Hacker News Reader

Comment by buboard | Hacker News Reader

> even rodents can outperform state of the art models at general tasks

Rodents can't play Go or a lot of other humanly-meaningful tasks. We don't need to build an artificial cell. A cell is too many components that by blind luck happened to find ways to work together, this is as far from efficient design as can be. The same way we don't build two-legged airplanes, we don't need anything that's close to the wet spiky mess that happens in human brains. It's more likely that we have all the ingredients already in ML, and we need to connect them in an ingenious way and amp up the parallelism.

View on HN

AlphaZero has coded all the rules for the respective three games, they do a tree search and their neural network output layer has exactly n neurons for max(n) possible moves. Although it's impressive they don't teach it heuristics and strategies, it's a very specific task.

What about pigeons predicting breast cancer with 99% probability, rats learning to drive cars, monkeys building tools?

Rodents stand a bigger chance at learning Go than AlphaZero spontaneously building stone tools and driving cars.

You are talking about AlphaGo. AlphaZero was not given any prior knowledge of the game and is trained exclusively through self-play -- and it outperforms Monte Carlo tree search-based systems such as AlphaGo and Stockfish in chess 100-0 with a fraction of the training time.

AlphaZero is also capable of playing Chess, Shogi and Go at a super-super-human.

As impressive as AlphaZero surely is, I don't think it ever got a proper comparison to Stockfish. It was running on a veritable supercomputer while Stockfish was running in a crippled mode on crippled hardware.

Not working in this area but the abstract of the AlphaZero paper [0] seems to disagree about your /any prior knowledge/ point: "Starting from random play, and given no domain knowledge except the game rules, AlphaZero achieved within 24 hours a superhuman level of play in the games of chess and shogi (Japanese chess) as well as Go, and convincingly defeated a world-champion program in each case."

[0] https://arxiv.org/abs/1712.01815

This is my point exactly. The model is trained without any prior domain knowledge at all. It only has access to a game world where the constrains in the world is a representation of the game's rules.

You can view these as optimized pattern recognizer regexes. You start with a blank fully connected graph and it eventually converge on a useful function. That graph has many paths encoded in it that represents specific optimal game play.

Isn't this how the neurons and synapses in our brain work, though?

Maybe... there’s some other properties of biological neurons we don’t capture in NNs currently.

The natural environment encodes "all the rules" for real animals, too. You need some constraints or else there is nothing to be learned. One could say that every survival task is also specific , but is a slight variation of previously learned one.

> pigeons predicting breast cancer with 99%

pigeons contain 340M neurons (with dendrites and all, giving them higher computational capacity than ANN units).

> Rodents stand a bigger chance at learning Go

They probably don't ; probably because they can't understand the objective function and their brain capacity is limited

Scientist have just recently taught rats how to play hide-and-seek for fun. Other scientists have found out that slime mold will model the Japanese railroad system. I wouldn't be surprised if rodents (plural) instinctively have a go strategy once someone figures how to make an analog game for them.

its probably safe to assume that even if rodents are behaviorally trained to follow complex rules, they are mostly pattern-matching, and are lacking higher-level abstraction and communication models like humans do. If they did they would at least attempt to communicate with us, like we do with them. In such a case, an elephant that plays go by patternmatching is no different from a neural network that learned by patternmatching

The problem with the analogy is that the car, by far, is not a general transportation device. Practically, most cars are solving a very constrained transportation problem: moving on roads that humans made.

We don't have anything remotely close to a wetware-enabled transportation device, something that can move on flat land, climb mountains, swim in bodies of water, crawl in caves, hide in trees.

Within the constrained problem, the machine exceeds humans. But generally, the wetware handles moving around much better.

Same with AI: in a constrained problem, the AI can excel (beat humans in chess and go). But I doubt we will see a general AI any time soon.

> constrained problem

human AI also evolved by solving constrained problems, one at a time. Life existed before the visual system , but once this was solved it moved on to do other things. In AI we have a number of sensory systems seemingly solved: Speech recognition, visual object recognition, and we are closing to certain output (motor) systems: NLP text synthesis systems seem a lot like the central pattern generators that control human gait, except for language. What seems to be missing is the "higher-level ", more abstract kernels that create intent, which are also difficult to train because we don't have a lot of meaningful datasets. Or maybe , we have too big datasets (the entirety of wikipedia) but we don't know how to encode it in a meaningful way for training. It's not clear however that these "integrating systems" are going to be fundamentally different to solve than other subsystems. It certainly doesn't seem to be so in the brain, since neocortex (which hosts both sensory and motor and higher level systems) is rather homogeneous. In any case, it seems we 're solving problems one after another without copying nature's designs, so it's not automatically true that we need to copy nature in order to keep solving more.

> In AI we have a number of sensory systems seemingly solved: Speech recognition, visual object recognition,

Do you have examples of those systems which are competitive in general use rather than specialized niches? The cloud offerings from Amazon, Google, etc. are good in the specific cases they’re trained on but fall off rapidly once you get new variants which a human would handle easily.

There are many vision models where classification is better than human. I m not sure what you mean 'fall of rapidly'; they do fail however for certain inputs where humans are better. But we 're talking about models that contain 6 to 7 orders of magnitude less neurons than an adult brain.

It's also interesting in the context of how we build our technology in general: we constraint our environments just as much we develop tools that operate in them. E.g. much as cars were created for roads, we adapted our communities and the terrain around them by building roads and supporting infrastructure. A lot of things around us rely on access to clean water at pressure, which is something we built into our environments, etc.

> A cell is too many components that by blind luck happened to find ways to work together

Can't tell if sarcasm.

carbon chemistry + thermodynamics

!= "luck"

so you think cells had some insight on how to evolve themselves?

more like caused to happen by the Creator.

Who created the creator?

Reply on news.ycombinator.com