Cybiote

As you point out, that the relative ordering remained stable means progress being made is not simply overfitting to the test set. The negative finding from the paper you link to is in how much accuracy drops considering how slight modifications to the test set were. In their own words:

> We view this gap as the result of a small distribution shift between the original CIFAR-10 dataset and our new test set. The fact that this gap is large, affects all models, and occurs despite our efforts to replicate the CIFAR-10 creation process is concerning.

> Nevertheless, the accuracy of all models drops by 4 - 15% and the relative increase in error rates is up to 3×. This indicates that current CIFAR-10 classifiers have difficulty generalizing to natural variations in image data.

It remains to be seen if others can think up ways to maintain performance that the authors did not manage.

The phrasing of pendulum swinging in relation to symbolic was also used by Micheal Jordan here (search for pendulum): https://medium.com/@mijordan3/artificial-intelligence-the-re...

It's clear that AIs based solely on FOL are unlikely, but it's also likely to be true that any system needing to solve problems whose exact solution can't be found in steps polynomial in inputs, will require ideas that are similar to the core of the older approaches. There are problems where wide-ranging search can't be avoided.

Other nice perks of the hybrid approach are data efficiency, compact specification, easier composition, the ability to grow or alter your representation and generate new inferences on the fly (I mean inferences as a result of learning or conditioning on new information and not what people use when they mean prediction). If you learn a new fact, you can go back and explicitly work out the consequences on all the other facts and the inferences generated from them as well as generate new inferences that might not have existed before. These are more easily done with "symbolic" representations.

Combining with the strengths of DL based approaches: more robust, learns complex mappings, can learn non-trivial computations, can exploit indirectly specified structure, can approximate probability densities if trained correctly, will get the best of all worlds.

Here's a short presentation on the topic from a recent workshop: (and if you can, it's worth checking out the other presentations too) https://www.youtube.com/watch?v=_9dsx4tyzJ8

Yeah, it's the same general principle of using a model to cheaply speed up policy learning. An advantage to their approach however, is that it learns a latent space and generalizes better.

The VAE learns a compressed vector and the latent variables are somewhat meaningful. The VAE can also be sampled from and is not just a table of memorized examples. The RNN maintains coherence with actions and observations of previous time-steps and a separate controller is also learned. The end result is their approach is richer and more flexible.

9.9 times out of 10, I'd be in strong agreement with you. This time however, I think the use of the term "dream" is not as egregious as would be typical. When we dream, we often believe we're conscious and multiple senses will seem to corroborate this. It must be the case that we're interacting with a generative world model. Hippocampal replay is hypothesized to, in part, facilitate what their model is doing when they call it dreaming.

Hallucination on the other hand, is a less defensible use here. Hallucination refers to when predictions overwhelm sense input. Since their agent was not behaving in a way uncorrelated with its inputs, such as its predictive model overriding input data, it doesn't qualify as hallucination.

Analogies can help but if taken too far, will provide only illusionary knowledge from superficial similarities. Consider, if you trace out the hierarchies induced by interactions in biological networks and then do the same for an OS's call graph, you find quite different topologies reflecting their different priorities. In computers, in part, efficiency and reuse. In biology, robustness, minimized interdependencies, developmental stability and more.

There are things that concern our electronics that do not matter for biology and there are many things biology must allow for that our hardware cannot. Tying things down to the peculiarities of our systems too narrows the scope of applicable models and understanding. Sometimes it is more useful to think in terms of stochastic differential equations than in terms of operating systems.

Linking biology to specific concepts in computer architecture or worse, with OS design, is indeed a stretched analogy but biology is rich with computation. I'll try to be short.

A cell must sense and respond to its environment. One aspect of this is regulation of gene activity by interactions of for example, transcription factors. When interaction types can be well enough approximated as either of inhibition or activation, you can model them with boolean networks and when levels matter, some have found moderate success with recurrent neural networks under a restriction on what NN nodes represent, to ease interpretability.

That is not the same thing as saying your genome unrolls neural networks. What it is actually saying is that the complexity of the best performing neural network indicates the richness of the underlying computation, and the predictive accuracy of the model captures the functional equivalence of the respective computations (in the cell and in the neural network). The learned model is the instance and neural networks are the class, it is a mistake to place emphasis on neural network, they are merely a way of packaging the computational model of interest.

A cursory review here: https://www.sciencedirect.com/science/article/pii/S001048251...

If you have time for videos: https://www.youtube.com/watch?v=vA9D727AGHI , https://www.youtube.com/watch?v=ZX-GrO2qXQM

More relevant to this article is the study of cancer within the framework of evolutionary game theory, as done here: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2768082/

There is a precise correspondence between that formulation and one done in terms of online sequential prediction or the so called experts algorithm. All of these are models and still quite limited at that but biology is complex and the future lies in the development and leveraging of these rich connections.

