David's presentation, teasing three things apart 20:41; video data compression & prediction 21:20; coding efficiency in video and optic system 23:30; used predictive coding in the brain (retina), efficient transmission 25:00; predict forward in time, delays and tennis example (sensory processing or motor is ahead of time) 26:00; silent backwards movie 27:18, forward prediction; inference, top-down + bottom-up, bayesian process & generative 28:30; all three scenarios all happen, sensory processing 31:00; Theory of cortical function integrated all three 33:38; summary "predictive docing - no time representation" "predictive processing - avoid explain away", so not hierarchical, but seperate 34:49

Michael's presentation philosophical perspective, how people receive 52:49; Bayesian - sensory to perception how? Helmholtz unconscious inference 53:30; Bayesian's success - a range of object tracking illusion - unifying theory 54:40; how brain implements bayesian? approximately 55:50; predictive coding is one strand, others might be sampling distribution which matches posterior 56:50; bayesian does not restrain by predictive coding or any structure 58:06; behavioral evidence side by side, bayesian is more general and satisfying 58:41; evidence binocular rivalry defeats predictive coding, but paper does not offer model OR using sampling explanation can work as well, with model 1:01:25; summary - predictive coding might be implementation of generic Bayesian inference, but minimization of PE is not the principle (too soon) 1:06:23

Andy's presentation 5:00; HPC 7:35, explain away prediction error 8:11; action changes the world to fit prediction 9:32; PE generative model 10:50; pyramidal cells 11:30; precision weighted formula 12:40; representation unit does not activate new signals 13:35; precision weighed reliability, etc., suppression and ENHANCEMENT, dopamine's job 15:20; Different properties of prediction and PE 16:50; defect reasons 17:50; summary 18:10; Caution 19:25

@raresmircea

2 жыл бұрын

Great man, thank you

Audience question. Learning, not inference, is the feedback connection? Prediction over time error is signal for learning - not backwards, but temporal difference reinforcement learning "target propogation" 1:19:55; contextual state is PC, bayesian does not have action, PC is the missing part? b1, ACC top-down projections inhibition map, turn on and off precise connectivity 1:22:26; claims of priority of prediction or sensory? T loss brain prediction machine principle - optional, surviving? generative model is centered around prediction and info is the additive, and choosing the next action is the principle and minimize PE -> update model 1:26:00; perceptual system is estimation based on bayesian norm, desire or cost function? stomach metaphor 1:30:30; prefrontal and frontal cortex and bottom up and top down, no hard hold, too many layers, whole brain? models are applied to the whole brain 1:33:00; the form of feedback is only one type of feedback (executive control, cognitive control, modulatory), two clusters vs. 1 group of neuron that do the two jobs (error and prediction) 1:35:46; simplest organism PC occurs? birds and mammals (not plants maybe, b/c free energy rules), efficient transfermation also must exist to make use of limited resources 1:38:00; representation unit but prediction is just the prediction, 1) PE, 2) prediction (top down), 3) representation (bottom up & top down)? representation unit holds generative model 1:40:00; PC can explain imagination, memory, etc. what are the differences between cognitive functions? strength of PC is the overlap between the functions , seperate noise from signal. How to explain consciousness? no answer. causal explanation of why some characteristics will be identified as different things & the stimuli are different 1:42:00; what is the PE for imagination, dream? there is no PE, no state to state transtion. the generative model is the driving force of transition 1:46:30; priority heavy lifting? All initially bottom-up to then top-down recursive? Why is priority subject dependent, not intrinsic in the environment? you always deal with prediction, through evolution, when PE signals changes arousal state when no prediction works. turning the knob 1:48:16; where does the model come from? implicitly a deep net and weights and weighted sum, it is a model when training is done. That model can be run backwards and forwards, adversarial network. turning the knob for a more bayesian or PC approach depending on the situation 1:52:45; Why surprise seeking, not model building? artificial curiosity, improve the model over a long term, in the short term by exploring 1:54:46; Human being different from animal PE b/c time? if moving in space, the spacial info can be predicted statistically, low level. Predicting light intensity with relations between pixels. So at low basic sensory level 1:56:30. Conclude 1:59:25

Lucia's presentation info processing frame 36:50; hierarchical algorithmic, computational, implementation & direction 37:40; recap Andy, post synaptic increased gain - attention 38:40; evidence in single units with amplitude&confidence - rats learning visual pattern 39:50; evidence in humans auditory 43:00; evidence of hypothesis in higher area in ML faces orientation signals die away - monkeys 44:12; ML as AL, higher level does not care about orientation 47:10; predictions from higher areas - monkey faces 48:07; only feedback & error together has signal, not other places with only feedback 49:40; evidence in cortical layers - human superficial layer is where PE is encoded 50:00; summary - unknown but there is evidence, but different systems of function? other bayesian inference can explain? 50:52

This is incredible, this format of debates is so good as a way to really understand the topic deeper

All defeating argument: "Friston thinks.."

@MrRunebro

3 жыл бұрын

Heeger looked so disappointed when they Andy Clark said "Friston and I think your model is just hierarchical predictive coding."

Thank you

Great video. The problem for me was that it did not seem that Heeger and Rescorla understood predictive coding in depth.

Finish presentations. Deep convolutional neural nets, but no explaining away, so is it a key property 1:08:22; what do generative models do? Generate percept with sensory all the way to the back, and sending error, and being asymmetric (wiring can be flipped) and yes Friston thinks the same LOL 1:10:02; falsifiable predictive coding with 3 levels, instead of a model that is super generic no implementation, comparing models 1:12:00; PE does not say much, maybe negotiation and reinforce rather than "explain away" as it is thrown around under different context 1:13:20; influence of higher to lower and asymmetry is the core to processing, PE, feedback model did not show enhanced activity or less selective 1:15:29; pull apart explain away from adaptation, and two predictions from the paper (b/c no precision weighting that decides top or bottom dominance) 1:18:36

Fuck that intro strawberry demonstration blew my mind. No red pixels yet couldn't unsee red strawberries

min 14:06 how long should the decay period be set for?

Great video and talks. Really helps to understand where the discussion on prediction and perception stands. Thanks! Also this picture: NYC end of 2018. Scientists. Climate change, pollution, and excessive waste production is a massive global problem. 4 renown scientists drinking water from small plastic bottles. Organizers can do better. Are there glasses and pitchers on NYU?