4 ms·

I've yet to see an "AI" that is not just memorizing data.

KorfmannArno 6y ago
But then again, how does something like the following work? https://twitter.com/GoogleAI/status/1293970520753369088 https://twitter.com/GoogleAI/status/1293970520753369088 Any idea how it could be fooled?
- antipaul 6y ago
  Via a fancy “adversarial” patch https://www.theverge.com/2019/4/23/18512472/fool-ai-surveillance-adversarial-example-yolov2-person-detection https://www.theverge.com/2019/4/23/18512472/fool-ai-surveill...
- nl 6y ago
  What do you mean by "fooled"? I'm very familiar with the BlazeFace and FaceMesh models (which are related to this in Google's MediaPipe framework). They have weaknesses - they aren't designed for running upside down for example, so if they get data that is that way oriented they will tend to fail. They aren't designed for "life" detection, so you can show printed pictures of a face and it will detected it. But they give you confidence scores etc, so if you give it something like a caricature of a face it will return reasonable confidence numbers indicating it isn't as sure of its predictions.
nl 6y ago
Then you haven't really looked. Most credible machine learning systems work well on unseen data, which by definition isn't memorizing.
- KorfmannArno 6y ago
  Can you link an example you find to generalize particularly well?
  - nl 6y ago
    Sure. Apple's Face Unlock. It generalises to almost every face, and the ones it doesn't its failure mode is safe. Or something like word embeddings. Works incredibly well, and most "failure" modes are around things like bias, where the behavior reflects the real world. Or something like AlphaZero. Not only is every new game of Go it plays brand new, it learnt to play Chess without knowing the rules. That just isn't memorization. https://deepmind.com/blog/article/alphazero-shedding-new-light-grand-games-chess-shogi-and-go https://deepmind.com/blog/article/alphazero-shedding-new-lig...
    - KorfmannArno 6y ago
      https://arxiv.org/pdf/1805.12177.pdf https://arxiv.org/pdf/1805.12177.pdf I only read the abstract.
      - nl 6y ago
        Not sure what you think this shows, but there's a lot of reasons why the results they show don't really matter much - or at least might actually reflect an accuracy approximatly what a human would also achieve. Their headline claim is a "1 pixel change reduces accuracy by 30%". The test process for that number is this: We choose a random square within the original image and resize the square to be 224x224. The size and location of the square are chosen randomly according to the distribution described in (Szegedy et al., 2015). We then shift that square by one pixel diagonally to create a second image that differs from the first one by translation by a single pixel. So... they are taking a random square, downsampling to 224, moving and then predicting on that subset of the original image, and measuring the performance against the accuracy of the original prediction. What this seems to show is that "CNNs aren't as accurate at making predictions on subsets of an image as on the whole image". This is of course to be expected, and is exactly how a human would perform. There are a bunch of other criticisms too. Read the ICLR (for which it was rejected) reviews: https://openreview.net/forum?id=HJxYwiC5tm https://openreview.net/forum?id=HJxYwiC5tm
    - PaulHoule 6y ago
      Word embeddings suck. Take a look at the graphs under " 2. Linear substructures" in https://nlp.stanford.edu/projects/glove/ https://nlp.stanford.edu/projects/glove/ and note that they all involve looking at a small number of points. It is easy to reproduce plots like that but if you try to increase the number of points the result breaks down completely. It is a curve fitting problem: for a small enough set of points compared to the number of dimensions, you can find a matrix that projects a set of random points to an exactly specified set of points in the plane. If you relax the problem to something like "put colors on the left side, put smells on the right side" you will get better than random performance from that kind of model, but not that much better than random. Word embeddings are a strategy that approaches an asymptote. Systems that are destined to low performance will perform better if you use a word embedding, but they throw away information up front that makes high performance impossible.
- visarga 6y ago
  > Most credible machine learning systems work well on unseen data, which by definition isn't memorizing. Sorry, but no. ML models don't generalise well outside the training data, but they can interpolate inside. This question becomes very interesting in the case of GPT-3 which has had a huge corpus of text to train on, so it's probably seen 'everything'. It's still memorising for GPT-3 but also learning to manipulate data, like software algorithms.
  - nl 6y ago
    ML models don't generalise well outside the training data, but they can interpolate inside. I'm unsure if you just misstated this or don't know, but this is wrong. ML models don't generalise well on data outside the distribution of their training data. But that's an entirely different thing, and doesn't mean at all they are memorising data. Imagine something training on the US unemployment rate until 2020 being hit with the COVID rate. It wouldn't know what to do, but that doesn't mean it wouldn't work fine on a rate of 5.342% even if it had never seen that rate before. This is a simplified example, but applies to everything. GPT-3 generation of text does pull from memorised training data. There's a lot of stuff going on there, and amongst other things there has never really been a system that does textual generation well. It's also hugely overparameterised, so lots of potential for overfitting. I don't think it's a good example of a "good" AI system - it's very interesting, full of potential, but there are lots of issues.
visarga 6y ago
Just memorising data is simple, use a file on disk. The hard part is recognising data when it is slightly different than what the model 'memorised', deciding which of the millions of things it learned best fits the answer we desire.

