Fundamental problem with deep learning and neural networks in general.

  1. The solutions that fit training data are infinite. We don't have precise mathematical equation that is satisfied by only a single one and that we can say generalizes best. Simply speaking we don't know which generalizes best.

  2. Optimizing weights is not a convex problem, so we never know we end up with a global or a local minimum.

So why not just dump the neural networks and instead search for a better ML model? Something that we understand, and something that is consistent with a set of mathematical equations? Linear and SVM do not have this mathematical drawbacks and are fully consistent with a a set of mathematical equations. Why not just think on same lines (need not be linear though) and come up with a new ML model better than Linear and SVM and neural networks and deep learning?

  • 40
    $\begingroup$ If you find it, people will. $\endgroup$ Aug 11 '17 at 3:08
  • 24
    $\begingroup$ "Why not come up with...?" You wouldn't believe how many researchers are busy trying to do exactly that! They just haven't had success so far. $\endgroup$ Aug 11 '17 at 6:13
  • 33
    $\begingroup$ "All models are wrong but some are useful" and nns are certainly useful. $\endgroup$
    – josh
    Aug 11 '17 at 8:05
  • 15
    $\begingroup$ @RajeshDachiraju - it is a old idiom, but I was perhaps a bit vague. You asked why not throw away NNs because they are not perfect. My retort is that they are not perfect, but they are USEFUL. People use them to autodrive cars, translate foreign languages, tag videos, in conservation of whales and even to apply those rubbishy snapchat filters with dog ears to your photos! e.g. they work, so we continue to use them :) $\endgroup$
    – josh
    Aug 11 '17 at 10:40
  • 13
    $\begingroup$ You know what is wrong as well: Newtonian mechanics. Quantum Mechanics. Relativity. All the physics is wrong (there is not one single model describing everything, all have their flaws). Chemistry is completely wrong with so many things (describing an atom is always just a good approximation but never exact). The only exactly true thing in the world is math. Pure math. Everything else comes close to the right answer. Should we throw away the rest? (starting from your computer built with wrong laws?). No. Again: all models are wrong, but some are useful. $\endgroup$
    – Mayou36
    Aug 11 '17 at 12:09

10 Answers 10

  1. Not being able to know what solution generalizes best is an issue, but it shouldn't deter us from otherwise using a good solution. Humans themselves often do not known what generalizes best (consider, for example, competing unifying theories of physics), but that doesn't cause us too many problems.

  2. It has been shown that it is extremely rare for training to fail because of local minimums. Most of the local minimums in a deep neural network are close in value to the global minimum, so this is not an issue. source

But the broader answer is that you can talk all day about nonconvexity and model selection, and people will still use neural networks simply because they work better than anything else (at least on things like image classification).

Of course there are also people arguing that we shouldn't get too focused on CNNs like the community was focused on SVMs a few decades ago, and instead keep looking for the next big thing. In particular, I think I remember Hinton regretting the effectiveness of CNNs as something which might hinder research. related post

  • 1
    $\begingroup$ I like the last paragraph particularly. $\endgroup$
    – Rajesh D
    Aug 11 '17 at 4:25
  • 10
    $\begingroup$ Do you have a citation for point #2? $\endgroup$
    – DrMcCleod
    Aug 11 '17 at 11:20
  • $\begingroup$ @DrMcCleod : to me point 2 looks more like jingoism. Just in a lighter sense. $\endgroup$
    – Rajesh D
    Aug 11 '17 at 13:17
  • 6
    $\begingroup$ @DrMcCleod there's a lot of work that suggests that local minima are very close to global minima and that saddle points instead are the issue. See this paper for a discussion of saddle points and this paper for why local minima aren't necessarily bad. $\endgroup$
    – jld
    Aug 11 '17 at 14:13
  • 1
    $\begingroup$ I would prefer only one theatre, I expect. But suppose I know that I will enjoy pretty much any movie almost as much as the one movie I really want to watch. Then I will not be disappointed when there are 10 theatres and I have to pick one at random, because I know any theatre and movie will leave me satisfied. $\endgroup$
    – shimao
    Sep 7 '17 at 6:16

As the comments to your question point out, there are a lot of people working on finding something better. I would though like to answer this question by expanding the comment left by @josh

All models are wrong but some are useful (Wiki)

The above statement is a general truth used to describe the nature of statistical models. Using data that we have available, we can create models that let us do useful things such as approximate a predicted value.

Take for example Linear Regression

Using a number of observations, we can fit a model to give us an approximate value for a dependent variable given any value(s) for the independent variable(s).

