Some argue that synthetic data can make AI systems better

https://spectrum.ieee.org/synthetic-data-ai

by headalgorithm • 4 years ago

83 78 4 years ago

spectrum.ieee.org

If a computer program is generating the training data, aren't you just training the AI to do the same thing as the already existing computer program does?

throwawaynay2• 4 years ago

Generating a realistic city and a self driving AI are two wildly different tasks

SomewhatLikely2• 4 years ago

Even a simpler task like image classification such as: does the picture contain a lion. Imagine you have 3d model of a lion. You can render it from lots of different angles, lighting conditions, backgrounds, stretched out, curled up, etc. You know the ground truth classification on all renderings is that the picture contains a lion, but being able to generate images of lions is a very different task from recognizing lions in images.

ausbah• 4 years ago

at least in the case of reinforcement learning, no. just because you can simulate the problem doesn't mean you know how to optimally solve it - ex driving a car

btdmaster• 4 years ago

Not necessarily: https://en.wikipedia.org/wiki/Data_augmentation

ynfnehf• 4 years ago

You can train an AI to do the inverse of the existing program (as is the case for the self-driving described in the article.) Take some input, generate output using the existing program, and then train the AI with the input/output reversed.

tomp• 4 years ago

No. We had realistic 3D graphics 20 years ago.

I'm not aware of any true 3D computer vision system that could reliably play those games (from just vision).

csee5• 4 years ago

Not at all. The existence of the Gran Turismo game (a simulator) is not the same thing as an AI that can play Gran Turismo.

stingraycharles• 4 years ago

But a better analogy would be if an AI generated a computer game that another computer can learn to play.

In the end, it’s more like an anonymization layer than anything. If a computer is trained to generate input data for other computers to train with, there’s not a lot special going on.

taeric5• 4 years ago

But.... if you were to train a model on the simulator of the game, you would expect it to pick up on the rules programmed into the simulator.

This is really no different than it picking on the rules embedded in the data gathering of real world data. Any implicit and hidden decisions in that space would be expected to find their ways into the ML.

8note• 4 years ago

as a contrast to the other comments, you will still get biases from where the simulation differs from reality.

eg. if your simulated traffic lights dont blink at 60hz, the model trained on it wont know to handle it

wolverine8762• 4 years ago

I thought the same thing ...

Simulations are daily parts of life, in research and development from real-world simulations in research (e.g., climate) and industry (often in spreadsheets); to the theory of gravity (gravity simulated with mathematics) - and every other theory of science, social science, and humanities; to develping your iPhone app on your laptop or just reading the train schedule or using a mapping program.

So what is the difference? Those simulations were built from reality. The Theory of Gravity was built from and confirmed with empirical observations, not from someone else's simulation of gravity! That essential foundation of science, reality (it's not science otherwise), is what is missing.

Also, we already have the problem of our biases and preconceived notions infecting training data, and AI becoming a simulation of that rather than reality. By then training on 'simulated data' (yikes!), we seem to create more of a loop.

rapjr9• 4 years ago

If machine learning (ML) is trained on human behavior it can never be better (for some measure of better) than people are. So if racism is widespread, it will influence decisions trained into a ML algorithm. That raises the question, can we train an ML algorithm to make the decisions we _want_ rather than base its decision making on what people currently do. Training on generated data might be one way to do that, to build in implicit biases because we want more fairness in decision making than people currently exhibit. But then who gets to choose what those "goal" biases are? Someone could add a bias that improves life for people of male gender and worsens it for everyone else for example. This seems like a really important problem that has no clear answer. It's very much related to electing politicians where the choice of politician is also a choice of future goals. We don't know of any objective way to always decide what goals are best for everyone (voting certainly is not objective since many voters are not aware of all possible policy/goal implications and can be tricked into voting against their own interests). Yet it seems certain that researchers are working to introduce selective biases into algorithms, if just to adjust for known biases that are problematic. As an opaque input into an opaque algorithm, biases become invisible in deployments and could become very difficult to reverse or fix later. Even with continuous learning, if people's behavior is the input, the result will never be better than people are currently, which can create problems for the future. Some kind of intentional bias aimed towards goals seems necessary, yet also seems very dangerous since it can introduce biases decided by a very small set of people.

