When not to use deep learning
hyperparameter.space
hyperparameter.space
If I take a multiple-choice exam and always answer "C", then I have a good chance at getting more than 25%.
For image recognition, I think the classifier is doing the real work (trying to actually answer the question), and the deep learning is just seeing if the answer matches the pattern of expected answers.
Somehow, this actually works. I think that it's because true randomness is hard to find.
The problem that I've found is that it's really difficult to teach deep learning. I'm making a Chinese-English teaching tool ( http://pingtype.github.io ) and sourcing my translations from Google Translate. I find a lot of mistakes in my dictionary that obviously came from Google's model getting the word spacing wrong. I can fix it in my own dictionary immediately. If I submit the correction to Google, it just changes some weightings, and hundreds of people will have to submit the same correction before their deep learning will finally catch on that it needs to change something.
No, your naive understanding is not correct.
'Deep learning', and by that I mean a neural network, works at a super high level by generalising some input (say X) into some related output (lets say, Y).
It's not just random choice; it's like defining a programming function:
foo(x) -> y { ... }
...but ultimately, you raise a valid point.
Training models is incredibly time consuming, and incorporating 'corrections' as new training data is extremely non trivial.
It's one of the issues with deep learning.
I would even go as far as to say, 'You need to regularly update your training data with new examples and counter examples' as a 'when not to use deep learning'.
Putting in 这是什么东西 ("What is this thing") in the Chinese box yielded the translation "this is Why tender east west", which is a _character for character_ translation of the phrase. That's not especially meaningful or useful given that Chinese words are polysyllabic - I hope this is a temporary bug.
By contrast, Google Translate gives me the much more sensible "What is this" as a translation.
You can click Advanced -> Regional -> Simplified to Traditional.
Entering 這是什麼東西 gives "this is what thing", which has correct word spacing.
Humans seem to solve problems with a combination of learned stock knowledge, induction, and constrained improvisation.
Constrained improvisation is the most interesting part of that process, and the one we know least about.
It's one thing building a system that asymptotically improves over millions of trials, and then sending out a press release claiming your system is as smart as a human.
It's another building a system that learns a domain as efficiently as a human.
Compare the relatively small number of games played/analysed by a Go master on their way to master status, compared with the number of simulated games played/analysed by AlphaGo.
ML is still a rather naive form of constrained brute forcing. It's a long way short of efficient learning.
Uh... Bayesian can do the above at in term of expert domain. They call it elicitation in the Bayesian world.
I think you're overall generalizing all learning techniques. And also it's not like we actually really know how the human brain learn. Psychology is a field with huge uncertainties and you can see that in their research papers with correlation values. So the concept of learning may be out dated and/or we are still learning about what makes us learn.
This isn't a problem with deep learning, but a problem with verifying a data source. You are trusting yourself to submit a proper correction, but Google doesn't know you from John.
To use another example, imagine that someone says 他的 is "they're", another person says it's "their", and still another person says it's "there". Your approach would just accept all three in a row, wouldn't it?
Google's approach presumably trusts that the crowd + what it sees on the web is correct and thus attempts to verify that a submission is typical language use before actually putting it into play.
I'm not convinced that Google Translate is as major of a focus at Google as other things.
Perhaps I'm wrong, and it's just a hard problem, but the translations I've seen haven't improved as much over the years as I expected given progress in other areas of AI.
They may be more focused on adding new languages than improving existing ones. Don't know.
Japanese translation is pretty good. Chinese is bad. I guess Chinese is harder because nobody really uses any phonetic writing for reading beyond grade schools. And then there is traditional/simplified, and, I imagine, other differences between how different regions use characters. Baidu's translate is better than Facebook/Google in some cases.
I wonder if translation just isn't as sexy as image recognition and self driving cars, therefore research dollars aren't as focused on it.
It's a very hard problem. There are a ton of people at Google, baidu, and elsewhere working on it. (Source: I'm a part timer on Google Brain)
I'm not sure about USA, but it's been a major focus of very large EU grant programs due to the obvious multilinguality of EU and the explicit goal to move towards a single European market by reducing barriers in trade, including language barrier.
It also has many commercial use cases and thus has always had quite a lot of people and teams working on it compared to other fields of ML or NLP.
The problem is that it's hard. Every 0.1% of progress has historically required a lot of work.
More research into theory, more / better data from professional translators, and more efficient implementations from engineers.
I think this is quite profound and inspiring.
Although perhaps it is only one half of it. Concretely, deep learning finds patterns, the best patterns are derived from the highest bandwidth signal, often this is the input.
Geoff Hinton has argued that the task, solving the problem, is a low bandwidth signal.
Hinton aphorises [approximately] 'If you want to learn computer vision first learn to do computer graphics, i.e. a generative model.' - this is about the bandwidth of the data signal.