But then again, how does something like the following work? https://twitter.com/GoogleAI/status/1293970520753369088 https://twitter.com/GoogleAI/status/1293970520753369088 Any idea how it could be fooled?

antipaul 6y ago
Via a fancy “adversarial” patch https://www.theverge.com/2019/4/23/18512472/fool-ai-surveillance-adversarial-example-yolov2-person-detection https://www.theverge.com/2019/4/23/18512472/fool-ai-surveill...
nl 6y ago
What do you mean by "fooled"? I'm very familiar with the BlazeFace and FaceMesh models (which are related to this in Google's MediaPipe framework). They have weaknesses - they aren't designed for running upside down for example, so if they get data that is that way oriented they will tend to fail. They aren't designed for "life" detection, so you can show printed pictures of a face and it will detected it. But they give you confidence scores etc, so if you give it something like a caricature of a face it will return reasonable confidence numbers indicating it isn't as sure of its predictions.

nl 6y ago

Then you haven't really looked. Most credible machine learning systems work well on unseen data, which by definition isn't memorizing.

KorfmannArno 6y ago
Can you link an example you find to generalize particularly well?
- nl 6y ago
  Sure. Apple's Face Unlock. It generalises to almost every face, and the ones it doesn't its failure mode is safe. Or something like word embeddings. Works incredibly well, and most "failure" modes are around things like bias, where the behavior reflects the real world. Or something like AlphaZero. Not only is every new game of Go it plays brand new, it learnt to play Chess without knowing the rules. That just isn't memorization. https://deepmind.com/blog/article/alphazero-shedding-new-light-grand-games-chess-shogi-and-go https://deepmind.com/blog/article/alphazero-shedding-new-lig...
  - KorfmannArno 6y ago
    https://arxiv.org/pdf/1805.12177.pdf https://arxiv.org/pdf/1805.12177.pdf I only read the abstract.
    - nl 6y ago
      Not sure what you think this shows, but there's a lot of reasons why the results they show don't really matter much - or at least might actually reflect an accuracy approximatly what a human would also achieve. Their headline claim is a "1 pixel change reduces accuracy by 30%". The test process for that number is this: We choose a random square within the original image and resize the square to be 224x224. The size and location of the square are chosen randomly according to the distribution described in (Szegedy et al., 2015). We then shift that square by one pixel diagonally to create a second image that differs from the first one by translation by a single pixel. So... they are taking a random square, downsampling to 224, moving and then predicting on that subset of the original image, and measuring the performance against the accuracy of the original prediction. What this seems to show is that "CNNs aren't as accurate at making predictions on subsets of an image as on the whole image". This is of course to be expected, and is exactly how a human would perform. There are a bunch of other criticisms too. Read the ICLR (for which it was rejected) reviews: https://openreview.net/forum?id=HJxYwiC5tm https://openreview.net/forum?id=HJxYwiC5tm
  - PaulHoule 6y ago
    Word embeddings suck. Take a look at the graphs under " 2. Linear substructures" in https://nlp.stanford.edu/projects/glove/ https://nlp.stanford.edu/projects/glove/ and note that they all involve looking at a small number of points. It is easy to reproduce plots like that but if you try to increase the number of points the result breaks down completely. It is a curve fitting problem: for a small enough set of points compared to the number of dimensions, you can find a matrix that projects a set of random points to an exactly specified set of points in the plane. If you relax the problem to something like "put colors on the left side, put smells on the right side" you will get better than random performance from that kind of model, but not that much better than random. Word embeddings are a strategy that approaches an asymptote. Systems that are destined to low performance will perform better if you use a word embedding, but they throw away information up front that makes high performance impossible.
visarga 6y ago
> Most credible machine learning systems work well on unseen data, which by definition isn't memorizing. Sorry, but no. ML models don't generalise well outside the training data, but they can interpolate inside. This question becomes very interesting in the case of GPT-3 which has had a huge corpus of text to train on, so it's probably seen 'everything'. It's still memorising for GPT-3 but also learning to manipulate data, like software algorithms.
- nl 6y ago
  ML models don't generalise well outside the training data, but they can interpolate inside. I'm unsure if you just misstated this or don't know, but this is wrong. ML models don't generalise well on data outside the distribution of their training data. But that's an entirely different thing, and doesn't mean at all they are memorising data. Imagine something training on the US unemployment rate until 2020 being hit with the COVID rate. It wouldn't know what to do, but that doesn't mean it wouldn't work fine on a rate of 5.342% even if it had never seen that rate before. This is a simplified example, but applies to everything. GPT-3 generation of text does pull from memorised training data. There's a lot of stuff going on there, and amongst other things there has never really been a system that does textual generation well. It's also hugely overparameterised, so lots of potential for overfitting. I don't think it's a good example of a "good" AI system - it's very interesting, full of potential, but there are lots of issues.

visarga 6y ago

Just memorising data is simple, use a file on disk. The hard part is recognising data when it is slightly different than what the model 'memorised', deciding which of the millions of things it learned best fits the answer we desire.