[link]
This paper is to mitigate the scene bias in the action recognition task. Scene bias is defined as the model only focusing on scene or object information without paying attention to the actual activity. To mitigate this issue, the author proposed 2 additional types of loss: (1) scene adversarial loss that helps the network to learn features that are suitable for action but invariant to scene type. Hence, reduce the scene bias. (2) human mask confusion loss that prevents a model from predicting the correct action (label) of this video if there is no person in this video. Hence, this can mitigate the scene bias because the model can not predict the correct action based on only the surrounding scene. https://i.imgur.com/BBfWE17.png To mask out the person in the video, they use a human detector to detect and then mask the person out. In the above diagram, there is a gradient reversal layer, which works as follows: In the forward pass, the output is similar to the input. In the backward pass, the output is equal to the input times 1. https://i.imgur.com/hif9ZL9.png This layer comes from Domain Adaptation. In domain adaptation, there is a need to make the distribution of the source and the target domain distinguishable. So, in this work, they want to make the action distribution and the scene distribution distinguishable, which is why they train the action classifier and scene classifier in an adversarial way. https://i.imgur.com/trNJGlm.png And by using the Gradient reversal layer, for the training instances, the action predictor will be trained for predicting the labels of the training instances. The feature extractor will therefore be trained to minimize the classification loss of the action predictor and maximize the classification loss of the scene predictor. As a result, the action will be sceneagnostic. 
[link]
The typical model based reinforcement learning (RL) loop consists of collecting data, training a model of the environment, using the model to do model predictive control (MPC). If however the model is wrong, for example for stateaction pairs that have been barely visited, the dynamics model might be very wrong and the MPC fails as the imagined model and the reality align to longer. Boney et a. propose to tackle this with a denoising autoencoder for trajectory regularization according to the familiarity of a trajectory. MPC uses at each time t the learned model $s_{t+1} = f_{\theta}(s_t, a_t)$ to select a plan of actions, that is maximizing the sum of expected future reward: $ G(a_t, \dots, a_{t+h}) = \mathbb{E}[\sum_{k=t}^{t+H}r(o_t, a_t)] ,$ where $r(o_t, a_t)$ is the observation and action dependent reward. The plan obtained by trajectory optimization is subsequently unrolled for H steps. Boney et al. propose to regularize trajectories by the familiarity of the visited states leading to the regularized objective: $G_{reg} = G + \alpha \log p(o_k, a_k, \dots, o_{t+H}, a_{t+H}) $ Instead of regularizing over the whole trajectory they propose to regularize over marginal probabilities of windows of length w: $G_{reg} = G + \alpha \sum_{k = t}^{t+Hw} \log p(x_k), \text{ where } x_k = (o_k, a_k, \dots, o_{t+w}, a_{t+w}).$ Instead of explicitly learning a generative model of the familiarity $p(x_k)$ a denoising autoencoder is used that approximates instead the derivative of the log probability density $\frac{\delta}{\delta x} \log p(x)$. This allows the following backpropagation rule: $ \frac{\delta G_{reg}}{\delta_i} = \frac{\delta G}{\delta a_i} + \alpha \sum_{k = i}^{i+w} \frac{\delta x_k}{\delta a_i} \frac{\delta}{\delta x} \log p(x).$ The experiments show that the proposed method has competitive sampleefficiency. For example on Halfcheetah the asymptotic performance of PETS is not matched. This is due to the biggest limitation of this approach, the hindering of exploration. Penalizing the unfamiliarity of states is in contrast to approaches like optimism in the face of uncertainty, which is a core principle of exploration. Aiming to avoid states of high unfamiliarity, the proposed method is the precise opposite of curiosity driven exploration. The appendix proposes preliminary experiments to account for exploration. I would expect, that the pure penalization of unfamiliarity works best in a batch RL setting, which would be an interesting extension of this work. 