Hinton: [1] "Each image has much more information in it than a typical label... Each image puts a lot of constraint on the identity function. Whereas if I give you an image and a label and I try and get the right answer I don't get much constraint on the mapping from image to label. The bits of constraint on that mapping imposed by training example are just the number of bits to say what the answer is which is not very many."
Wow, I love this sentence!
Plato: A man is a featherless biped
Diogenes (plucking a chicken): Here's your man, Plato!
Plato (deep learning): A man is a flat-nailed featherless biped
Superman is an alien.
> I haven’t found a way to properly articulate this yet but somehow everything we do in deep learning is memorization (interpolation, pattern recognition, etc) instead of thinking (extrapolation, induction, etc). I haven’t seen a single compelling example of a neural network that I would say “thinks”, in a very abstract and hard-to-define feeling of what properties that would have and what that would look like.
> All the while I'm thinking: this thinking process this person goes through as he analyzes this data: THAT is what Machine Learning SHOULD do
-- Andrej Karpathy
Deep learning for image recognition works because our visual world is made up of structured hierarchical features: Dark/Light, Texture, Edge, Part of Object, Object, Scene. Deep learning layers create increasingly higher-level features in a computationally feasible way.
I searched for "face detection" and got "face recognition" videos. I felt like a linear model would have been more useful.
A related problem is "one shot learning", [1].
https://github.com/Newmu/dcgan_code/raw/master/images/faces_...
The other widely reported vector algebras in a semantic space were discovered by Mikolov et al when producing ~300 dimensional vectors for a billion word Wikipedia corpus.
If one performs vector algebra and ~= is near by cosine distance then using Mikolov's Vectors[3].
King - Man + Woman ~= Queen
France - Paris + Gernmany ~= Berlin
Karpathy shows trained nets can be assembled like lego across modalities, slice off the classifier to reveal the rich semantic 'thought vector' layer of an Imagenet trained Alexnet, plug in a RNN sentence generator using word2vec and ( some over simplification ... ) you get a convincing image captioner [6].
The thought vectors are akin to high level representations of the world and can cross modalities . Text to Images using thought Vectors ( from hnnews discussion [7] )
So the vectors of though are in some way a an AI mentalese or encoding of a symbolic representation of the world derived from the data and can ( again drastic over simplification ) transfer modalities and even between previously unlinked languages [8]
Also see Anything2Vec https://gab41.lab41.org/anything2vec-e99ec0dc186
[1] https://youtu.be/izrG86jycck?t=25m58s
[2] The paper Geoff Hinton is reffering to : Sequence to Sequence Learning with Neural Networks by Ilya Sutskever, Oriol Vinyals, Quoc V. Le https://arxiv.org/abs/1409.3215
[3] Efficient Estimation of Word Representations in Vector Space by Tomas Mikolov, Kai Chen, Greg Corrado, Jeffrey Dean https://arxiv.org/abs/1301.3781
[4] Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks Alec Radford, Luke Metz, Soumith Chintala https://arxiv.org/abs/1511.06434
[5] https://code.facebook.com/posts/1587249151575490/a-path-to-u...
[6] SF Machine Learning: Automated Image Captioning with ConvNets and Recurrent Nets by Karpathy https://youtu.be/ZkY7fAoaNcg?t=38m31s
[7] https://news.ycombinator.com/item?id=12366684
[8] https://github.com/Babylonpartners/fastText_multilingual
https://arxiv.org/abs/1601.01705v4 (Learning to Compose Neural Networks for Question Answering) comes close to breaking this barrier.
"One of the strongest reasons for drawing the line of separation clearly and broadly between science and art is the following:—That the principle of classification in science most conveniently follows the classification of causes, while arts must necessarily be classified according to the classification of the effects, the production of which is their appropriate end."
Essays on some unsettled Questions of Political Economy
http://www.gutenberg.org/ebooks/12004?msg=welcome_stranger#E...
Deep Learning is finding associated effects. It does not find the underlying causes. It is a mode of technical rather than scientific advance.
For instance, in cases where deep neural networks aren't desirable or don't outperform classical approaches, I'm a big fan of boosted decision trees, due to their accuracy on many real-world datasets, their ease-of-use, and the existence of great open source implementations. xgboost (which routinely wins Kaggle competitions) and Spark MLLib both have high-performance distributed training algorithms for gradient boosted trees. And as far as hyperparameter searches go, there just aren't as many parameters to optimize. (And frameworks like Spark are already fantastic for embarrassingly parallel tasks like hyperparameter searches.)
Well it exists, but I wouldn't describe it as high performance in either accuracy or speed.
I'm a big fan of Spark, but Spark ML needs some love from people who actually use it.