RicoElectrico• 4 years ago

The fact that you can compute y = f(x) doesn't imply you already know x = f⁻¹(y).

WalterBright2• 4 years ago

If you want an AI that can recognize cats, and train it with computer generated pictures of cats, you may wind up with an AI that only recognizes cat pictures from that particular generator.

renewiltord• 4 years ago

An obvious counterexample. I was young in 2005 when Tesseract was open sourced. I wanted to use it to do something. But I decided to trial it out first by writing something in Notepad and then screenshotting it and trying. Synthetic data! But no, I didn’t know how to make the existing computer program turn image into text.

SomewhatLikely2• 4 years ago

Pretty clickbaity. Lots of "some argue", "some say", "has estimated", and "striving to", but not much substance about actual successes. I believe both Tesla and Cruise are working in this direction but there are serious issues to be worked out. I also vaguely remember some work on pose estimation being helped by generating renderings. Going over real successes would make for a more convincing article.

chestervonwinch• 4 years ago

https://en.wikipedia.org/wiki/Weasel_word

MaxBarraclough• 4 years ago

My thought exactly. Also, HackerNews submissions are generally meant to use the title from the article.

fxtentacle2• 4 years ago

Odd, my impression is that everything moves in the opposite direction. Good synthetic data is expensive. Recording the real world is free.

Just a variational autoencoder with a discrete latent space is good enough to learn a usable phoneme recognition and pronunciation model from raw WAV files with unsupervised learning.

And clip shows is just for far you can get without supervision. So what's the point in paying for artificial data if you can solve the problem without it?

krapht• 4 years ago

What. That really depends on your task. In my field, real data is extremely expensive and if we could generate believable synthetic data, we'd save a ton of money.

hervature• 4 years ago

if we could generate believable synthetic data, we'd save a ton of money.

It sounds like you are saying good synthetic data is more expensive?

nomel4• 4 years ago

Recording the real world is free.

Waiting for edge cases to occur, in the real world, is definitely not free.

vhold• 4 years ago

With synthetic data you can also generate the thing you are trying to solve with the AI in the first place, like generating a depth map, and an object classification map to go with a simulated image.

With real world data a human will have to label it.

fxtentacle2• 4 years ago

Yes but AIs trained on synthetic depth maps tend to learn the oddities and noise patterns of the rendering, not actually the depth recognition part.

That's why so many AIs that work almost perfectly on the Sintel benchmark (synthetic flow and depth data) then fail to transfer over to KITTI (real images with lidar).

zhobbs• 4 years ago

Recording real world data can be cheap for some use cases, but often labeling it is very expensive.

variaga• 4 years ago

Recording real-world data may be cheap (not "free") in some circumstances, but properly labeling real-world data so it will be useful can be incredibly expensive.

Even if you're doing completely unsupervised learning, you will still need correctly labeled data at some point to validate your model.

And - I tell you as someone who has actually tried to do this - getting a wide range of high-quality voice recordings across multiple speakers is NOT CHEAP, and trying to train a model that will work under real-world conditions using "whatever dataset I downloaded from kaggle.com" is a fool's errand.

daenz• 4 years ago

Some years ago I worked at a startup that was doing OCR on paper receipts. As part of my application to the company, I wrote a synthetic training data generator[0] to generate a range of CG receipts, along with pixel-perfect accuracy of labeled XY bounding boxes for each letter. Generating synthetic training data allows for a high degree of flexibility to the shape of your data. It allows you to focus on strengthening edge cases where you just don't have enough real world data.