Whether it's okay to ignore word order really depends on your goals. In my work, tracking word order to build a sentence representation is necessary to get composition rules correct. You can get away with ignoring order surprisingly often but when it's needed, it's vital. When aggregated across an entire corpus, order ends up mattering a lot, in an absolute sense (something looking only at accuracy numbers will hide).

Compare the following:

1) You must state the benefit.

2) You must benefit the state.

3) Feed fish vs. Fish feed.

Sometimes limits in inference:

1) A loves B.

2) Cats eat fish.

3) Cats are mammals.

Sometimes the sequence preceding or following will be important:

1) Problems with learning

2) Learning with problems

3) It has a learning problem

4) A learning problem for it.

Or consider an example from:

https://www.researchgate.net/publication/2335962_How_Well_Ca...

1) It was not the sales manager who hit the bottle that day, but the office worker with the serious drinking problem.

2) That day the office manager, who was drinking, hit the problem sales worker with the bottle, but it was not serious.

Order ends up mattering when you want to capture nuance or power higher order inference.

An important observation is that the probability of death drops very rapidly with drops in speed. A rough calculation shows that at 38 mph and a near instant reaction (as a robot should) would within ~1 second of applied breaking force, allow going from a ~70% chance of death to ~5%.

A very rough calculation: v(t) = -μgt + v_0. To reach 32 kph or 20 mph, t ~= 0.82/μ, guessing a fair μ=0.8 and the chart here https://ec.europa.eu/transport/road_safety/specialist/knowle...

As of this posting, if you put a "what" in front of the query, it returns the correct average distance. I suspected the parser, even if it was learned, is sensitive to the query format.

Despite what the press teams of large companies will tell you, our ability to model language is still in its early infancy.

This is a poor example considering evolution is anything but brute force trial and error. Each organism can be considered either an hypothesis or a collection of hypotheses about the background environment (https://arxiv.org/pdf/0911.1763v3.pdf, https://aeon.co/essays/consciousness-is-not-a-thing-but-a-pr...).

The algorithms which allow AlphaGo or the Libratus Poker AIs to achieve superhuman play have a direct correspondence with natural selection (and learning rate with selection strength). There is idea sharing between evolutionary game theory and learning, up to algorithms to play extensive form games with imperfect information such as this, based on replicator dynamics: http://dl.acm.org/citation.cfm?id=2617448

Evolution, per genome trajectory, also improves in its ability to evolve, as seen in evolvability.

Returning to the grandparent's original lament on stochastic gradient descent, I do agree with them. I suspect the need for better has not been seen due to non-supervised and incremental learning remaining minor areas of study. Artificial separation between learning and prediction allows "hacks" like batch-norm to somewhat suffice. It does seem unlikely that we will never need to take into account curvature of information manifolds while learning. Note that evolution uses curvature too.

Anything order than SGD has proven expensive but there are promising approximations, as found in KFAC or Projected Natural Gradient Descent, allowing me to close this post with yet another link between learning and evolution.

Correction to the above: I stated Deepmind applied Neural Nets+MCTS and achieved ok results. I was actually misremembering two David Silver (Deepmind) papers as one. Smooth UCT modified UCT (popular brand of MCTS) to be able to handle imperfect information games. MCTS does not converge under imperfect information. Smooth UCT is strong at limit poker. Limit is much simpler than no-limit.

Neural Fictitious Self Play based on fictitious play (invented 1950s), is an approach to reinforcement learning using neural nets for function approximation. Typical RL methods like DQN are highly exploitable. Against strong programs, NFSP did okay, with a win rate of -50 mbb/h against the best bot it played against.

Looking not just at Deepmind, there's Deepstack. It's similar to AlphaGo OG, combining CFR+Neural nets. Deepstack did not win convincingly against humans at 2 player no limit hold em.

The general point I'm trying to make here is that Chess and Go are closer to checkers than to poker, which is itself a constrained game with known rules. I mention all this and this Deepmind paper: https://arxiv.org/pdf/1711.00832.pdf, to provide a sense of scale to those talking about smoke and fire alarms.

The difficulty of imperfect information is from cross cutting through information sets and partial observability. With perfect information games like chess or Go, one can solve subgames with guarantees that the equilibrium is the same as for the full game. This is not the case for games like poker, which is why they have been difficult. In addition to that, for n > 2 players, there are no longer theoretical guarantees about converging to a nash equilibrium, which makes designing theory guided algorithms harder. Though empirical performance with n=3 of CFR is encouraging, I know of no results for n > 3.

