-neurons-stained-golgiâ €…Slide 16 of 61: https://medium.com/@jayeshbahire/the-artificial-neural-networks-handbook-part-4-d2087d1f583e Output Activation function Sum Bias Weights Inputs https://www.dreamstime.com/pyramidal-cell-cerebral-cortex-neurons-stained-golgiâ €…Slide 17 of 61: https://medium.com/@jayeshbahire/the-artificial-neural-networks-handbook-part-4-d2087d1f583e Output Activation function Sum Bias Weights Inputs https://www.dreamstime.com/pyramidal-cell-cerebral-cortex-neurons-stained-golgiâ €…Slide 18 of 61: https://medium.com/@jayeshbahire/the-artificial-neural-networks-handbook-part-4-d2087d1f583e Output Activation function Sum Bias Weights Inputs https://www.dreamstime.com/pyramidal-cell-cerebral-cortex-neurons-stained-golgiâ €…Slide 19 of 61: https://medium.com/@jayeshbahire/the-artificial-neural-networks-handbook-part-4-d2087d1f583e Output Activation function Sum Bias Weights Inputs https://www.dreamstime.com/pyramidal-cell-cerebral-cortex-neurons-stained-golgiâ €…Slide 20 of 61: https://medium.com/@jayeshbahire/the-artificial-neural-networks-handbook-part-4-d2087d1f583e Output Activation function Sum Bias Weights Inputs https://www.dreamstime.com/pyramidal-cell-cerebral-cortex-neurons-stained-golgiâ €…Slide 21 of 61: https://medium.com/@jayeshbahire/the-artificial-neural-networks-handbook-part-4-d2087d1f583e Output Activation function Sum Bias Weights Inputs https://www.dreamstime.com/pyramidal-cell-cerebral-cortex-neurons-stained-golgiâ €…Slide 22 of 61: https://medium.com/@jayeshbahire/the-artificial-neural-networks-handbook-part-4-d2087d1f583e Output Activation function Sum Bias Weights Inputs https://www.dreamstime.com/pyramidal-cell-cerebral-cortex-neurons-stained-golgiâ €…Slide 23 of 61: The brain cannot be doing backpropagation (at least not the way we do it) https://mattmazur.com/2015/03/17/a-step-by-step-backpropagation-example/Slide 24 of 61: The brain cannot be doing backpropagation (at least not the way we do it) Forward pass https://mattmazur.com/2015/03/17/a-step-by-step-backpropagation-example/Slide 25 of 61: The brain cannot be doing backpropagation (at least not the way we do it) Forward pass Backwards pass https://mattmazur.com/2015/03/17/a-step-by-step-backpropagation-example/Slide 26 of 61: The brain cannot be doing backpropagation (at least not the way we do it) Forward pass Backwards pass https://mattmazur.com/2015/03/17/a-step-by-step-backpropagation-example/ ✓ ✓Slide 27 of 61: The brain cannot be doing backpropagation (at least not the way we do it) Forward pass Backwards pass https://mattmazur.com/2015/03/17/a-step-by-step-backpropagation-example/ ✓ ✓Slide 28 of 61: The brain cannot be doing backpropagation (at least not the way we do it) Forward pass Backwards pass https://mattmazur.com/2015/03/17/a-step-by-step-backpropagation-example/ Forward pass Neurons send spikes, not real-valued outputs…Slide 29 of 61: The brain cannot be doing backpropagation (at least not the way we do it) Forward pass Backwards pass https://mattmazur.com/2015/03/17/a-step-by-step-backpropagation-example/ Forward pass Neurons send spikes, not real-valued outputs Backwards pass…Slide 30 of 61: The brain cannot be doing backpropagation (at least not the way we do it) Forward pass Backwards pass https://mattmazur.com/2015/03/17/a-step-by-step-backpropagation-example/ Forward pass Neurons send spikes, not real-valued outputs Backwards pass…Slide 31 of 61: Quite different on the level of implementation < ✓ 'ish ✗Slide 32 of 61: Algorithm and representationSlide 33 of 61: Hippocampus and experience replay HippocampusSlide 34 of 61: Hippocampus and experience replay HippocampusSlide 35 of 61: Hippocampus and experience replay HippocampusSlide 36 of 61: Hippocampus and experience replay HippocampusSlide 37 of 61: Hippocampus and experience replay Hippocampus DQN algorithm has state memory buffer and experience replay mechanism: - store experiences while interacting with the environment - randomly sample and replay experiences from memory while learning Replay bufferSlide 38 of 61: Vision: hierarchy of layers Deep Neural Networks Reveal a Gradient in the Complexity of Neural Representations across the Ventral Stream https://www.jneurosci.org/content/35/27/10005 Comparison of deep neural networks to spatio-temporal cortical dynamics of human visual object…Slide 39 of 61: Vision: hierarchy of layers Deep Neural Networks Reveal a Gradient in the Complexity of Neural Representations across the Ventral Stream https://www.jneurosci.org/content/35/27/10005 Comparison of deep neural networks to spatio-temporal cortical dynamics of human visual object…Slide 40 of 61: Vision: hierarchy of layers Deep Neural Networks Reveal a Gradient in the Complexity of Neural Representations across the Ventral Stream https://www.jneurosci.org/content/35/27/10005 Comparison of deep neural networks to spatio-temporal cortical dynamics of human visual object…Slide 41 of 61: Vision: hierarchy of layers Deep Neural Networks Reveal a Gradient in the Complexity of Neural Representations across the Ventral Stream https://www.jneurosci.org/content/35/27/10005 Comparison of deep neural networks to spatio-temporal cortical dynamics of human visual object…Slide 42 of 61: Vision: hierarchy of