Outrageously Large Neural Networks: The Sparsely-Gated Mixture-Of-Experts Layer
arxiv.org
arxiv.org
After AlphaGo, I tasked myself with creating a neural network that would use Q-Learning to play Reversi (aka Othello).
At that point, I had already utilized Q-Learning (the tabular version, not using a neural network) for some very simple and mostly proof-of-concept projects, so I understood how it worked. I read up only perceptrons, relu, the benefits/disadvantages of having more/fewer layers, etc.
Then I actually started on the project, thinking "I know about Q-Learning, I know about neural networks, now I just need to use Keras and I'll have a network ready to learn in about twenty lines of python."
Boy was that naive. Regardless of how much you understand the CONCEPTS of neural networks, actually putting together an effective one that matches the problem state perfectly is so, so difficult (especially if there are no examples to build off of). How many layers? Dropout or no, and if so, how much? Do you flatten this layer, do you use relu, do you need a SECOND neural network to approximate one part of the q-function and another to approximate a different part?
I spent MONTHS messing with the hyperparameters, and got nowhere because I'm doing this on a desktop pc without CUDA, so it takes days to train a new configuration only to find out it hardly "learned" anything.
At one point after days of training, my agent actually had a 90% LOSE rate against an opponent that played totally randomly. To this day I am baffled by this.
I went into the project thinking "I have this working with a table, the q-learning part is in place -- just need to drop in a neural net in place of the table and I'm good to go!" It's been almost a year and I still haven't figured this thing out.
That’s not easily paid.
Also, how cost effective is it to use cloud GPUs for real world machine learning?
For my needs I have a machine with a 2011-v3 socket, and four GTX1080 GPUs. Warms up my man cave pretty nicely in winter. I also have access to about a hundred GPUs (older Titans, Teslas, newer 1080s and Pascal Titans) at work that I share with others.
Now, regarding Titan. Titan is actually not that much faster than GTX1080, so in terms of raw speed there's no reason to pay twice as much. BUT, it has 4GB more RAM, which lets you run larger models. NVIDIA rightly decided that for a $100+/hr deep learning researcher $600 is not going to be that big of a deal, and priced the card accordingly. If your models fit into 8GB, you'll be better off buying two 1080's instead.
As to me, I'm thinking of replacing at least one of my 1080s with a Titan, to be able to train larger models.
On a purely TFLOP or even TFLOP/watt basis, it doesn't currently make sense to buy anything that doesn't run Pascal.
A 4GB GTX1050 is ~$180. A p2 instance on Amazon is $0.9/hour. The cost effectiveness depends on if you have a PC already.
[1] https://blog.keras.io/building-powerful-image-classification...
TPD Specs: GTX 1050 75W GTX 1080 180W Titan 250W
Triple these numbers for the overall machine's PSU spec.
So even the Titan costs $0.03 per hour to run, or maybe $0.10 if the rest of the machine is flat-out.
edit: $0.19 kWh for San Francisco residents. Domestic rates.
Nice! You should have just added a bit at the end to invert whatever answer it got, and you would have had a winner.
But more seriously, I think that we will become more clever with designing genetic algorithms to evolve the neural networks as part of the training process rather than trying to build our own from scratch every time. I vaguely recall there is some research being done on that front already.
That's the joke, but to be fair, inversion is not a binary concept. Negation is binary, but inversion is more general and has the 2D interpretation of reflecting something across X=Y.
2. Almost always,.5 is usually good, but you can tune this.
2. You usually flatten right before the final layers.
3. Yes, relu is preferred.
4. You're referring to double q networks, although this helps, optimality bounds are even better.
You'll learn eventually, don't worry. Cheers and welcome to deep learning! ;)
The newer trend appears to be fully-convolutional networks even for classification, since they appear to overfit less, compared to flattening+dropout.
In the end, it was a reduction of the training rate (with SGD) that made things work in what felt like magic.
I've started reading the deep learning text book from Ian Goodfellow now (http://www.deeplearningbook.org). Hoping a solid foundation will build some intuitions for reasoning about these hyper parameters.
Just curious is this offer for other people also? I am working on this https://github.com/deependersingla/deep_portfolio (first version open sourced) and before open-sourced this https://github.com/deependersingla/deep_trader (255 stars). I was using GTX 980-TI locally but system crashed two weeks ago. I am basically trying to find a RL agent which automatically optimize momentum strategy look_back period and generate portfolio according to optimisation one set on reward function. More can be read on project Readme. Thanks.
Also, what we know about good CNN design from supervised learning land doesn't seem to apply to reinforcement learning land, because you're mostly bottlenecked by credit assignment / supervision bitrate, not by a lack of a powerful representation. Your ResNets, batchnorms, or very deep networks have no power here.
SL wants to work. Even if you screw something up you'll usually get something non-random back. RL must be forced to work. If you screw something up or don't tune something well enough you're exceedingly likely to get a policy that is even worse than random. And even if it's all well tuned you'll get a bad policy 30% of the time, just because.
Long story short your failure is more due to the difficulty of deep RL, and much less due to the difficulty of "designing neural networks".