[link]
Kumar et al. propose an algorithm to learn in batch reinforcement learning (RL), a setting where an agent learns purely form a fixed batch of data, $B$, without any interactions with the environments. The data in the batch is collected according to a batch policy $\pi_b$. Whereas most previous methods (like BCQ) constrain the learned policy to stay close to the behavior policy, Kumar et al. propose bootstrapping error accumulation reduction (BEAR), which constrains the newly learned policy to place some probability mass on every non negligible action. The difference is illustrated in the picture from the BEAR blog post: https://i.imgur.com/zUw7XNt.png The behavior policy is in both images the dotted red line, the left image shows the policy matching where the algorithm is constrained to the purple choices, while the right image shows the support matching. **Theoretical Contribution:** The paper analysis formally how the use of outofdistribution actions to compute the target in the Bellman equation influences the backpropagated error. Firstly a distribution constrained backup operator is defined as $T^{\Pi}Q(s,a) = \mathbb{E}[R(s,a) + \gamma \max_{\pi \in \Pi} \mathbb{E}_{P(s' \vert s,a)} V(s')]$ and $V(s) = \max_{\pi \in \Pi} \mathbb{E}_{\pi}[Q(s,a)]$ which considers only policies $\pi \in \Pi$. It is possible that the optimal policy $\pi^*$ is not contained in the policy set $\Pi$, thus there is a suboptimallity constant $\alpha (\Pi) = \max_{s,a} \vert \mathcal{T}^{\Pi}Q^{*}(s,a)  \mathcal{T}Q^{*}(s,a) ]\vert $ which captures how far $\pi^{*}$ is from $\Pi$. Letting $P^{\pi_i}$ be the transitionmatrix when following policy $\pi_i$, $\rho_0$ the state marginal distribution of the training data in the batch and $\pi_1, \dots, \pi_k \in \Pi $. The error analysis relies upon a concentrability assumption $\rho_0 P^{\pi_1} \dots P^{\pi_k} \leq c(k)\mu(s)$, with $\mu(s)$ the state marginal. Note that $c(k)$ might be infinite if the support of $\Pi$ is not contained in the state marginal of the batch. Using the coefficients $c(k)$ a concentrability coefficient is defined as: $C(\Pi) = (1\gamma)^2\sum_{k=1}^{\infty}k \gamma^{k1}c(k).$ The concentrability takes values between 1 und $\infty$, where 1 corresponds to the case that the batch data were collected by $\pi$ and $\Pi = \{\pi\}$ and $\infty$ to cases where $\Pi$ has support outside of $\pi$. Combining this Kumar et a. get a bound of the Bellman error for distribution constrained value iteration with the constrained Bellman operator $T^{\Pi}$: $\lim_{k \rightarrow \infty} \mathbb{E}_{\rho_0}[\vert V^{\pi_k}(s) V^{*}(s)] \leq \frac{\gamma}{(1\gamma^2)} [C(\Pi) \mathbb{E}_{\mu}[\max_{\pi \in \Pi}\mathbb{E}_{\pi}[\delta(s,a)] + \frac{1\gamma}{\gamma}\alpha(\Pi) ] ]$, where $\delta(s,a)$ is the Bellman error. This presents the inherent batch RL tradeoff between keeping policies close to the behavior policy of the batch (captured by $C(\Pi)$ and keeping $\Pi$ sufficiently large (captured by $\alpha(\Pi)$). It is finally proposed to use support sets to construct $\Pi$, that is $\Pi_{\epsilon} = \{\pi \vert \pi(a \vert s)=0 \text{ whenever } \beta(a \vert s) < \epsilon \}$. This amounts to the set of all policies that place probability on all nonnegligible actions of the behavior policy. For this particular choice of $\Pi = \Pi_{\epsilon}$ the concentrability coefficient can be bounded. **Algorithm**: The algorithm has an actor critic style, where the Qvalue to update the policy is taken to be the minimum over the ensemble. The support constraint to place at least some probability mass on every non negligible action from the batch is enforced via sampled MMD. The proposed algorithm is a member of the policy regularized algorithms as the policy is updated to optimize: $\pi_{\Phi} = \max_{\pi} \mathbb{E}_{s \sim B} \mathbb{E}_{a \sim \pi(\cdot \vert s)} [min_{j = 1 \dots, k} Q_j(s,a)] s.t. \mathbb{E}_{s \sim B}[MMD(D(s), \pi(\cdot \vert s))] \leq \epsilon$ The Bellman target to update the Qfunctions is computed as the convex combination of minimum and maximum of the ensemble. **Experiments** The experiments use the Mujoco environments Halfcheetah, Walker, Hopper and Ant. Three scenarios of batch collection, always consisting of 1Mio. samples, are considered:  completely random behavior policy  partially trained behavior policy  optimal policy as behavior policy The experiments confirm that BEAR outperforms other offpolicy methods like BCQ or KLcontrol. The ablations show further that the choice of MMD is crucial as it is sometimes on par and sometimes substantially better than choosing KLdivergence. 