0. https://www.arwmoffat.com/work/synthetic-training-data

nomel4• 4 years ago

When developing anything in the real world, you don't wait for edge cases to happen naturally. You force them to happen, and look at the response. This is the rigor of engineering. Has that been lost?

snek_case3• 4 years ago

There's this dogmatic idea in the machine learning field that only real world data is valuable.

Frost1x• 4 years ago

I often here "edge case" as a defense as to why something hasn't been developed or handled. There's an ongoing trend in software to avoid high hanging fruit and focus on the low hanging big and fast returns. I just accept it now although it bothers me everytime I see it happen.

The beauty and allure of science and engineering, to me, is understanding something so well you can predict its future behavior and then predicting it so well you can prevent undesired behaviors and create or improve desired behaviors, within some set of bounds of course.

zwaps3• 4 years ago

Well, to test the system sure. That's normal.

However, the issue not quite as simple when training a model. For instance, the increasingly popular Transformer architecture pushes around internal representations when fed new training samples. These representations are not "causal parameters" - they do not represent unchanging mechanisms in real life. Instead, they depend on data.

The question thus becomes: How prominent should edge cases be in the sample? A ML model trained on many edge cases may be more robust for edge cases we feed it, but it may be more unstable on real-world data (in essence, the model may have new edge cases when new data comes in that would not have been edge cases before).

All of this comes from the fact that the parameter of the model are not "causal" for the underlying latent DGP. ML, in constrast to stats, does not yet have much theory regarding how to think about parameters relative to the sample of observations one trains on. The question of "causal" identification in stats is easier to tackle, because one can usually define a "scientific" or "structural" model where parameters have meaning, and so we can reason whether we capture "real, causal parameters" (which would be compatible with edge cases) or "reduced form parameters" (which would generally depend on the sample we have). ML will likely get there. Right now, people think in terms of "data shift" (which they shouldn't) and generalizing (missing the point a bit). But we will get there.

By the way, this issue arose because ML systems where treated like engineering systems instead of statistical models. For better or worse.

Fricken• 4 years ago

In autonomous vehicle development 'edge cases' are the things that engineers would never conjure up via thought experiment. They need to be discovered through real world testing.

Once an edge case has been discovered, they can then artificially generate myriad subtle variations on the edge case, which they use to train their systems.

bpodgursky• 4 years ago

This is relevant to healthcare startups where it is extremely difficult to get your hands on enough real Protected Health Information to do any interesting ML work (unless you are already part of an enormous company with PHI... and even then it is harder than you'd think).

rosenjcb2• 4 years ago

I worked at a health insurance company and had access to a lot of data (for ML research). It was impossible to lend that data to contractors for nearly any reason and this frustrated us.

sbrother• 4 years ago

"Some argue"? We've been doing this for the last ten years and that's just as far as my career goes back.

rosenjcb2• 4 years ago

You're right. This isn't new or controversial. I don't know why HN is so weak on data science. Maybe software devs are weak on math in general.

fhood• 4 years ago

What a strange article. I think that a company like Nvidia could probably provide a great deal of additional value to the world of scientific computer modeling. And I think that simulations can work really well to assist in training models, I don't really understand why that would be up for debate, they already do. What I don't understand is talking about "simulating the entire world down to atomic interactions" or teleporting to Mars via data collected from..."sensors".

Little sections of this, particularly wrt pushing towards more standardized systems for building computer models make sense and seem like a worthwhile goal, but most of this reads like nonsense to me.

freddealmeida• 4 years ago

One of my companies has a patent in this space (Neuri) (rather defunct now). It worked exceptionally well for time series data.

Used similar work at Ascent.ai for robotics data (visual mostly) and that worked very well.

In fact I don't think we ever really saw an issue with this approach.

NalNezumi• 4 years ago

Ascent as in the startup in Tokyo? Didn't the self driving car approach fail so miserably they had to 180 degree pivot to P&P robotics applications?

Interesting, but Sim2real have had some lab success for the past few years. Making it work in real application seems to be way more trickier, especially profitably

YeGoblynQueenne4• 4 years ago

> The solution is to just have more data and better data.

