9 ms·
Do neural networks dream of electric sheep?
- minimaxir 9y ago@picdescbot is a Twitter bot which uses the same Microsoft Azure Computer Vision API endpoint for image captioning: https://twitter.com/picdescbot https://twitter.com/picdescbot It's more accurate than you'd think, and as this article notes, the mistakes are indeed funny: https://twitter.com/picdescbot/status/968561437126938625 https://twitter.com/picdescbot/status/968561437126938625
- GuiA 9y agoIt's okay, humans make these kind of mistakes too. A friend of mine has a young son (~2-3 years old), and a cat named Mono. Her son knows the cat is named Mono - he plays with her everyday. But when they go out for a walk, any 4 legged animal he sees is also a "Mono". Fortunately for her son, his developing, extremely plastic brain will soon know how to differentiate Mono the cat from a random dog on the street (unlike neural networks which we will need to entirely redesign, rebuild, and retrain to get similar progress).
- nonbel 9y ago>"unlike neural networks which we will need to entirely redesign, rebuild, and retrain to get similar progress" I don't think you would need to redesign or rebuild anything for that. You would need to train the network on additional examples, which I supposed you could call retraining (although it need not be from scratch), but that is the same in the case of the child.
- dvaun 9y agoYour friend's son is over-extending "Mono" to other similar animals. This is actually a normal part of a child's linguistic development. My son also does the same thing with "dodo" for dinosaur. At first, he applied it only to big, scary dinosaurs that made large sounds—later on, it applied to any animal that had some large body and jaws, e.g. a shark. Finally, he learned to differentiate between the majority of animals (save for a few big, scary-looking animals that are still "dodo") and can name sharks, dinosaurs, and bears separately!
- comex 9y agoSo in other words, "dodo" is (slowly) going extinct from his vocabulary?
- dvaun 9y agoYes. I don't know what the exact term for the process is, though, since I'm not knowledgeable about linguistics and childhood development beyond what I've learned from raising my son. It's very exciting to watch and hear the progress!
- comex 9y agoI don't actually know the terminology, either; I was just making a really bad pun. ;p
- DFHippie 9y agoIt could be that your friend's son has an overly generous sense of what Mono is. It could be that he knows precisely which individual entity is Mono but misunderstands what "Mono" denotes/refers to. It could be somewhere in between. If a neural network has a representation of entities in the world apart from language referring to those entities, that's would be awesome. I'm guessing we're not there yet though.
- _emacsomancer_ 9y agoIt could also just be limited vocabulary and that his outside use of "Mono" means something like "a creature which is somewhat similar to Mono".
- DFHippie 9y agoIt could be he knows who Mono is and he knows "Mono" just refers to Mono, but there are a lot of other things in the world that he wants to talk about and he knows the grownups will be able to piece thing together. I do that often enough. "Fred or whoever, that guy, did that thing ..." If neural networks were doing that, that would be more awesome still.
- sp332 9y agoYeah but it's good to keep in mind what kinds of errors a neural network can make. It will help in training and debugging them. Also it helps to keep a little skepticism now that "deep learning" gets waved at everything like a magic wand.
- John_KZ 9y agoThey're not the same type of mistake. I don't remember the deteails, but when I was reading papers on caption generation, the part of producing coherent sentences seemed more like a hack that happens to kinda work (usually) rather than a robust solution.
- taneq 9y ago> unlike neural networks which we will need to entirely redesign, rebuild, and retrain to get similar progress You could argue that the child's neural network will be redesigned and rebuilt as his brain matures.
- pitchups 9y agoThis example, and others like it point to the central weakness of neural networks for image recognition: No matter how much data you feed it, they never really develop concepts or abstractions of what the objects it is classifying really represent or mean. The weight and biases that get fine tuned by gradient descent, are no more than a highly complex function mapping the input pixels to discrete classes. While this may well represent how the visual cortex works at the lowest level, what appears to be missing are higher levels of abstraction and meaning. Perhaps machine learning needs to be coupled with some of the older paradigms of AI which included modeling, logic, reasoning, to achieve understanding. As of right now, a well trained convolutional neural network is no more than a mechanical pattern matching algorithm on steroids.
- felippee 9y agoThough I agree with the spirit of what you are saying, I would add that it is perhaps not the weakness of neural networks per se, but weakness of the current architectures and training (supervised, hence verbalized) paradigms. I think we could do much more to improve things, if we stopped pushing for extra % on benchmarks and instead rethink the problems we'd like to solve and approach them from a new angle.
