My view is throw it all away and start again. —Hinton[x]
⭐ 😮 ❗ 💡 😎 ❤ hi all. blogged about a curiosity-based AGI theory in early 2018 after being inspired by recent Deepmind Go advances. the title was “novelty detection/ seeking”. the short word for that is curiosity. “curiously” the word “curious” didnt appear in the essay a single time. thinking back, suspect my thinking was that maybe the term was too advanced/ bold at the time. its a relatively abstract concept not even fully understood in neurobiology or psychology. it also crosses the species boundaries to other animals besides humans ie a general biological consideration also.
but now its time to use that word. the machine learning field has discovered curiosity in a big way… both locally and more globally.
the futures already here, its just not evenly distributed. —Gibson
the field of ML is very vast, grown rapidly in the last few years (esp wrt deep learning), and its not easy to keep track of these days. its something of a minor obsession for me and track it daily and over several years in this blog, almost since the beginning, via hundreds of links a year. did not run into some key references myself, and that shows how broad the field is and how hard it is even for dedicated/ committed individual researchers to keep up. but theres another element, curiosity was, and to some degree still is, “flying below the radar”.
this blog is timed based on some renewed attention and traction. the researchers in the field are definitely starting to notice something. they are still scattered but theyre the leaders, the pioneers, and suspect a mass herd shift is in the near future/ horizon (say within a few years) just like what happened with ML/ deep learning explosion/ wave itself in the last few years.
so in brief, thats exciting! the Curiosity Paradigm of Intelligence is gaining unmistakable signs of traction. this blog tracks some of those new milestones. yes its been studied for decades and by others, but the core/ nearly radical new theory/ proposal here (not entirely espoused by the following researchers, but aligned/ close) is that curiosity is necessary and sufficient for intelligence.
have been advocating/ promoting/ proseletyzing it myself in cyberspace heavily and got a lot of views on my blog. can be sure it is influencing some. also promote it on reddit and for that, earned some serious resistance there, some battle scars in cyberspace (oops/ yikes “promote” is nearly a 4-letter word in certain quarters of cyberspace that supposedly uphold/ glorify/ exist/ have entire business models based on user-generated-content!). but also a lot of excellent/ positive/ pointed feedback from redditors. thx for that guys!
the first key breakthru is Deepak Pathaks work. he did a talk in May 2018. hes working on his Phd… (someone get this man his Phd asap! think this guy is going to have a big career!) found his mid 2017 work after writing 2 major blogs myself, the 2nd meant to be a very comprehensive survey (and it was!). am a bit embarrassed to have not found it sooner, in retrospect a glaring oversight. somehow my idiosyncratic search terms (maybe not incl “curiosity”) didnt bring it up showing how critical individual terms are. it was widely publicized in 2017 in popular science circles eg Wired, MIT Tech Review (my first intro), Quanta etc (even engadget, WSJ, new scientist, economist, digital trends, and NYPost! how could this be missed by me, astonishing…) but there wasnt much reaction by other researchers it seems. he had a very good PR dept but it seems other scientists were not really in on the breakthru.
Pathak has gone on to do further breakthru work just announced at the end of october on a large scale study across many video games by open AI. as us insiders know, and cited in this blog (but not mentioned directly) Montezumas Revenge has been a real challenge/ holdout game for other MLs, recently “shown”/ thought beyond immediate capability. this new study shows superior-human performance by an ML algorithm that doesnt have access to scores (along with all the other video games). the world has not recognized the massive significance of that feat…
this is extraordinary, ML algorithms that solve video games without access to scores are a very new area of research. my memory was that in early 2017, there were almost no such cases reported in the literature, and hence one of my proposed challenges in the area in the earlier blog (#3 item). much to my surprise the basic idea had mostly already been demonstrated in mid 2017 by Pathak, and has also been further confirmed/ strongly-to-even-vastly extended this year. it looks like with a bold sweep Pathak et al have already proven that “scoreless” or “score blind” ML algorithms can possibly “solve” virtually all video games. this suggests a deep principle in play in not only learning but video game design. popular games are well liked maybe because they trigger this curiosity instinct in humans, if you ask me. ie they are directly triggering this deep brain instinct that crosses to other animals! and now the AI research juggernaut is onto the scent. fantastic!