Burnham, K. P.; Anderson, D. R. (2002), Model Selection and Multimodel > Inference: A Practical Information-Theoretic Approach (2nd ed.):

"A model is a simplification or approximation of reality and hence will not reflect all of reality. ... Box noted that “all models are wrong, but some are useful.” While a model can never be “truth,” a model might be ranked from very useful, to useful, to somewhat useful to, finally, essentially useless."

Deviations from our model (as can be seen in the image above) appear random, some observations are below the line and some are above, but our regression line shows a general correlation. Whilst deviations in our model appear random, in realistic scenarios there will be other factors at play which cause this deviation. For example, imagine watching cars as they drove through a junction where they must turn either left or right to continue, the cars turn in no particular pattern. Whilst we could say that the direction the cars turn is completely random, does every driver reach the junction and at that point make a random decision of which way to turn? In reality they are probably heading somewhere specific for a specific reason, and without attempting to stop each car to ask them about their reasoning, we can only describe their actions as random.

Where we are able to fit a model with minimal deviation, how certain can we be that an unknown, unnoticed or immeasurable variable wont at some point throw our model? Does the flap of a butterfly’s wings in Brazil set off a tornado in Texas?

The problem with using the Linear and SVN models you mention alone is that we are somewhat required to manually observe our variables and how they each affect each other. We then need to decide what variables are important and write a task-specific algorithm. This can be straight forward if we only have a few variables, but what if we had thousands? What if we wanted to create a generalised image recognition model, could this realistically be achieved with this approach?

Deep Learning and Artificial Neural Networks (ANNs) can help us create useful models for huge data sets containing huge amounts of variables (e.g. image libraries). As you mention, there's an incomprehensible number of solutions which could fit the data using ANNs, but is this number really any different to the amount of solutions we would need to develop ourselves through trial and error?

The application of ANNs do much of the work for us, we can specify our inputs and our desired outputs (and tweak them later to make improvements) and leave it up to the ANN to figure out the solution. This is why ANNs are often described as "black boxes". From a given input they output an approximation, however (in general terms) these approximations don't include details on how they were approximated.

And so it really comes down to what problem you are trying to solve, as the problem will dictate what model approach is more useful. Models are not absolutely accurate and so there is always an element of being 'wrong', however the more accurate your results the more useful they are. Having more detail in the results on how the approximation was made may also be useful, depending on the problem it may even be more useful than increased accuracy.

If for example you are calculating a persons credit score, using regression and SVMs provides calculations that can be better explored. Being able to both tweak the model directly and explain to customers the effect separate independent variables have on their overall score is very useful. An ANN may aid in processing larger amounts of variables to achieve a more accurate score, but would this accuracy be more useful?

  • 6
    $\begingroup$ You make some good points, but the fact that "in many cases our observations and predictions will not sit exactly on the fitted line" is not an apt demonstration of the "all models are wrong" slogan. In linear regression we are modelling E(Y|X) and thus points not lying exactly on the line do not demonstrate a deficiency in our model. Randomness is prespecified and expected; the model is not "wrong" when we observe deviations from the fitted line. $\endgroup$
    – klumbard
    Aug 11 '17 at 12:48
  • $\begingroup$ @klumbard Thanks for the comment. I have updated my answer with more detail which explains my reasoning behind using this as an example. I took a more philosophical approach in my answer and spoke in more general terms rather than specifics, this is my first post in this community so apologies if this is not the place to do so. You seem knowledgeable about the specifics, could you elaborate on your comment a bit more? The question I have is, where deviations do not demonstrate deficiency, is a regression model with an R-squared of 0.01 also not "wrong"? $\endgroup$
    – Carrosive
    Aug 11 '17 at 14:57
  • 2
    $\begingroup$ My only issue with your post is the way you phrase "...as in many cases our observations and predictions will not sit exactly on the fitted line. This is one way in which our model is often 'wrong'...". I'm simply saying that the specification of the model includes an error term and so the fact (alone) that the observed data do not fall on the fitted line does not indicate model "wrongness". This might seem like a subtle semantic distinction but I think it's important $\endgroup$
    – klumbard
    Aug 11 '17 at 15:16
  • 1
    $\begingroup$ The salient point, which you do address, is that all models are wrong because of omitted variable bias as well as misspecification of the functional form. Every time you write down a regression model and perform inference on the estimates, you are assuming you have correctly specified the model, which is never the case. $\endgroup$
    – klumbard
    Aug 11 '17 at 15:17
  • 1
    $\begingroup$ @klumbard Oh I can see where you are coming from now. So although the model produces estimates that are unlikely to be completely accurate, we can measure the error term to state how much the real values may deviate from the estimates, and thus it would be incorrect to say that the model is inherently wrong. I'll take that part out of my answer, I think my point is better explained in the part I added after it. Thanks for explaining :) $\endgroup$
    – Carrosive
    Aug 11 '17 at 15:55