[link]
Salman et al. combined randomized smoothing with adversarial training based on an attack specifically designed against smoothed classifiers. Specifically, they consider the formulation of randomized smoothing by Cohen et al. [1]; here, Gaussian noise around the input (adversarial or clean) is sampled and the classifier takes a simple majority vote. In [1], Cohen et al. show that this results in good bounds on robustness. In this paper, Salman et al. propose an adaptive attack against randomized smoothing. Essentially, they use a simple PGD attack to attack a smoothed classifier, i.e., maximize the cross entropy loss of the smoothed classifier. To make the objective tractable, Monte Carlo samples are used in each iteration of the PGD optimization. Based on this attack, they do adversarial training, with adversarial examples computed against the smoothed (and adversarially trained) classifier. In experiments, this approach outperforms the certified robustness by Cohen et al. on several datasets. [1] Jeremy M. Cohen, Elan Rosenfeld and J. Zico Kolter. Certified Adversarial Robustness via Randomized Smoothing. ArXiv, 1902.02918, 2019. Also find this summary at [davidstutz.de](https://davidstutz.de/category/reading/). 
[link]
This paper is a bit provocative (especially in the light of the recent DeepMind MuZero paper), and poses some interesting questions about the value of modelbased planning. I'm not sure I agree with the overall argument it's making, but I think the experience of reading it made me hone my intuitions around why and when modelbased planning should be useful. The overall argument of the paper is: rather than learning a dynamics model of the environment and then using that model to plan and learn a value/policy function from, we could instead just keep a large replay buffer of actual past transitions, and use that in lieu of modelsampled transitions to further update our reward estimators without having to acquire more actual experience. In this paper's framing, the central value of having a learned model is this ability to update our policy without needing more actual experience, and it argues that actual real transitions from the environment are more reliable and less likely to diverge than transitions from a learned parametric model. It basically sees a big buffer of transitions as an empirical environment model that it can sample from, in a roughly equivalent way to being able to sample transitions from a learnt model. An obvious counterargument to this is the value of models in being able to simulate particular arbitrary trajectories (for example, potential actions you could take from your current point, as is needed for Monte Carlo Tree Search). Simply keeping around a big stock of historical transitions doesn't serve the use case of being able to get a probable next state *for a particular transition*, both because we might not have that state in our data, and because we don't have any way, just given a replay buffer, of knowing that an available state comes after an action if we haven't seen that exact combination before. (And, even if we had, we'd have to have some indexing/lookup mechanism atop the data). I didn't feel like the paper's response to this was all that convincing. It basically just argues that planning with model transitions can theoretically diverge (though acknowledges it empirically often doesn't), and that it's dangerous to update off of "fictional" modeled transitions that aren't grounded in real data. While it's obviously definitionally true that model transitions are in some sense fictional, that's just the basic tradeoff of how modeling works: some ability to extrapolate, but a realization that there's a risk you extrapolate poorly. https://i.imgur.com/8jp22M3.png The paper's empirical contribution to its argument was to argue that in a lowdata setting, modelfree RL (in the form of the "everything but the kitchen sink" Rainbow RL algorithm) with experience replay can outperform a modelbased SimPLe system on Atari. This strikes me as fairly weak support for the paper's overall claim, especially since historically Atari has been difficult to learn good models of when they're learnt in actualobservation pixel space. Nonetheless, I think this push against the utility of modelbased learning is a useful thing to consider if you do think models are useful, because it will help clarify the reasons why you think that's the case. 