Earlier this year, DeepStack, a system combining neural nets with search, competed live against humans without any side being dominant. Search policy guided training might improve its results, which are impressive compared to even 5 years ago, but this highlights how much more demanding imperfect information games are.

Let's break this down and consider things carefully. To informed researchers, what is most surprising here is not that the AlphaGo Zero algorithm beat stockfish but that MCTS managed to outperform Alpha-beta search. I'll venture a hypothesis as to why this was.

Informed skepticism would have discounted MCTS against alpha-beta search but wouldn't have put much stock into the idea that Neural Networks couldn't learn better features than what has been painstakingly handcrafted. We know that given sufficient data and an appropriate architecture, neural nets have achieved better local minima than humans. This shouldn't be surprising anymore. A structurally adapted searcher will always do better in its adapted to domain. A Cat is so good at being a cat, it doesn't even have to think about how to cat. Choice of optimization method, input pre-processing, loss function, hyper-parameters and architecture together define a search space, a structural prior and how to navigate.

Returning to alpha-beta vs MCTS, my view is that earlier work on the chess search space being ill-suited to MCTS has not been invalidated once you account for the synergy between the neural net and search method brought about by the imitation learning approach. What might be happening here is the neural net not only learns to correct when it goes out of bounds, it also learns to account for missteps of MCTS!

The AlphaGo Zero Chess Program is clearly smarter than stockfish from the perspective of its ability to better navigate the search space but before talking about fire alarms there are some things to note.

Assuming the paper, AlphaGo zero does well if you hold compute fixed and adjust time, but how does it do as you move along both compute and time? This is of relevance to the general community, especially if AlphaGoZero skill degrades gracefully enough to allow it to be a better tutor than current engines.

Contrary to the no fire alarm claim, we should see sudden improvements everywhere due to how close joint, structured prediction, reinforcement and imitation learning are to each other. Unexpected improvement across a broad class of problems is a fire alarm. Right now, POMDP or games with hidden information and multiple interacting agents are still very difficult. Structured prediction is still difficult. Granted, this was before AGZ, but Neural Nets+MCTS had to be modified to Neural Self-Play before it could work just ok in poker-like games.

What we should take away is the power of combining searching and learning. I'll argue that what is now being called expert iteration was presaged in an antique 2006 paper [1] where Hal Daume et al discuss the power of a learning algorithm trained to imitate a search computed policy. Even with limited compute and data, you can use similar ideas under the learning to search framework. The imitation approach is what's consistently yielded great results, whether applied to neural nets or logistic regression.

[1] http://www.umiacs.umd.edu/~hal/docs/daume06searn-practice.pd...

https://link.springer.com/content/pdf/10.1007/s10994-009-510...

I disagree. I found his meaning clear. The core of the essay is that thinking with the right abstractions provides greater boons to reasoning capacity than comparatively trivial differences in human IQ. To make things a bit more concrete, take any game; Go, Poker, Chess. If these were truly (approached as) combinatorial problems then humans could not stand a chance at them. But over time, people have learned patterns, compacted them and shared them such that the average level of play is what masterful players might have once struggled to reach.

Lisp provided concepts captured enough structure that it cut through a lot of combinatorial complexity in the space of possibilities, making reasoning about the construction of computer programs and algorithms a lot less random.

A complaint might be he left out what aspects of Lisp provided that combinatorial boost but I'm sure many HN users could furnish those reasons. But one easy observation is the sheer number of languages and language concepts that are rooted in it. Newtonian physics is not the best physics but it fundamentally changed the way people reasoned about the world. Lisp is not the best language but it fundamentally changed how people reasoned about computer programs.

One objection one might make is that the real Maxwell/Heaviside Equations can instead be found in the Lambda Calculus/ISWIM. Despite the name, Lisp did not originally derive from the lambda calculus but just about all typed functional languages and future versions of lisp are of the LC. But a quick answer to that is influential languages like SASL and Meta Language were prototyped in and strongly influenced by Lisp (cognitive boost).

As someone of skill in both topics, I was similarly confused but my misapprehension remained for a few seconds more. As luck would have it, the first sentence my eyes alighted on was Trees and Concrete Data.

That a process is non-conscious does not necessitate that it is primitive or incapable of sophisticated inference and computation. Sophistication and conscious reportability are absolutely not anti-correlated. Indeed, if I were to lean one way, it would be against the things we consciously perceive as difficult, inferring from the Moravec Paradox.

Non-consciously accessible processes play an important role in our ability to reason, predict and control behavior in a manner Evolution could not anticipate. They are no less us.

