Artificial networks model systems with a set of inputs and outputs and expected behavior. To train a network for modeling such systems, hundreds, thousands, or millions of example inputs-output pairs may be required. This is called a labelled data set, and the network and its optimization algorithm are meant to find a set of network parameters that best match the I/O of the artificial network with the I/O of the system.

Are there any systems, for which sufficient labelled data sets exist, that have yet to be successfully modeled with artificial networks of any type (recurrent, deep, convolution, etc)?

  • $\begingroup$ That's a list of encyclopedic size. Listing what they do well is a much more tractable question. $\endgroup$ – Douglas Daseeco Jul 29 '18 at 9:21
  • $\begingroup$ @FauChristian: name 3 examples of that list is so huge $\endgroup$ – Martin Thoma Sep 27 '18 at 6:01
  • $\begingroup$ @MartinThoma, here are a few among thousands: Arriving at the next word in a patterned series of words, rendering a plot or scene, an ant climbing, comprehension of an arbitrary imperative, vehicle control in an unexpected situation, prime number generation, self-repair, suggesting movies based on customer beliefs, astrodynamic calculation, simulation, a bee collaborating to construct a honeycomb, learning a new voice from sparse input, inference, floating point division, bipedal walking on rock, auto-routing a PC board, bar code decoding, boot up, a fly finding blood, CNC, landing a jet, ... $\endgroup$ – Douglas Daseeco Sep 27 '18 at 22:49
  • 1
    $\begingroup$ Rendering a scene: GANs; Recommendation systems, voice learning. For most of the others: There are no labeled datasets of sufficient size. Most of the problems you describe are reinforcement learning problems. While I agree that the results so far are not super impressive, they are still far better than random. $\endgroup$ – Martin Thoma Sep 28 '18 at 4:54
  • $\begingroup$ Also, some of the task can be approached with ML, but have a perfect solution with non-learning algorithms. I know there are some toy problems which use them to visualize their capabilities, but nobody in their right mind would actually want to apply ML for those problems in production. floating point division is an example for this. $\endgroup$ – Martin Thoma Sep 28 '18 at 9:34

I haven't seen any dataset where some standard models worked and neural networks utterly failed.

For columnar data (e.g. Excel files / database dumps / CSV files) which contain structured data usually tree-based models like random forests and gradient boosting work better, but neural networks are also usually way better than random.

If you demand other things, e.g. explanations for the decision then Bayesian models might give you an easier time. Or for baselines/simple implementations linear models. Or for real time applications...


If i understand your question correctly , you are asking if there exists functional datasets for which there are no proven solutions based on neural networks which give substantial accuracy.

there are many such problems for which we have data in abundancy, question answering would be one such a thing , you still cant devise a neural network architechture that reads through entire principia mathematica and then complete theorems , and point cloud processing is a also a big hurdle for neural networks considering the highly irregular datastructure , even if you voxelize a point-cloud it would be infeasible to train large convolutional networks on it . (there is also rapid progress in this direction ,, point cloud processing).

geoffrey hinton mentioned in an AMA before 3 years that we will see neural networks that will answer questions based on videos in the next five years , but still video-question answering seems to be far away from the present technology.

Graph datasets are also one such area where still neural networks research is in infancy (refer http://www.inference.vc/how-powerful-are-graph-convolutions-review-of-kipf-welling-2016-2/)


I still havent seen a good program for recommending me good movies, songs and books based on my previous lkes. Youtube recommendation system is shit. Since youtube is owned by google and since youtube has most content and since google is investing in AI a lot I would conclude this problem remains largely unsolved.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy