Building Machines That Learn and Think Like People

the parameters across the edges in the graph represent, is logically unable to run any other program. Lee Sedol is a Universal Machine in the Turing sense, any statement contradicting this makes no mathematical sense. > We limit software to specific domains only on account of efficiency, not because algorithms are fundamentally limited. It is well known within the literature that these models do not make best available use of information when learning. They are exceedingly inefficient in their incorporation of new information. Issues include improper adjustment of learning rates, not using side information to constrain computation, having to experience many rewards before action distributions are adjusted in the case of reinforcement learning, samples per example in supervised learning. Note that animals are able to learn without explicit labels and clear 0/1 losses. Humans and animals generally, even in the supervised regime, are vastly more flexible in the format the supervision can take. For an example, look into the research on how children are able to generalize from ambiguous explanations as "that is a dog" and why difficulty in learning color from this kind of "supervision" shows just what priors are being leveraged to get that kind of learning power. See here for an excellent overview of limitations in our current approaches to AI: https://arxiv.org/pdf/1604.00289v2.pdf > "Learning without forgetting" That's a great paper but it does this by minimizing prediction error drift by comparing before and post performance on the old task while learning the new. I do not know that this method will scale with increasing task numbers, considering Neural Networks are already difficult and energy-time consuming enough to train as is.

scene is parsed in an explicitly interpretable way, with bounding boxes for the various objects on which reasoning will commence. This 'bridge' theme between natural language and the underlying task space is really powerful, and it probably makes sense to figure out how we include them for systems that have nothing to do with natural language. https://arxiv.org/abs/1901.11390 contains another great example of how interpretable such models can be, especially if they are generative. Take a look at those segmentations! Lastly, https://arxiv.org/abs/1604.00289 lays out this vision in a lot more detail.