Yet, together with the machine, given no other recourse, I would have been able to extract the essence. As an example, I originally missed your english text and thought the task was to translate the german. I ran the garbled german through a translation engine and got something that made not much sense. It was still enough for me to guess that the intended meaning should not be too distant from "a bilingual speaker would leave the machine translator in the dust".

For what it's worth, I consequently ran this by a native german speaker. It's perfectly reasonable to rank them, understanding is not a binary proposition.

An AI can be like an electric bicycle, multiplying your existing ability. It can be like a car, working side by side. Or it can be like a self-driving car, leading your driving skills to rot with time.

You're assuming the only form is the last kind. As a human, I most prefer the electric cycle but I expect most will tend towards some kind of car. For the foreseeable future, only cars and cycles will exist because we don't yet know how to achieve deep understanding.

The algorithms understand in the sense of a set of rules to follow, transforming inputs to desired outputs. The rulebook doesn't merely consist of a look up table of all sentence pairs therefore at least some kind of understanding must be present. The algorithms can't turn words into stories with actors and predictions about their internal state, so I agree the understanding is shallow and far from human like.

Humans can often, even if not all choose to, step outside the system by integrating various modalities, propagating errors, to realize they're being forced to learn garbage. The machines can't reason outside the system. Compare with an animal of pure instinct (which are less common than many have supposed) whose behavior displays an implicit understanding of its environment's statistics.

DeepL translate: Das heißt: wenn zwischen Sprache A und Sprache B unbedingt etwas übersetzt werden muss, wird 2017 ein Zweisprachiger den Boden _absolut mit maschineller Übersetzung wischen.

At first I accidentally left in the underscore with Bing Translate I got:

Zum Witz: Wenn etwas absolut übersetzt werden muss zwischen Sprache a und Sprache B, in 2017 wird eine zweisprachige Person noch _absolutely MOP die floor_ mit der maschinellen Übersetzung.

Removing them, Bing Translate: Zum Witz: Wenn etwas absolut übersetzt werden muss zwischen Sprache a und Sprache B, in 2017 eine zweisprachige Person wird immer noch absolut wischen den Fußboden mit der maschinellen Übersetzung.

I don't speak German but mixing and matching between the three, I'm guessing the ranking for this example is Bing, DeepL, Google.

Don't have anything relevant to say except to make a fun note that we can change your second sentence, keeping it correct, while increasing its ambiguity. If we restate it as:

Are they talking about basically getting the most out of the current ML type systems?

Since Deep learning is a subset of machine learning, the sentence retains correctness but there are now two equally valid interpretations.

Okay, I do have something relevant to say. I don't see what advantage there is in raw sensory data streams, models can be trained off-line to operate on sensor streams just fine. What we actually want are systems that learn adaptively and on-line. To do that well, they'd need to also be data efficient.

As with almost all things machine learning, python libraries are where you will find most of the action.

I can recommend https://spacy.io for a low fuss solution to get you up and running quickly.

Oversimplifying quite a bit, if spacy focuses on syntax, then gensim focuses on semantics (https://en.wikipedia.org/wiki/Distributional_semantics). Gensim has an active community and is well documented.

https://radimrehurek.com/gensim/

If you have the data, can spend a few days experimenting and if you want something that can be orders of magnitude faster than deep learning to train, there's vowpal wabbit. Prediction speed is blazing. Results can be nearly state of the art but with a cost that's a great deal less. It's C++ but with bindings for many languages. It's very poorly documented.

https://github.com/JohnLangford/vowpal_wabbit/wiki/Learning%...

I've never taken a gander at Facebook's fastText.

I agree with you. I remember skimming the paper and finding it interesting but wondered if selecting that term might set too many imaginations running. I don't have a problem with their calling it imagination, though.

But consider that many would object if I stated jumping spiders have an active imagination. Yet, this is not far fetched if you accept that imagination includes planning against a learned model.

Insects are known to be capable of learning. They need to be able to remember routes or learn which locations to prefer or avoid. Jumping spiders are known for their ability to carry out and hold complex plans in their head. Though a form of imagination, most would hesitate to call it that.

(1) http://science.sciencemag.org/content/355/6327/833, https://youtu.be/exsrX6qsKkA?t=44s

Dependent or even refinement type systems are where a great deal of research into proof carrying code is done (for refinement types, since systems are restricted to decidability, the proofs are automatic but more limited).

Although it's true that there are many examples of multiple independent discoveries (supporting the zeitgeist theory of invention), this does not fully explain the nature of discovery and invention.