Also listen to karpathy, he knows his stuff, rather than random me :D
RL is about learning expected-reward-maximizing policies for environments that you get to interact with. Common benchmarks currently mostly include games (e.g. ATARI, AlphaGo, VizDoom), physics-based animation (e.g. https://www.cs.ubc.ca/~van/papers/2016-TOG-deepRL/index.html), or (simulated) robotics-like tasks (e.g. MuJoCo). But the core algorithms (such as policy gradients) can be used more generally in settings that don't necessarily look like environments as usual, but where you want to train a network with stochastic nodes, such as in hard attention, etc.
RL is a funny area; A lot of AI researchers get excited about it (mostly motivated by its promise as the formalism that leads to AGI), and yet despite the hype it has so far had very little impact in the industry so far (the Google data center application possibly being an exception, though it was more "RL" than RL, with quotes). It has some promise for Robotics in the real world, but not applied directly and naively. The way that will play out is likely through behavior cloning on human demonstrations or on outputs of trajectory optimizers from simulation, or possibly RL fine-tuning in simulation transferred to real world. But it's still quite early to tell.
On this topic, fun story, the most impressive robots I'm aware of right now are from Boston Dynamics and as they mentioned at this year's NIPS they use ZERO machine learning. Forget deep learning or even deep reinforcement learning. Zero Machine Learning.
I gave a talk last week about some of our RL experiments @ OpenAI and someone came to me after the talk, described their (straight forward) supervised learning problem and asked me how they can apply RL to it. This, to me, is an alarming sign of damaging hype to the community. You don't use RL for your SL problems. You can if you really want to (e.g. reward = 1.0 if you guess the correct label or -1.0 otherwise), but you really don't want to. You're lucky, use your labels, business as usual.
On the other hand, maybe evolution of a nervous system is analogous to some form of neural learning, but on a huge timescale, and maybe the scale is proportional to the searchspace complexity.
edit: Well, reflexes are also propagated by neurons, but on a short circuit and thus readily modeld by PDI etc. Noticing the similar (?) reliance on differentials and coefficients, the advantage is that a subset of PDI has perfect solutions.
However, any time you have a reward signal (like the score of the game) in a multi-step decision problem, like a game where you take actions sequentially (e.g. once per turn), you need RL machinery to make sense of the data. Maybe you take an action now, and you might only reap the reward of that action in the future. So how do you "label" the action right now? You label it with some measure that takes into account the future of the reward signal.
So some human plays a game and gets a super high score. You only see that they got a high score at the end of the game. How do you go back and label the 150 actions that led you to the score? That's the part that is RL.
If this is roughly what it is, it sounds like a logical place to apply some kind of automated meta-programming. I recently stumbled across a Python ML framework that I think used Jinja2 templates for this purpose...
Reinforcement learning's goal is either to estimate each action's expected utility (possibly using neuron networks), or to directly learn what the best action to take is in any given situation, without bothering with utility estimation.
Classification/supervised learning is essentially about learning labels. We have some examples that have already been labeled: cats vs. dogs, suspicious transactions vs legitimate ones, As vs Bs vs...Zs. From those examples, we want to learn some way to assign new, unlabeled instances, to one of those classes.
Reinforcement learning, in contrast, is fundamentally about learning how to behave. Agents learn by interacting with their environment: some combinations of states and actions eventually lead to a reward (which the agent "likes") and others do not. The reward might even be disconnected from the most recent state or action and instead depend on decisions made earlier. The goal is to learn a "policy" that describes what should be done in each state and balances learning more about the environment ("exploration", which may pay off by letting us collect more rewards later) and using what we know about the environment to maximize our current reward intake (exploitation). Games are a particularly good test-bed for reinforcement learning because they have fairly clear states (I have these cards in my hand, or that many lives, etc), actions ("hit me!", "Jump up") and rewards (winnings, scores, levels completed). There's also an obvious parallel with animal behavior, which is where the name originated.
In both cases, neural networks are useful because they are universal function approximators. There's presumably some very complex function that maps data onto labels (e.g., pixels onto {"DOG", "CAT"}) for supervised learning, and states onto action sequences for reinforcement learning. However, we usually don't know what that is, and can't fit it directly, so we let neural networks learn it instead. However, you can do both supervised learning and reinforcement learning without them (in fact, until recently, nearly everyone did).
However, the network typically doesn't get "rewritten" on the fly. Instead, it does something like estimate the value of a state or state-action pair.
(I've obviously gone wrong somewhere here... Just walking you through my thought process)
The agent "learns" by stumbling around and interacting with its environment. At the beginning, its behavior is pretty random, but as it learns more and more, it refines its "policy" to collect more rewards more quickly.
Brute force is certainly possible for some situations. For example, suppose you're playing Blackjack. You can calculate the expected return from 'hitting' (taking another card) and 'standing' (keeping what you've got), based on the cards in your hand and the card the dealer shows.