Nah, sorry, that is just trying to put out the fire by throwing fuel at it. The big, big weakness of neural networks right now is their reliance on gigantic datasets that require gigantic computational resources. Neural nets need those because they can't generalise to unseen data. So people try to get them to see as much data as possible during training. They still overfit, but if they can overfit to a diverse enough dataset, then they can be useful in practice, even if that's only to solve narrow, specific instances of a problem (like in the domino recognition system in the article).

To address this weakness what is needed is to find ways to make neural nets less reliant on data, not to find ways to make more data. Make neural nets capable of generalising robustly to unseen data, from few training instances. Then you don't need to train in a simulation. Of course that would require a radical rethink of how deep neural nets are trained (even perhaps whether they remain "deep", or whether they are trained using gradient descent, the sources of their data-hunger). Trying to make more data by simulation is only kicking the can down the road and the only effect it can possibly have is to push the time at which the real limitation must be really addressed even further down the line so that an other generation of researchers has to deal with it while the current generation can keep getting their papers published and their grants granted.

See Vladimir Vapnik's challenge to the machine vision community:

https://youtu.be/bQa7hpUpMzM?t=492

To summarise: learn to identify MNIST digits from 60 examples of each class, rather than 6000, while retaining current accuracy. (My words now:) Improve sample efficiency to improve neural nets. Neural nets have shown a remarkable ability to work well when large amount of resources are available. Now, do like everyone else does in computer science and try to make them (sample) efficient.

teruakohatu• 4 years ago

learn to identify MNIST digits from 60 examples of each class, rather than 6000

More like 60,000 per class once augmentation has generated a bunch of new samples from each orginal!

treesprite823• 4 years ago

To summarise: learn to identify MNIST digits from 60 examples of each class, rather than 6000,

SOTA accuracy on a similar but more challenging problem (5-shot 20-way rather than 60-shot 10-way) appears to be around 99.6%: https://paperswithcode.com/sota/few-shot-image-classificatio...

while retaining current accuracy.

Depends how strict you're being with this. There's room for the gap to shrink, but I think on average classifiers (whether organic or machine) with a large number of examples to go off of will always perform at least marginally better than classifiers with fewer examples.

YeGoblynQueenne4• 4 years ago

> SOTA accuracy on a similar but more challenging problem (5-shot 20-way rather than 60-shot 10-way) appears to be around 99.6%: https://paperswithcode.com/sota/few-shot-image-classificatio...

I'm aware of results like that but they're just kicking the can down the road with the other foot: they push the problem of training with big data to the pre-training stage and then claim to do "few-" or "one-shot" learning at the end, or even "zero-shot" which is egregious abuse of terminology (and it's very sad that it's accepted terminology). It's like the Aesop's fable where the sparrow hid in the eagle's feathers and jumped up at the last moment to claim "I'm the bird that flies the highest!".

Vapnik's point in the interview I linked above is that you should not need a lot of data for anything, including pre-training. His challenge is for the community to find what he calls good "predicates" which are primitive functions (I think of them as feature detectors) that can be composed into a good statistical invariant, a function representing a high-level concept while having good out-of-dataset generalisation ability. His claim is that if you have a bunch of good predicates and good invariants, then you don't need a lot of data, because the generalisation ability of the good predicates makes up for it. Or, seen another way, lots of data is needed _in the case when_ a good predicate is not known. In a certain way, transfer learning, or meta-learning in the case of the paper you link, is a step towards the right direction, but the reliance on big data for pre-training suggests that the models are still not learning good representations that generalise well - so they still need big data to make up for it.

twayt• 4 years ago

As long as (stochastic) gradient descent and random initialization is the state of the art, more data will be better.

kd5bjo• 4 years ago

Random initialization is already starting to give way. One of the more promising approaches right now is pretraining with synthetic data to obtain “good” initial weights. Once that converges, you’ve got a network that internally recognizes interesting features which is a bettter-than-random starting point for training with real-world data.

cf. https://arxiv.org/abs/2106.05963

zwaps3• 4 years ago

This is always an interesting discussion. I think what you say is probably right "in practice" - that is, it is where ML should move now to improve its ability to generalize.