Until that happens, just use XGB (which now has Spark integration[1])
[1] http://dmlc.ml/2016/03/14/xgboost4j-portable-distributed-xgb...
Basically, my point is that model interpretation strongly depends on how accurate your model is, and because deep learning models are so much better than linear models for some tasks, it makes sense to use them - even if your primary goal is interpretability.
That said, I do believe that if you ever care at all about interpretation, you should almost never be using multilayer perceptrons (which have recently become part of the widening umbrella term "deep learning"), because they rarely work better than decision tree models or basic linear models (and MLPs are generally less or equally as interpretable when compared to traditional methods).
DeepLIFT (the method I linked in my original comment: https://github.com/kundajelab/deeplift), takes a Keras model (with Theano or TensorFlow backend) as input and provides feature importance scores for any desired layer of the network (raw data inputs, inputs to dense layers following convolution, etc.). Keras-Vis (https://github.com/raghakot/keras-vis) is another nice package that allows for easy visualization of saliency maps and convolutional filters. Perturbing inputs and looking at the effect on the output of the network is another technique people use pretty often.
I think there's a lot of room for this space to become easier to use, especially for newer deep learning practitioners. To that point, I definitely agree with the author of this blog post.
Random Forests can give feature importance, but that does not account for interactions between features. So, in the end, you don't know how a model made a decision (it could be because there is a feature with high importance, but it could also be because there is an informative interaction between lower importance features).
If you want to compare deep learning with linear models, you should leave image data out of it. Compare them on structured data and bag of words.
MLP's and boosted decision trees, in my experience, definitely beat decision tree and linear models, on structured data. But they lack longterm robustness (complex forecasting models need constant retraining, which can hamper their adoption by business units) and don't pass regulation (it is not enough to say "has_asthma" is a high-importance feature).
In finance and health care, interpretability is enormously valued. It is a constant trade-off between accuracy and interpretability.
A long time ago, Caruana made hospital triage models, with neural networks being the clear winner in generalization performance. Instead, they opted for a simple logistic regression when productionizing. Why?
> [...] patients with pneumonia who have a history of asthma have lower risk of dying from pneumonia than the general population. Needless to say, this rule is counterintuitive. But it reflected a true pattern in the training data: patients with a history of asthma who presented with pneumonia usually were admitted not only to the hospital but directly to the ICU (Intensive Care Unit). The good news is that the aggressive care received by asthmatic pneumonia patients was so effective that it lowered their risk of dying from pneumonia compared to the general population. The bad news is that because the prognosis for these patients is better than average, models trained on the data incorrectly learn that asthma lowers risk, when in fact asthmatics have much higher risk (if not hospitalized).
http://people.dbmi.columbia.edu/noemie/papers/15kdd.pdf
Though there is nothing holding you back from using both simple linear, and complex non-linear models at the same time: Only when the models severely disagree do you pick the interpretable model. Or use the linear model to find data issues, like those mentioned above, that are tremendously obscured (if not impossible to identify) when only using deep learning in a train-test framework.
The model learned correctly in this scenario. If you go to the hospital for pneumonia it is apparently in your best interest to claim a history of asthma.
So what they wanted to know is the POD|"No hospital" but they clearly collected data about POD|"Hospital" (since it included ICU admission, etc).
The problem is they measured the wrong thing and then misinterpreted their results. Worse, it looks like the study was designed to be this way!
I see that it has also gotten mainstream news coverage as some kind of lesson about the dangers of machine learning. The real problem is they didn't have data that could answer the question they had, P(Death|No hospitalization), so instead they fit models to answer a different question, P(Death|Hospitalization).
Then they didn't like that the complex models answered the second question too well, so they used simpler ones that made it easier to manually filter out any results that didn't make sense as answers to the first question (which isn't one they could answer to begin with).
No model they fit is safe. You could only use one limited to domains where P(Death|No hospitalization) ~ P(Death|Hospitalization), which isn't something they assessed.
As a Masters student now training deep models for a little while now, I think this point is underemphasized. Doing something novel (so, not just image classification) requires a TON of engineering, not to mention the research considerations. And there are so many tiny decisions and hyperparameters, that even when I thought I had considerable domain knowledge I found it very lacking. I guess it should not be surprising given that 'Deep Learning' refers to a very broad set of models only related by having a learned hierarchical representation. There are a few problems where you can use existing deep learning almost off the shelf (most notably image classification, segmentation), but for most applications I think we're not there yet. As long as this remains true (which I suspect will be for a long time), SVMs and decision trees and linear models are still definitely worth knowing and understanding.
I mean if you can do it for small data and it was good then we would be seeing it dominate kaggle in all problem domains. Maybe the small data problems belong to other algorithm (such as tree base and forest, SVM).
disclaimer - I'm bias for tree base algorithm in medium and small data since it is my thesis.