more recently after writing reddit comments on these breakthrus, have been tipped off on Juergen Schmidhubers work by redditors. Schmidhuber is one of the (somewhat unheralded it seems) oldtimers whos been working on key ideas eg in neural networks long thru the coldest stretches of the AI winter. he has been working on it since 1990 in published work! it does not seem to be cited much elsewhere, and honestly it was not on my radar either after very deep searching. should not feel too bad because his wikipedia entry current cites none of it! his papers are highly relevant but current listed citations seem to end in 2014. wondering, whats the status of this work? is it still ongoing? is he or his cohorts still working on it?
Lillicrap/ Deepmind + Savinov/ Google Brain
in a new/ very recent breakthru, Deepmind + Google Brain researchers Lillicrap and Savinov are now interested in curiosity based learning.[20,21,22] they seem not as far along with some maze-exploring + other additional tasks ideas, but think its all definitely steps in the right directions. some of Lillicraps other new work on “transporting value” seems to me to fit into the curiosity paradigm by looking at familiar features based on memory and using it to drive rewards/ motivation. “rewards” are a very big topic in AI and current have somewhat mixed feelings on coupling the idea with curiosity because of maybe some misleading aspects (conventional rewards seem to be highly tied with external cues whereas the critical curiosity paradigm as outlined by me is all about measuring internal knowledge and its growth), but oh well it works for now (maybe serving as a provisional bridge).
at times it might seem the ML community is large, but on closer look, actually to the contrary there are only a very few researchers in the world working on ideas that will lead directly to AGI. they are working mostly at the big companies, DeepMind, Open AI, Google brain, and at this point all 3 are now looking into curiosity-based intelligence, and theres billions of dollars, strong competition, and some of the worlds greatest minds riding on getting new breakthru results. so its not yet a movement in the larger sense but its getting closer.
“The [Big] Powers That Be” dont yet quite yet have the target directly/ exactly in their laser sights, but theyve now at least peripherally “scanned it”. the earliest seeds/ flickers/ glimmers are here, they are now discernable/ identifiable, they have now arrived! lets see how long before the mass stampede happens/ triggers! dont forget you heard it here 1st!
games + AI, university depts, etc
elsewhere, as usual am tracking lots of AI developments and just pounded all my links into form (stay tuned for that), but heres 3 somewhat-related sections that are really buzzing and want to include in this immediate/ timely/ flash blog. all this research ties in heavily into games + AI which is really thriving esp due to OpenAIs focus on it.[a2] recently Dota has (“largely/ mostly”) fallen to Deepmind/ Open AI research and it made big headlines in 2018. it does look like these team-oriented games require substantially more complex attacks/ architectures/ capabilities than Googles historic/ decisive/ stunning (but now ~2yr old!) Go success, although havent seen a comparison along those exact lines, on edge of my seat to see that analysis!
other extraordinary news is that the deep learning and data science revolutions are starting to impact normally conservative universities announcing very large new departments/ entire massive buildings/ facilities at MIT, UCSB, Carnegie Mellon, Berkeley, Peking.[a3] (at this point china is clearly really ahead of the game worldwide wrt A(G)I prioritization/ funding/ national initiatives, have covered that in many blogs.) theres a lot of activity in Canada also. $1B at MIT, staggering!
last in [a4] am tracking some of my favorite thought leaders/ writers in the area with some really great/ thoughtful essays, Brooks, Piekniewski, Sutskever, and dissident Marcus.