Consider that many of the examples, when examined carefully, fail to maintain the independence constraint. If the inventors were not in communication with each other but were consulting the same sources, or studying the same past failures (which was often) the case for independence becomes weaker. Nonetheless, one can still make the argument that since all the pieces were in place, that the inventor was less important. Three things stick a wrench in this: people who were truly ahead of their time, when there are many qualified yet with only a few succeeding and when there are outstandingly prolific individuals.

1) A notable example of someone ahead of their time is Hermann Grassmann, who toiled for many years in relative obscurity, quietly nursing a zoo of groundbreaking ideas. Gregor Mendel is another example. His discoveries took 30 more years to be independently rediscovered by 3 different individuals when the time was more ripe.

2) At each point in time, one can argue that many individuals met the criteria, from an intellectual and knowledge standpoint, to produce an invention yet only a handful do. There were plenty of people just as or even more intelligent than Einstein. Some even got very close to making the breakthroughs he did but none ever quite did. It seems genius individuals are able to grasp enough of a space to take steps that are locally suboptimal but end up escaping local optima for far better optima. There's a great deal of chance to this and it's not an issue of just out of the box thinking and high IQ (but sadly for the rest of us, it's likely still an innate quality).

3) It's difficult to look at the works of Gauss and Euler and not think there was something special about them. Genius individuals tend not to be one hit wonders and usually have something of an outsized effect on their field. While it's true that there is a rich get richer effect, the ability of geniuses to generate novel ideas and fields of investigation slightly outpaces their peers. And when advantages multiply, which is likely the case for productivity, these unique individuals end up dominating.

In the language of machine learning, geniuses seem able to sample more widely and are better at learning to learn. Even though it's likely that they're only a bit better at those things and they don't think qualitatively differently than the average person, it still ends up making a large amount of difference.

All in all, 2) impacts 3) which makes it so multiple independent discoveries can more readily occur. This doesn't mean the importance of environment can be discounted. Genius only appears under conditions of stability and sharing, the availability of gifted mentors and a rich set of freely shared and interacting ideas.

I was directly replying to the asked question. But, it is not just the developmental timeline that was off. There's also its hypothesis of what children learn and how they reason or what they are capable of that is inaccurate.

Child centered, problem centered is vague. More practically, the big things are: you want material that's novel to the learner but also somewhat familiar (this is true, regardless of age). Children's attention is more distributed and worse at blocking out irrelevancies. Therefore, more complex tasks (such as math) should be presented in a way such that the signal meant to be learned can be extracted with minimal ambiguity (less extraneous information in presentation). Few other things have replicated.

It's nothing at all like that. I already mentioned the Selection task above, it's a very googleable term. It's old and there is yet no theory fully explaining why people fail at it as they do. Pragmatic reasoning schemas are the nearest attempt at an explanation. https://en.wikipedia.org/wiki/Wason_selection_task

The main failures of Piaget's work is that it assumes a smooth progression in reasoning facilities with age and secondly, that its notion of formal operations rests too much on deductive logic. But humans do not really reason logically: affirming the consequent and denying the antecedent are fallacies often encountered. However, as the 19th century logician Peirce once stated, Not the smallest advance can be made in knowledge beyond the stage of vacant staring, without making an abduction at every step.

A famous counter-example is an abstract selection task which adults fail with high probability. Yet, when a structurally identical set of rules are given to ~9 years olds but couched in the language of permission, the children are able to pass with high probability. Stating the rules in terms of what the individual has experience with (or altering the linguistic phrasing) significantly increases the probability of solving much more than adjusting the age. So we have that on one hand, people develop capacities sooner (and more unevenly) than his theory supposed and that on another, some stages of reasoning aren't ever reached.

amasad is wrong to imply that the field has not progressed beyond the ideas of Piaget. Although, people do often get the important parts of Papert's ideas wrong by focusing too much on the computer part of computer-based learning. The computer should be a means to an end, a tool to help the learner explore more possibilities, make the abstract more graspable and amplify one's ability to ask better questions.

Top down predictions are likely mediated by feedback connections from higher to lower areas. Functions include possibly encoding a generative prior for prediction, speeding up inference. They also play an important role in coding more informative error signals than simple derivatives and are part of how the brain learns even as it predicts.

It's not that simple. The creativity is that the combination of rollouts, policy and value networks allow for more efficient traversal of the search space. Which gets you better exploration of possible paths, meaning more options than a human considered and therefore more creativity.

As others have pointed out, SGD requires differentiability. The next best match I can think of is that you're actually thinking of subgradient methods, which mostly see use in convex optimization problems.

They might be talking about Business Rules Engines which are elaborations of the Forward Chaining style of inference engine.