examination of where this technology came from; Stanley, Thrun’s robotic car that was acquired by Google and is the basis for Waymo’s tech. Thrun notes; > “In the last Grand Challenge, it didn't really matter whether an obstacle was a rock or a bush, because either way you'd just drive around it," says Sebastian Thrun, an associate professor of computer science and electrical engineering. "The current challenge is to move from just sensing the environment to understanding the environment." http://news.stanford.edu/pr/2007/pr-junior-021407.html More precisely, to interact and function in the wild, the robot needs to understand, distinguish, and model the properties of wildly different entities. And these entities aren't just other people - they are everything in its environment. And this includes random events like; people walking into traffic, objects falling onto the road, sudden rain and hail, or just potholes. This is not a problem that can be trivially solved with better sensors or cheaper LIDAR. It is a conceptual problem that can not be trivially brute forced nor simulated in advance as each unique permutation is just different enough to confound existing models. Even if such different and confounding events are six sigma in nature and have a probability of occurring 0.00034% for every mile driven, then at 3.22 trillion miles/year (https://www.npr.org/sections/thetwo-way/2017/02/21/516512439...) Americans will experience 10,948,000 anomalies per year. Or, about 20.8 events per minute. 20.8 anomalies per minute is a pretty big number. That’s a potential accident every 3 seconds or so. And that’s a six sigma estimate. It is hard to emphasize how hard six sigma performance is for the current generation of autonomous vehicle hardware. You might dispute the numbers, but they are meant to be an illustration and an intuitive explanation for why this problem is Hard with a capital H. Unlike some professionals, I don’t believe Strong AI is necessary to solve this problem, but Weak AI is still a form of AI. Technology that's just out of reach today. At this point, people counter with Google’s cars. After all, seeing is believing, and haven’t they been seen to run successfully for millions of miles at this point? Yes, but... Waymo/Google's success is hard to replicate. No one else can bend the rules far enough to achieve their headway. Most of us forget that before starting the project, Google used LIDAR to create highly detailed maps of every environment their cars will ever be in. Engineers then captured standard human driving on that route over and over across different hours. These elements were then combined within their model to pre-compute routes before any real world driving happened. Google was able to achieve its miraculous results because of its ownership of GMaps, Street View and the resources to deploy a fleet to map everything in advance. No one else has these resources. No one else can acquire them cheaply. And even Alphabet, a trillion-ish dollar corporation can't replicate this package for the entire world. Maybe someone will figure out a way to create precise sub-eighth-of-an-inch level 3D models of environments using drones and satellite images, but until that time this technology is too expensive to deploy. It is hard to emphasize how important a distinction this is. The "only" thing a Google/Waymo robot does in real-time is that it checks the difference between the recorded dataset and whatever the real-time LIDAR + camera feed is saying. Or in other words, for every point of the route, Google has generated an expected set of values, and any deviations from these are used by the car to make real-time decisions. https://spectrum.ieee.org/automaton/robotics/artificial-inte... However, even this partial solution is tempting. Because it works. But, sadly, that's not the whole story. What gets lost in the re-telling of this story is that this technique means that the cars can't work in unstructured environments. They are unsafe for most roads, especially uncontrollable, unpredictable and narrow residential roads. It's why these cars exist inside strict geo-fences even when they're ferrying passangers around, like in Phoenix. Where the car seems to mostly serve the downtown area with very little residential coverage (caveats apply, I couldn't find a map). But, even getting this far is a huge technical achievement, even if it's not the fundamental breakthrough we've been looking for. We should be prudent, but optimistic. There are a lot of smart people out there who are working on this problem right now. They are figuring out really clever ways to deal with these edge cases and they will solve this problem. Even if it's not on the hype cycle's schedule. With that in mind, I'd like to suggest a better model for thinking about Driverless Vehicles is another Google service, Translate. It's quite good, but when was the last time someone translated an entire book with it (and it came out to be readable/sensible)? The system's performance - due to its statistical roots - seems to be fundamentally asymptotic w.r.t. the data presented to it. It can't be improved beyond a certain point even if you throw more data at it. I suspect that the current generation of driverless vehicles will be the same. http://people.csail.mit.edu/brooks/papers/elephants.pdf http://arxiv.org/pdf/1604.00289.pdf Before the needed breakthrough happens, it is likely that the cutting edge won't advance beyond Level 3 in the next decade; > I would guess Tesla’s position on this would be that most of the time, yes, you can rely on it, but because Tesla has no idea when you won’t be able to rely on it, you can’t really rely on it. https://spectrum.ieee.org/cars-that-think/transportation/sel... And that's okay. We might not have driverless cars by 2020, but we probably will by the mid 2030's and that sounds pretty darn fantastic to me.

have here works pretty well". But it also seems to be that those edge cases are where human cognition truly shines. (ex. naturalized categorization, metaphor in language, novel problem solving) There is some progress, but at this stage its unclear how to import the philosophical discussions to an actual theory. I'll add that it's very easy to be an "amateur cognitive scientist" because you just need some self-awareness to be able to construct some folk theory of the mind that works most of the time. But a model that "kinda works" isn't going to cut it for AGI because eventually as you push it down the line further your misalignment will get worse and worse. AGI would probably need some massive conceptual breakthrough in cognitive science and not in computational methods (though they are obviously interrelated) cf: https://www.youtube.com/watch?v=EUjc1WuyPT8 (tbh I haven't read the bottom two, but I personally never thought that modern NN methods can achieve AGI, so I didn't bother to read their rationale asides from a skim) https://srconstantin.wordpress.com/2017/02/21/strong-ai-isnt... https://arxiv.org/abs/1604.00289 I am: a cognitive science student that has /some/ understanding in recent ML methods P.S. Still, my dream to be an AGI researcher! I have 0 prior experience but if someone out there is reading is also doing AGI research or something similar, contact me! I have absolutely no official credentials I must add however.

human-like learning and thinking machines will have to reach beyond current engineering trends in both what they learn, and how they learn it. Specifically, we argue that these machines should (a) build causal models of the world that support explanation and understanding, rather than merely solving pattern recognition problems; [...] https://arxiv.org/pdf/1604.00289.pdf Some others papers related to these ideas: - Learning a Theory of Causality http://stanford.edu/~ngoodman/papers/LTBC_psychreview_final.... - Theory-Based Causal Induction: http://cocosci.berkeley.edu/tom/papers/tbci.pdf -