So...brute force works for simple tasks, but in a lot of situations, it's hard to enumerate all possible states (chess has something like 10^47 possible states) and state-action pairs. It's also difficult to "assign credit"--you rarely lose a chess game just because of the last move. These make it difficult to brute-force a solution or find one via analysis. However, the biggest "win" for using RL is that it's applicable to "black box" scenarios where we don't necessarily know everything about the task. The programmer just needs to give it feedback (though the reward signal) when it does something good or bad.
Furthermore, depending on how you configure the RL agent, it can react to changes in the environment, even without being explicitly reset. For example, imagine a robot vacuum that gets "rewarded" for collecting dirt. It's possible that cleaning the room changes how people use it and thus, changes the distribution of dirt. With the right discounting setup, the vacuum will adjust its behavior accordingly.
The obvious heuristic in the endgame - maximize the number of your pieces - is actually reversed in the opening and midgame, where you want to minimize your number of pieces (other considerations like edge play and parity being equal). So it's possible your NN managed to do worse than random because it learned an endgame heuristic and generalized it improperly.
If so, you may want to consider explicitly wiring in the turn counter as an input.
This is indeed baffling. My best guess: perhaps your error function was reversed? Getting to 90% loss sounds like it would require training.
I'm training a sequence-to-sequence model and have been tuning hyperparameters for the last 2-3 months. I'm making progress, but painfully slowly due to the large time it takes to train and test models (I have a local Titan X and some Tesla K80's in a remote cluster, to which I can send models expecting a latency of 3-4 days of queue and a throughput of around 4 models running simultaneously on average - probably more than many people can get, but still feels slow for this purpose) and the fact that hyperparameter optimization seems to be little more that blind guessing with some very rough rules of thumb. The randomness also doesn't help, as running the same model with different random seeds I have noticed that there is huge variance in accuracy. So sometimes I tweak a parameter and get improvements, but who knows if they are significant or just luck with the initialization. I would have to run every experiment with a bunch of seeds to be sure, but that would mean waiting even more for results and my research would be old before I got to state of the art accuracy.
Maybe I'm just not good at it and I'm a bit bitter, but my feeling is that this DL revolution is turning research in my area from a battle of brain power and ingenuity to a battle of GPU power and economic means (in fact my brain doesn't work much in this research project, as it spends most of the time waiting for results for some GPU - fortunately I have a lot of other non-DL research to do in parallel so the brain doesn't get bored). In the same line, I can't help but notice that most of the top DL NLP papers come from a very select few institutions with huge resources (even though there are heroic exceptions). This doesn't happen as much with non-DL papers.
Good thing that there is still plenty of non-DL research to do, and if DL takes over the whole empirical arena, I'm not bad at theoretical research...
Set all your random seeds to something predefined, such as 42. Even though the exact randomness is OS-specific, this will at least rule out lucky runs from real hyperparameter improvements.
I'm afraid that I cannot go into more specific details right now, but you can get more stable training and faster convergence with a better initialisation strategy.
Fixing the starting randomness seems to be like that old adage of a man with two clocks doesn't know what time it is so he throws away one of them.
Of course, with better tuning one can obtain better optima in general (that's what the whole field is doing) and it's possible that I'm not applying the best techniques and I would get less randomness if I had a better model. But as far as I understand, even the best models can converge to different local optima depending on initialization.
Getting TS to work was hard. Ended up opting for Keras/Theano, TS would be slower for the same problems CPU only
Then getting all tensor shapes to "fit together", to run, then to converge, getting the right non-linearity for each layer, then finding out overfitting is a B! with a capital B
It is hard
Why the broadspread interest in a technology so unpredictable, whose payoff is so little and that requires so much investment in hardware/developer time? Wouldn't you be better off learning tools you can understand and that can be used to build reliable and predictable programs?
Why not warn graduate students:
"You'll be working on this project for about two years. We don't know what to tell you about how to solve the problems involved. We don't have any good general guidelines; this field is changing all the time; nobody knows how these things work except in very broad terms (there is no explanatory power). Sometimes it takes years to find something useful. Luck is the key - if you're unlucky, you're screwed."
"Nothing useful will come of your work when you're done. At completion you'll be dumber than when you started, because you will not have learned anything useful except possibly 'patience', a trait not valued in our field and sometimes viewed as equivalent to 'stupidity' or 'stubbornness'. You will, as a result of working with deep learning/NN, forget that sometimes you must cut your losses and quit exploring a particular solution path. At the end, you _may_ get a Master's degree, if you can show you made a good effort or even some progress. Three months after graduation, everything you have done will be obsolete and no firm will hire you: those familiar with NN because your knowledge is obsolete; those not in NN because they see no particular value in your training."
Can someone explain whether this solves the problem of identifying multiple pathologies within the same image? For example, if a patient has pneumonia and a pleural effusion, could this method activate two subnetworks and come up with a consensus diagnosis based on both those networks? Or does only one pathway get activated at a time?
Especially interesting given that most Kaggle competitions are won with mixture of experts models.
Does mixture-of-experts works well the other way around, as a way to minimize power and hardware in common sized problems ?
And would it work in low resolution networks, like BinaryConnect ?