The global minimum may as well as be useless, so we don't really care if we find it or not. The reason is that, for deep networks, not only the time to find it becomes exponentially longer as the network size increases, but also the global minimum often corresponds to overfitting the training set. Thus the generalization ability of the DNN (which is what we really care about) would suffer. Also, often we prefer flatter minima corresponding to a higher value of the loss function, than sharper minima corresponding to a lower value of the loss function, because the second one will deal very badly with uncertainty in the inputs. This is becoming increasingly clear with the development of Bayesian Deep Learning. Robust Optimization beats Determinist Optimization very often, when applied to real world problems where uncertainty is important.

Finally, it's a fact that DNNs just kick the ass of methods such as XGBoost at image classification and NLP. A company which must make a profit out of image classification will correctly select them as models to be deployed in production (and invest a significant amount of money on feature engineering, data pipeline, etc. but I digress). This doesn't mean that they dominate all the ML environment: for example, they do worse than XGBoost on structured data (see the last winners of Kaggle competitions) and they seem to not still do as well as particle filters on time series modelling. However, some very recent innovations on RNNs may modify this situation.

  • 2
    $\begingroup$ Really? A downvote? That's a bit uncalled for. It is reasonable answer (+1). $\endgroup$
    – usεr11852
    Aug 20 '17 at 20:48
  • 5
    $\begingroup$ @RajeshDachiraju since you're apparently trying to infer what I would or would not be aware of, you would probably be interested in learning that people with considerable more understanding of neural networks and non-convex optimization that you seem to have, routinely talk about a single global minimum for neural networks. Among the huge pile of papers using this terminology, you could try reading this one and see if you understand where you're wrong. $\endgroup$
    – DeltaIV
    Oct 31 '17 at 15:42
  • 2
    $\begingroup$ @RajeshDachiraju: Thank you for explaining your reasoning, many people would just not bother. That being said, I think your reasoning for this is flawed and stems from misinterpreting a very particular phrase. I agree with DeltaIV that this standard terminology. $\endgroup$
    – usεr11852
    Oct 31 '17 at 22:09
  • 1
    $\begingroup$ @DeltaIV : My point is, there could be multiple weight vectors that have 0 loss on training data (ofcourse keeping architecture constant). Whole point of training is to obtain weight vector inst it? So I disagree with you. One of these weight vectors is extremely useful. But I request lets agree to disagree and end this conversation here. Regards Rajesh $\endgroup$
    – Rajesh D
    Nov 1 '17 at 12:04
  • 1
    $\begingroup$ @RajeshDachiraju yes, there may be, and they're all equally useless/uninteresting, because they correspond to overfitting the training set $\to$ very low generalization power, if any at all. I really suggest that you read the paper I linked to before, which explains very nicely why when using NN we really don't care about the global minimum value on the training set of the loss function. Also some material on preventing overfitting in NNs may be useful. $\endgroup$
    – DeltaIV
    Nov 1 '17 at 12:09

I think the best way to think about this question is through the competitive market place. If you dump deep learning, and your competitors use it, AND it happens to work better than what you used, then you'll be beaten on the market place.

I think that's what's happening, in part, today, i.e. deep learning seems to work better than anything for the whole lot of problems on market place. For instance, online language translators using deep learning are better than the purely linguistic approaches that were used before. Just a few years ago this was not the case, but advances in deep learning brought those who used to the leadership positions on the market.

I keep repeating "the market" because that's what's driving the current surge in deep learning. The moment business finds something useful, that something will become wide spread. It's not that we, the committee, that decided that deep learning should be popular. It's business and competition.

The second part, is that in addition to actual success of ML, there's also fear to miss the boat. A lot of businesses are paranoid that if they miss out on AI, they'll fail as businesses. This fear is being fed by all these consulting houses, Gartners etc., whispering to CEOs that they must do AI or die tomorrow.

Nobody's forcing businesses to use deep learning. IT and R&D are excited with a new toy. Academia's cheering, so this party's going to last until the music stops, i.e. until deep learning stops delivering. In the meantime you can dump it and come up up with a better solution.

