slack channel will be on? ![[Pasted image 20230825130716.png]] - one of the settings (chisight - 3d human ; jim, bill freeman, josh, - pairing with psychophysic); doesn't confabulate (not all domains, as accurately as domain) - mind and brain + fundamental limits bf embark on - scaling route is the following: beautiful (microscopic structure - digital logic to trillion of steps with low failure rate - all the up to software ecology / in 30 yrs) - ![[Pasted image 20230825131152.png]] - possibility that basic ingredient to change the history is here(natural - how intelligent agent can scale) - 5yr time scale - early prototype of chiexpertise - open for any feedback (profound risk) - military and drones (no way to improve science wo getting risk) - ![[Pasted image 20230825131511.png]] - at least as important (sci, not tech)![[Pasted image 20230825131823.png]] - how rational are we and where can machine help? - reverse engineering the first 1k-ms of perception (medical intervention - assist ppl understand what they do right or wrong - perceptual rationality, not cognitive rationality) - ![[Pasted image 20230825132113.png]] - goerge matheos - biological - ![[Pasted image 20230825132201.png]] - write ppl in gen, compile to network neuron, biological neurons are clocked - ![[Pasted image 20230825132238.png]] - how is it even possible? "brain-like" real time inference can happen - quantitative, gamma rythin (90hz) vs theta? (4hr) - ![[Pasted image 20230825132435.png]] - images of cortex highlighting its layer (broadly reported - comes in thru 4th layer) - cortex is more like firmware (locally re-program and simulate much / relative to indirec simulation)can not software - ![[Pasted image 20230825132523.png]] - biophysically implementation![[Pasted image 20230825132819.png]] - wet lab facility - to differenciate btw disambiguity - ![[Pasted image 20230825132852.png]] - real prediction correlation (most right: randomly permute)![[Pasted image 20230825132931.png]] - topologically sorting the graphical and assign to layer 4 and canonical (4-> 2,3 -> 5,6) - can follow the loop from ![[Pasted image 20230825133421.png]] - ![[Pasted image 20230825133442.png]] - diagnal is strongest ![[Pasted image 20230825133521.png]] - errors in allen projection atlas - ![[Pasted image 20230825133612.png]] - andrew bolton - injection (lp is too coarsely labeled - reverse loop) ![[Pasted image 20230825133626.png]] - no bug after correcting the confound ![[Pasted image 20230825133654.png]] - poisson process is so fast and simple (independent from what model is) - happy to verify 1. theory (facts from ) - disconnected facts 2. instance of how theory corrected the data - c.e.g. of theory abandon us, data save us estimating the state of the world (not decision; what actions!!! - sequential mc and control) -particle weights go to brain region where dopamin region - multiplication (expected reward) - weights go to exact right place !!! - ways to test theorys on how drugs interfer ![[Pasted image 20230825134106.png]] micro circuits for rolling dice (prob.prog. caught on more) - inference meta programming : automating (restricted - good time series, chiexperties) + ai/cognition complete (Think flexibly); sgd of large network (generic but non-economic) - ![[Pasted image 20230825135308.png]] - better automation for us to (orders of 10s written dsl - church - continuation fit into the - Tom night computer engineer; need at least a few hundreds before worrying about continuation) 1. statistical abstracts (arguments could be supported by data) 2. causes of malnutrition - couldn't even meta-agree whether they agree or disagree (factionalized camps) - micro data with 4milion (pharma, best university; too easy to poke holes) - system causal knowledge and data in the speed of human thoughts - ![[Pasted image 20230825135356.png]] - ![[Pasted image 20230825135936.png]] - ![[Pasted image 20230825140150.png]] - ![[Pasted image 20230825140201.png]] - ![[Pasted image 20230825140211.png]] - ![[Pasted image 20230825140217.png]] - ![[Pasted image 20230825140246.png]] - I only believe in stats that I doctored myself - why is sth interesting (reasoning and sampling bias) - what if this theta (biased by certaink - how biased would the data have to be - till the assumptions get falsified) - there are zones for much more debat - empirical questions they could be asked - implications in clinical research process ![[Pasted image 20230825140857.png]] - ![[Pasted image 20230825140929.png]] - ![[Pasted image 20230825140939.png]] - ![[Pasted image 20230825140948.png]] - ![[Pasted image 20230825140959.png]] - ![[Pasted image 20230825141006.png]] - ![[Pasted image 20230825141021.png]] - automation and scaling (cs approach) - ![[Pasted image 20230825141124.png]] - tradeoff (~ simple, general, accuracy) - freer (mit) , Dan roy (toronto), ackerman, ![[Pasted image 20230825141243.png]] - independent noise + bounded density -> computable (organizim managed to learn from data) - ![[Pasted image 20230825141504.png]] - system level automation ![[Pasted image 20230825141517.png]] - no control on error (only by accident they; don't control prob. of failure - engineering and equity issue when they are used broad) - pretty general domain (Circumscribe - automate generate) - ![[Pasted image 20230825141722.png]] - ![[Pasted image 20230825141908.png]] - NOT know exact posterior but "can sample arbitraily well calibrated" - ![[Pasted image 20230825142032.png]] - if you have prior then bi is slow - backwards (수도 marginal is faster than determinstic than stoch alg) - ![[Pasted image 20230825142208.png]] - ![[Pasted image 20230825142238.png]] - long incubation process (more ) - supporting qualified (founding thesis ) Zoubin Ghahramani’s tutorials @ ICML / NeurIPS ~2004 David MacKay — Ch 28 — Bayesian Occam’s Razor