[link]
In my view, the Lottery Ticket Hypothesis is one of the weirder and more mysterious phenomena of the last few years of Machine Learning. We've known for awhile that we can take trained networks and prune them down to a small fraction of their weights (keeping those weights with the highest magnitudes) and maintain test performance using only those learned weights. That seemed somewhat surprising, in that there were a lot of weights that weren't actually necessary to encoding the learned function, but, the thinking went, possibly having many times more weights than that was helpful for training, even if not necessary once a model is trained. The authors of the original Lottery Ticket paper came to the surprising realization that they could take the weights that were pruned to exist in the final network, reinitialize them (and only them) to the values they had during initial training, and perform almost as well as the final pruned model that had all weights active during training. And, performance using the specific weights and their particular initialization values is much higher than training a comparable topology of weights with random initial values. This paper out of Facebook AI adds another fascinating experiment to the pile of off evidence around lottery tickets: they test whether lottery tickets transfer *between datasets*, and they find that they often do (at least when the dataset on which the lottery ticket is found is more complex (in terms of in size, input complexity, or number of classes) than the dataset the ticket is being transferred to. Even more interestingly, they find that for sufficiently simple datasets, the "ticket" initialization pattern learned on a more complex dataset actually does *better* than ones learned on the simple dataset itself. They also find that tickets by and large transfer between SGD and Adam, so whatever kind of inductive bias or value they provide is general across optimizers in addition to at least partially general across datasets. https://i.imgur.com/H0aPjRN.png I find this result fun to think about through a few frames. The first is to remember that figuring out heuristics for initializing networks (as a function of their topology) was an important step in getting them to train at all, so while this result may at first seem strange and arcane, in that context it feels less surprising that there are stillbetter initialization heuristics out there, possibly with some kind of interesting theoretical justification to them, that humans simply haven't been clever enough to formalize yet, and have only discovered empirically through methods like this. This result is also interesting in terms of transfer: we've known for awhile that the representations learned on more complex datasets can convey general information back to smaller ones, but it's less easy to think about what information is conveyed by the topology and connectivity of a network. This paper suggests that the information is there, and has prompted me to think more about the slightly mindbending question of how training models could lead to information compressed in this form, and how this information could be better understood. 
[link]
VQVAE is a Variational AutoEncoder that uses as its information bottleneck a discrete set of codes, rather than a continuous vector. That is: the encoder creates a downsampled spatial representation of the image, where in each grid cell of the downsampled image, the cell is represented by a vector. But, before that vector is passed to the decoder, it's discretized, by (effectively) clustering the vectors the network has historically seen, and substituting each vector with the center of the vector it's closest to. This has the effect of reducing the capacity of your information bottleneck, but without just pushing your encoded representation closer to an uninformed prior. (If you're wondering how the gradient survives this very much not continuous operation, the answer is: we just pretend that operation didn't exist, and imagine that the encoder produced the clustercenter "codebook" vector that the decoder sees). The part of the model that got a (small) upgrade in this paper is the prior distribution model that's learned on top of these latent representations. The goal of this prior is to be able to just sample images, unprompted, from the distribution of latent codes. Once we have a trained decoder, if we give it a grid of such codes, it can produce an image. But these codes aren't oneperimage, but rather a grid of many codes representing features in different part of the image. In order to generate a set of codes corresponding to a reasonable image, we can either generate them all at once, or else (as this paper does) use an autoregressive approach, where some parts of the code grid are generated, and then subsequent ones conditioned on those. In the original version of the paper, the autoregressive model used was a PixelCNN (don't have the space to fully explain that here, but, at a high level: a model that uses convolutions over previously generated regions to generate