layers Deep Neural Networks Reveal a Gradient in the Complexity of Neural Representations across the Ventral Stream https://www.jneurosci.org/content/35/27/10005 Comparison of deep neural networks to spatio-temporal cortical dynamics of human visual object…Slide 43 of 61: Vision: hierarchy of layers Deep Neural Networks Reveal a Gradient in the Complexity of Neural Representations across the Ventral Stream https://www.jneurosci.org/content/35/27/10005 Comparison of deep neural networks to spatio-temporal cortical dynamics of human visual object…Slide 44 of 61: Vision: hierarchy of layers Deep Neural Networks Reveal a Gradient in the Complexity of Neural Representations across the Ventral Stream https://www.jneurosci.org/content/35/27/10005 Comparison of deep neural networks to spatio-temporal cortical dynamics of human visual object…Slide 45 of 61: Model of the world: grid cells The Nobel Prize in Physiology or Medicine 2014 "for their discoveries of cells that constitute a positioning system in the brain" http://www.scholarpedia.org/article/Grid_cells O'Keefe Moser Moser https://deepmind.com/blog/article/grid-cells…Slide 46 of 61: Model of the world: grid cells Entorhinal cortex The Nobel Prize in Physiology or Medicine 2014 "for their discoveries of cells that constitute a positioning system in the brain" http://www.scholarpedia.org/article/Grid_cells O'Keefe Moser Moser…Slide 47 of 61: Model of the world: grid cells Entorhinal cortex The Nobel Prize in Physiology or Medicine 2014 "for their discoveries of cells that constitute a positioning system in the brain" http://www.scholarpedia.org/article/Grid_cells O'Keefe Moser Moser…Slide 48 of 61: Model of the world: grid cells Entorhinal cortex The Nobel Prize in Physiology or Medicine 2014 "for their discoveries of cells that constitute a positioning system in the brain" http://www.scholarpedia.org/article/Grid_cells O'Keefe Moser Moser…Slide 49 of 61: Model of the world: grid cells Entorhinal cortex The Nobel Prize in Physiology or Medicine 2014 "for their discoveries of cells that constitute a positioning system in the brain" http://www.scholarpedia.org/article/Grid_cells O'Keefe Moser Moser…Slide 50 of 61: Model of the world: successor representations https://julien-vitay.net/2019/05/successor-representations/ Model-free RL + fast – inflexible to change http://gershmanlab.webfactional.com/pubs/Stachenfeld17.pdfSlide 51 of 61: Model of the world: successor representations https://julien-vitay.net/2019/05/successor-representations/ Model-free RL + fast – inflexible to change Model-based RL + adaptable – hard to learn transition probs reward probs…Slide 52 of 61: Model of the world: successor representations https://julien-vitay.net/2019/05/successor-representations/ Model-free RL + fast – inflexible to change Model-based RL + adaptable – hard to learn transition probs reward probs Successor representation for RL discounted future state…Slide 53 of 61: Model of the world: successor representations https://julien-vitay.net/2019/05/successor-representations/ Model-free RL + fast – inflexible to change Model-based RL + adaptable – hard to learn transition probs reward probs Successor representation for RL discounted future state…Slide 54 of 61: Model of the world: successor representations https://julien-vitay.net/2019/05/successor-representations/ Model-free RL + fast – inflexible to change Model-based RL + adaptable – hard to learn transition probs reward probs Successor representation for RL discounted future state…Slide 55 of 61: Dual DQN network MERLIN Architecture Predictive coding Visual attention ... Generative Adversarial Networks Predictive coding Successor representations Attention in CNNs Consolidation of experience Novel vs. routine circuits & & & &Slide 56 of 61: Replay buffer Curious similarities on the level of algorithm and representation discounted future state occupancy matrix expected state rewardSlide 57 of 61: Goal of computationSlide 58 of 61: Same goal, different strengths... merge! Fetz Lab University of Washington NeuroChip Neuralink Elon Musk San Franciso, CA http://www.csne-erc.org/sites/default/files/Neurochip.pdf https://depts.washington.edu/fetzweb/closed-loop-bci.html https://neuralink.com/Slide 59 of 61: Same goal, different strengths... merge! Fetz Lab University of Washington NeuroChip Neuralink Elon Musk San Franciso, CA http://www.csne-erc.org/sites/default/files/Neurochip.pdf https://depts.washington.edu/fetzweb/closed-loop-bci.html https://neuralink.com/Slide 60 of 61: Same goal, different strengths... merge! Fetz Lab University of Washington NeuroChip Neuralink Elon Musk San Franciso, CA http://www.csne-erc.org/sites/default/files/Neurochip.pdf https://depts.washington.edu/fetzweb/closed-loop-bci.html https://neuralink.com/Slide 61 of 61: Goal of the computation What is the purpose of computation? Algorithm and representation What representations does the system use? What processes are in use to manipulate representations? Implementation How is the system physically realized? The Brain and the Modern AI Almost…
1 / 61

The Brain and the Modern AI: Drastic Differences and Curious Similarities

Presented at DSAi Special Edition: Brain Computer Interfaces and Brain-inspired AI in Sydney