- 1. Reinforcement Learning with Prediction-Based Rewards
- 2. Large-Scale Study of Curiosity-Driven Learning
- 3. Curious AI learns by exploring game worlds and making mistakes | New Scientist
- 4. A sense of curiosity is helpful for artificial intelligence – Machine learning
- 5. Lifelong Learning via Curiosity and Self-supervision – The Robotics Institute Carnegie Mellon University
- 6. Computer Scientists Have Developed An AI That Gets Curious About Its Surroundings | Digital Trends
- 7. WHAT’S INTERESTING? ACTIVE EXPLORATION – CURIOSITY-DRIVEN REINFORCEMENT LEARNING – ARTIFICIAL CURIOSITY – INTRINSIC REWARD – INTRINSIC MOTIVATION – ARTIFICIAL CREATIVITY – LIFELONG DATA COMPRESSION – FORMAL THEORY OF SURPRISE – FORMAL THEORY OF NOVELTY – FORMAL THEORY OF INTERESTINGNESS – FORMAL THEORY OF CURIOSITY – FORMAL THEORY OF CREATIVITY
- 8. Why humans learn faster than AI—for now – MIT Technology Review
- 9. Curiosity May Be Vital for Truly Smart AI – MIT Technology Review
- 10. Super Curious Mario: Teaching AI to Keep Asking Questions | California Magazine
- 11. UC Berkeley researchers teach computers to be curious
- 12. Clever Machines Learn How to Be Curious (And Play Super Mario Bros.) | WIRED
- 13. Clever Machines Learn How to Be Curious | Quanta Magazine
- 14. AI learned how to be curious from playing Super Mario Bros.
- 15. Curiosity Is a New Power in Artificial Intelligence – WSJ
- 16. Curiosity-driven Exploration by Self-supervised Prediction
- 17. Curiosity Driven Exploration by Self-Supervised Prediction – YouTube
- 18. On “solving” Montezuma’s Revenge – Arthur Juliani – Medium
- 19. Large-Scale Study of Curiosity-Driven Learning
- 20. A new method to instill curiosity in reinforcement learning agents
- 21. ICLR 2019 submission
- 22. [1810.02274] Episodic Curiosity through Reachability
- 23. [1810.06721v1] Optimizing Agent Behavior over Long Time Scales by Transporting Value
- 1. AI bots trained for 180 years a day to beat humans at Dota 2 – The Verge
- 2. AI Is Dreaming Up New Kinds of Video Games – MIT Technology Review
- 3. AI trained on 3500 years of games finally beats humans at Dota 2 | New Scientist
- 4. Why Can a Machine Beat Mario but not Pokemon? – Towards Data Science
- 5. A team of AI algorithms just crushed humans in a complex computer game – MIT Technology Review
- 6. Bill Gates hails ‘huge milestone’ for AI as bots beat humans at Dota 2 – Business Insider
- 7. DeepMind’s AI agents exceed ‘human-level’ gameplay in Quake III – The Verge
- 8. Evolutionary algorithm outperforms deep-learning machines at video games – MIT Technology Review
- 9. OpenAI bots played 180 years of Dota 2 a day to beat human players | PC Games Insider
- 10. OpenAI bots thrash team of Dota 2 semi-pros, set eyes on mega-tourney • The Register
- 11. The End of Open AI Competitions – Towards Data Science
- 1. Machine Learning, for Science | The UCSB Current
- 2. Carnegie Mellon’s Andrew Moore to join Google Cloud as new head of AI later this year – GeekWire
- 3. Berkeley inaugurates Division of Data Science and Information, connecting teaching and research from all corners of campus | Berkeley News
- 4. MIT Is Opening a $1Bn AI College – SyncedReview – Medium
- 5. MIT reshapes itself to shape the future | MIT News
- 6. Peking University to Open New Campus With AI Focus – SyncedReview – Medium
- 7. African Master’s In Machine Intelligence – Facebook Research
- 1. [FoR&AI] Steps Toward Super Intelligence I, How We Got Here – Rodney Brooks
- 2. From Here to Human-Level Artificial General Intelligence in Four (Not All That) Simple Steps
- 3. Rebooting AI – postulates – Piekniewski’s blog
- 4. Who will figure out intelligence? – Piekniewski’s blog
- 5. [FoR&AI] The Seven Deadly Sins of Predicting the Future of AI – Rodney Brooks
- 6. OpenAI Founder: Short-Term AGI Is a Serious Possibility
- 7. Predictions of AGI Takeoff Speed vs. Years Worked in Commercial Software – Essays on Reducing Suffering
- 8. DeepMind’s founder says to build better computer brains, we need to look at our own – The Verge
- 9. How to Teach Artificial Intelligence Some Common Sense | WIRED