  • $\begingroup$ What about academic research funding? Can you please shed some light on it? $\endgroup$
    – Rajesh D
    Oct 31 '17 at 13:36
  • 2
    $\begingroup$ A lot of funding comes from the industry. The professors that get most money from the industry are the ones that are most influential in the academia. Universities take away a huge chunk of money that they get from the firms, so they love these professors. If you read this NYT article, you can get an idea of the frenzy in both academia and the industry $\endgroup$
    – Aksakal
    Oct 31 '17 at 13:48
  • $\begingroup$ very good reference to the market (+1): I said the same ("A company which must make a profit out of image classification will correctly select them as models to be deployed in production"). However, I would mildly disagree on the paranoia. It's a fact (not paranoia) that Waymo is poised to beat Tesla, Audi, and another car manufacturer whose name I can't recall now, and this is in large part due to the huge investments of Google in Deep Learning. Audi could have definitely used SIFT and SURF (well-tested computer vision technologies which are in no way related to Deep Learning), if they... $\endgroup$
    – DeltaIV
    Nov 1 '17 at 12:37
  • $\begingroup$ ...wanted. The superiority of DL with respect to SIFT, SURF and other geometry-based methods, when it comes to image classification, is a fact attested by five years of solid academic and industrial research. It's definitely not a panacea (see IBM Watson's failures), and there's some hype, but there are also hard, cold facts. $\endgroup$
    – DeltaIV
    Nov 1 '17 at 12:42
  • 2
    $\begingroup$ @DeltaIV ML definitely works in some applications, but I think that today's wide spread adoption of it is due to paranoia and hype to a large degree. Whether it's working or not CTOs are just going for it. I have friends who had no idea what I was talking about just a year ago, now they say that AI is the future, they're going to start implementations etc. $\endgroup$
    – Aksakal
    Nov 1 '17 at 13:52

There are excellent answers, mostly weighing in with the usefulness of DL and ANNs. But I would like to object the OP in a more fundamental way, since the question already takes for granted the mathematical inconsistency of neural networks.

First of all, there is a mathematical theory behind (most models of) Neural Networks. You could likewise argue that linear regression does not generalize, unless the underlying model is... well, linear. In neural algorithms, a model is assumed (even if not explicitly) and the fitting error is computed. The fact that algorithms are modified with various heuristics does not void the original mathematical support. BTW, local optimization is also a mathematically consistent, let alone useful, theory.

Along this line, if Neural Networks just constitute one class of methods within the whole toolbox of scientists, which is the line that separates Neural Networks from the rest of techniques? In fact, SVMs were once considered a class of NNs and they still appear in the same books. On the other hand, NNs could be regarded as a (nonlinear) regression technique, maybe with some simplification. I agree with the OP that we must search better, well founded, efficient algorithms, regardless you label them as NNs or not.

  • $\begingroup$ Problem with being inconsistent is that, one cannot ask simple questions like, When should one stop training and give up? Also lot of rumours like, 'Dropot', 'weight decay', 'ReLu' and various activations, batch normalization, max pooling, softmax, early stopping, various learning rate schedules and all permutations and combinations of these make the designer always in doubt whether to give up or not at some point. $\endgroup$
    – Rajesh D
    Aug 13 '17 at 0:52
  • 1
    $\begingroup$ @RajeshDachiraju The same could be said on penalty coefficients in exterior point optimization algorithms, or the step size in Runge-Kutta methods. The word "inconsistent" has a precise meaning in science that does not apply here. $\endgroup$
    – Miguel
    Aug 13 '17 at 8:42

I guess for some problem we care less for the mathematical rigor and simplicity but more for its utility, current status is neural network is better in performing certain task like pattern recognition in image processing.


There is a lot in this question. Lets go over what you've wrote one by one.

The solutions that fit training data are infinite. We don't have precise mathematical equation that is satisfied by only a single one and that we can say generalizes best.

The fact that there are infinite many solutions comes from learning problem being an ill-posed problem so there cannot be a single one that generalizes best. Also, by no free lunch theorem whichever method we use cannot guarantee that it is the best across all learning problems.

Simply speaking we don't know which generalizes best.

This statement is not really true. There are theorems on empirical risk minimization by Vapnik & Chervonenkis that connect the number of samples, VC dimension of the learning method and the generalization error. Note, that this only applies for a given dataset. So given a dataset and a learning procedure we know the bounds on generalization. Note that, for different datasets there are no and cannot be single best learning procedure due to no free lunch theorem.

Optimizing weights is not a convex problem, so we never know we end up with a global or a local minimum. So why not just dump the neural networks and instead search for a better ML model?