However, from a more statistical viewpoint, this approach is not "theoretically correct". The issue of unseen data arises for one simple reason: the model does not capture causal mechanisms in the DGP. The model instead learns a reduced form that represents the training sample. If the training sample is large, and the model expressive, this leads to incredible results. However, it fundamentally differs from a causal model in that it could not do true out-of-sample predicitions if these are counterfactual to the data. That is, it may work, but we can not rely on it. It may randomly fail (and not in dependence on robustness of smoothness of the model, but in joint dependence to the counterfactual DGP, if that makes sense)

This issue stands independent of the amount of data. Indeed, if you had a model that identified all relevant causal mechanisms, then of course more data is still better.

In some areas, ideas from causal statistical analyses seep in. Some physical systems are well described by ODE and PDEs - which can be learned by DNNs but retain their causal structure.

I see no general way forward, though. The approach you propose moves along the general idea of ML/AI - pose better challenges to the model and see if it works. This has worked well, perhaps it is the way to go.

YeGoblynQueenne4• 4 years ago

> The issue of unseen data arises for one simple reason: the model does not capture causal mechanisms in the DGP. The model instead learns a reduced form that represents the training sample.

Thanks for articulating it so clearly, that's very much how I think of it!

> This issue stands independent of the amount of data. Indeed, if you had a model that identified all relevant causal mechanisms, then of course more data is still better.

Well, my point -and, as far as I understand it, Vapnik's point, in my link above- is that if a model can generalise well, then it doesn't need a lot of data, even if a lot of data is available. In terms of a model that can identify causal relations, then it should be able to do so without lots of data. My intuition for that is that if the model needs lots of data it will inevitably learn to represent the sampling error in the data generation process (DGP?). A model that can learn good generalisations from few data on the other hand is in a sense immunised against sampling error because it can ignore most of the data.

Vapnik makes an analogy about a good teacher. He asks, what is it that a good teacher does that helps his students learn better? He answers it by saying that the good teacher gives his students "good predicates", what I would call background knowledge, from which the students can build good statistical invariants. But the students must also be able to build good statistical invariants _on their own_, otherwise it doesn't matter what good predicates the teacher can give them, the student's can't learn good invariants.

I think that's the point of the challenge - Vapnik is asking for learning from few examples (not unreasonably few, I think) as a way to demonstrate that a system is learning good underlying principles- his good predicates and statistical invariants. In the past the machine learning community has sidestepped the issue by finding ways to augment data where it was scarce and then claim progress in certain problems for which there were initially few examples (for example, Bongard problems and Winograd schemas) but Vapnik is not only calling for few data, but also for good predicates, which is I think a better defense than just asking for few-shot learning.

In a sense I think if we find that more data is better, then that means the system is still not learning the good bits that make more data unnecessary.

naveen99• 4 years ago

Dreams are basically synthetic data that train our brains while we sleep.

hoppla• 4 years ago

How cool would it not be if the computers would not only use synthetic data for training, but also simulating the outcome of their own actions before taking it. But not only it’s own actions - like a game of chess it would also include all the possible immediate actions of other actors

ouid• 4 years ago

well that just sounds like programming with extra steps.

ddingus• 4 years ago

Does this not depend on the source data?

Some datasets are easy to render, generate, whatever. In that scenario, sure! Seems like a solid case can be made, particularly where an analytic approach can speak to the data elements needed.

Other data needs to be sourced from the world. That's harder, and it's extremely likely artificial data is either too expensive to create at the fidelity, or lack of, to make economic sense, or the cases are too numerous for an analytic approach to be inclusive.

Some argue that synthetic data can make AI systems better

Related Stories