- dahart 9y agoThat’s all true. Sometimes I describe NNs as fancy least squares. Interesting things are bound to happen when you have a few hundred million parameters. Just to play devil’s advocate, we don’t know yet if the model is bad or if we’re just feeding the wrong kind of data, right? These optimization algorithms are good at interpolating; they do well with new data points that land inside the multidimensional convex hull of the training data. They can fail spectacularly when the new data to inference is outside that boundary. But humans aren’t that great at extrapolation either. Maybe NNs will be good enough when we show them everything there is, maybe what we think of as conceptual understanding is just as much simple interpolation of our experiences as our neural networks...?
- pitchups 9y ago>maybe what we think of as conceptual understanding is just as much simple interpolation of our experiences as our neural networks...? Possibly - we don't know. Sure if we could train one of these imagenet type classifiers with a million or billion images - close to all known objects in the universe, it may well be able to "recognize" everything. But that still doesn't solve the problem of abstraction or meaning, much less of intuition and generalization. Humans are able to generalize to new domains based on internal models of the world around us. The models used in RNNs for encoding word embeddings, seem to a bit closer to representing meaning. I agree though that NNs are evolving - we are in very early days, and who knows the NNs of the future may reveal that what we consider as understanding is no more than simple interpolation, as you suggest.
- ibdf 9y agoThis was a great read. Thank you!
- wordpressdev 9y agoReminds me of Philip K. Dick's book Do Androids Dream of Electric Sheep? made into the cult movie Blade Runner.
- danbruc 9y agoThat left me a bit disappointed about the current - I assume the tested detectors are at least somewhat state of the art - capabilities of neural network for object identification. But thinking about it also makes it somewhat obvious that they are flawed in this way. Humans identify objects by looking at how different parts are geometrically located and connected, possibly in a hierarchical fashion, and what basic shapes, colors and textures those parts have. A sheep is a body, four legs, hoofs, a tail, a head with its characteristic shape, the ears, mouth, nose and all those come with characteristic textures and colors. And because there are so many features and their relations, it is quite hard to fool humans, you can hide or change quite a few of them. We also have a lot of background knowledge, a bright orange sheep might be unusual but we also have a pretty good idea of how hard it is to change the color of a sheep. I naively expected neural networks to also learn those features but there is just no pressure for them to do so. They mostly see common objects in common situations and there just looking for a patch of wool-colored, wool-textured fur might be enough to identify the presence of sheep correctly almost all the time. Or if sheep are mostly depicted in a characteristic environment it might be good enough to just identify landscape features and ignore the sheep altogether. I would guess that it is in general not really feasible to come up with enough contrived, adversarial examples to force neural networks to learn the important parts and relations of different objects just by starring at many images. I think one would have to hard-wire some knowledge about space, spatial relations, occlusion, shapes and the like into a system to really get it to learn what a sheep is in a similar way as humans do without heavily increasing the risk of over-fitting.
- dahart 9y ago> Humans identify objects by looking at how different parts are geometrically located and connected, possibly in a hierarchical fashion FWIW this is what capsule networks are designed to do.
- notahacker 9y agoAlso, our grasp of sheep morphology is based on an abstract idea that the sheep is qualitatively different to and more significant than the surrounding fields and rocks which are roughly the same size, shape and colour as sheep. Unlike the ML process, this concept of sheep or more unfamiliar mammals as animate creatures which might be friend, foe or food exists independently of how many sheep I've seen before and the language used to classify them, and is probably instinctual rather than learned. Show me a couple of images of fields full of sheep classified as oveja or Schafe and I might make the same learning error as the ML process and think the word refers to the [general pattern of] surrounding field or hills. But show me a further image of oveja outside a field - even a close up of an oveja that doesn't resemble those in field photos in any way - and I'll grasp the meaning of the term straight away. Needless to say I'm also less likely to stumble over conceptual links between the names of animals and tastes of food, types of clothing etc which are independent of the living animal's morphology
- lucideer 9y agoMaybe I'm oversimplifying but this seems like an obviously simple* problem to solve: as humans, before we can identify what objects are, we start out with depth perception and object detection. If NNs were trained on a dataset of imagery where autodetected object outlines were tagged, rather than simply tagging or describing an entire image, and then run with built in object detection and depth perception, I suspect the results would be pretty good. I know the likes of Google and Facebook are already doing precisely this with human faces, but we'd need a more generalised object detection algorithm before the examples of sheep in the article would be reliably identified. * I used the word "simple", as distinct from "easy": for example, creating a training dataset might be a challenge.
- fizx 9y agoLike most obvious things, scientist have tried what is known in the field as "image segmentation". If you google it, you'll see a bunch of demos and papers. I haven't seen anything in the literature about incredibly effective implementations of 3D understanding and/or depth perception at the level you'd perhaps hope, but there is some progress being made.
- jquip 9y agoCategorical divisions and their names are being refined at that age. Which is why the most familiar name is allocated to the entire group of similar entities. Later on, categories become refined and also their namings.