Here there are few things that you need to keep in mind. Optimizing non-convex problem is not as easy as convex one; that is true. However, the class of learning methods that are convex is limited (linear regression, SVMs) and in practice, they perform worse than the class of non-convex (boosting, CNNs) on a variety of problems. So the crucial part is that in practice neural nets work best. Although there are a number of very important elements that make neural nets work well:

  1. They can be applied on very large datasets due to stochastic gradient descent.
  2. Unlike SVMs, inference with deep nets does not depend on the dataset. This makes neural nets efficient at test time.
  3. With neural nets it is possible to directly control their learning capacity (think of number of parameters) simply by adding more layers or making them bigger. This is crucial since for different datasets you might want bigger or smaller models.

Something that we understand, and something that is consistent with a set of mathematical equations? Linear and SVM do not have this mathematical drawbacks and are fully consistent with a a set of mathematical equations. Why not just think on same lines (need not be linear though) and come up with a new ML model better than Linear and SVM and neural networks and deep learning?

Dumping things that work because of not understanding them is not a great research direction. Making an effort in understanding them is, on the other hand, great research direction. Also, I disagree that neural networks are inconsistent with mathematical equations. They are quite consistent. We know how to optimize them and perform inference.


How about viewing neural networks from an experimental point of view? Just because we created them doesn't mean that we're obligued to understand them intuitively. Or that we're not allowed to play with them in order to have a better grasp of what they're doing.

Here's a couple of thoughts I have on them:

  • Structure: they are hierarchies. They are like trees that share inputs. The roots are the inputs and the leafs are the output layer. The closer the layer is to the outputs, the more relevant it is to them, the greater level of abstraction It contains (it's more about the picture than the pixels).
  • Functionality: they "play" with data, the modus operandi is to experiment with relationships in neurons (weights) until things "click" (the error margin is acceptable).

This is consistent with how we think. It's even consistent with how the scientific method operates. So by cracking neural networks we may also be solving the general question of what knowledge represents.


Don't forget, there is a vast field of research that use LMs, GLM, multilevel modelling. Lately Bayesian techniques and Hamiltonian Monte Carlo(the STAN community is really at the forefront of this) have come of age and a number of problems that are solved by STAN really easily and don't really need NNs or deep nets. Social Science research, Microeconomics are two(large) examples of such fields adopting Stan rapidly.

Stan models are very "readable". The coefficients actually have a posterior distributional interpretation and so do the predictions. The priors are part of the data generating process and don't need to be conjugate to be performant(like gibbs). The model fitting in stan is a delight, it actually tunes the pesky MCMC params automatically pretty darn well and warns you when the exploration is stuck with really nice visualizations.

If you haven't tried it already see awesome stan demos here).

At the end of the day I think people don't talk about this stuff so much because the research in this field and the problems are not so "sexy"/"cool" as with NNs.


What typically happens when there is no mathematical consistency (atleast in this case of neural networks)...when its not giving results as desired, on the test set, your boss will come back and say...Hey why don't you try Drop out (which weights,which layer, how many is your headache as there isn't mathematical way to determine), so after you try and hopefully got a marginal improvement but not the desired, your boss will come back and say, why not try weight decay(what factor?)? and later, why don't you try ReLU or some other activation on some layers, and still not, why not try 'max pooling'? still not, why not try batch normalization, still not, or atleast convergence, but not desired result, Oh you are in a local minimum, try different learning rate schedule, just change the network architecture? and repeat all above in different combinations! Keep it in a loop until you succeed!

On the other hand, when you try a consistent SVM, after convergence, if the result is not good, then okay, the linear kernel we are using is not good enough as the data may not be linear, use a different shaped kernel, try a different shaped kernel if you have any hunch, if still not, just leave it, its a limitation of SVM.

What I am saying is ,the neural networks being so inconsistent, that it is not even wrong! It never accepts its defeat! The engineer/designer takes the burden, in case it does not work as desired.

  • 3
    $\begingroup$ This doesn't seem to me to contain an answer to your own question. Do you think you could edit it to sound less like a rant, and make it clear in what way this explains why neural networks and deep learning may be more useful than an ML model (which seems to be your original question)? $\endgroup$
    – Silverfish
    Aug 12 '17 at 20:02
  • 1
    $\begingroup$ His point is that with SVM we know when we've done as well as we can, but with NNs we cannot know. Arguably, given the ease which DL is fooled, even metrics like error do not tell us how well the model is really doing. $\endgroup$
    – yters
    Aug 12 '17 at 21:25
  • 1
    $\begingroup$ @yters, yes but silverfish's comment was that this isn't an answer to why not dump DL. It's closer to a restatement of the question. I'd suggest merging it with the question. $\endgroup$ Oct 31 '17 at 13:13

Not the answer you're looking for? Browse other questions tagged or ask your own question.