a new region). In this paper, the authors took inspiration from the huge rise of selfattention in recent years, and swapped that operation in in lieu of the convolutions. Selfattention has the nice benefit that you can easily have a global receptive range (each region being generated can see all other regions) which you'd otherwise need multiple layers of convolutions to accomplish. In addition, the authors add an additional layer of granularity: generating both a 32x32 and 64x64 grid, and using both to generate the decoded reconstruction. They argue that this allows one representation to focus on more global details, and the other on more precise ones. https://i.imgur.com/zD78Pp4.png The final result is the ability to generate quite realistic looking images, that at least are being claimed to be more diverse than those generated by GANs (examples above). I'm always a bit cautious of claims of better performance in the imagegeneration area, because it's all squinting at pixels and making up somewhatreasonable but still arbitrary metrics. That said, it seems interesting and useful to be aware of the current relative capabilities of two of the main forms of generative modeling, and so I'd recommend this paper on that front, even if it's hard for me personally to confidently assess the improvements on prior art. 
[link]
Coming from the perspective of the rest of machine learning, a somewhat odd thing about reinforcement learning that often goes unnoticed is the fact that, in basically all reinforcement learning, performance of an algorithm is judged by its performance on the same environment it was trained on. In the parlance of ML writ large: training on the test set. In RL, most of the focus has historically been on whether automatic systems would be able to learn a policy from the state distribution of a single environment, already a fairly hard task. But, now that RL has had more success in the singleenvironment case, there comes the question: how can we train reinforcement algorithms that don't just perform well on a single environment, but over a range of environments. One lens onto this question is that of metalearning, but this paper takes a different approach, and looks at how straightforward regularization techniques pulled from the land of supervised learning can (or can't straightforwardly) be applied to reinforcement learning. In general, the regularization techniques discussed here are all ways of reducing the capacity of the model, and preventing it from overfitting. Some ways to reduce capacity are:  Apply L2 weight penalization  Apply dropout, which handicaps the model by randomly zeroing out neurons  Use Batch Norm, which uses noisy batch statistics, and increases randomness in a way that, similar to above, deteriorates performance  Use an information bottleneck: similar to a VAE, this approach works by learning some compressed representation of your input, p(zx), and then predicting your output off of that z, in a way that incentivizes your z to be informative (because you want to be able to predict y well) but also penalizes too much information being put in it (because you penalize differences between your learned p(zx) distribution and an unconditional prior p(z) ). This pushes your model to use its conditionalonx capacity wisely, and only learn features if they're quite valuable in predicting y However, the paper points out that there are some complications in straightforwardly applying these techniques to RL. The central one is the fact that in (most) RL, the distribution of transitions you train on comes from prior iterations of your policy. This means that a noisier and less competent policy will also leave you with less data to train on. Additionally, using a noisy policy can increase variance, both by making your trained policy more different than your rollout policy (in an offpolicy setting) and by making your estimate of the value function highervariance, which is problematic because that's what you're using as a target training signal in a temporal difference framework. The paper is a bit disconnected in its connection between justification and theory, and makes two broad, mostly distinct proposals: 1. The most successful (though also the one least directly justified by the earlierdiscussed theoretical difficulties of applying regularization in RL) is an information bottleneck ported into a RL setting. It works almost the same as the classificationmodel one, except that you're trying to increase the value of your actions given compressedfromstate representation z, rather than trying to increase your ability to correctly predict y. The justification given here is that it's good to incentivize RL algorithms in particular to learn simpler, more compressible features, because they often have such poor data and also training signal earlier in training 2. SNI (Selective Noise Injection) works by only applying stochastic aspects of regularization (sampling from z in an information bottleneck, applying different dropout masks, etc) to certain parts of the training procedure. In particular, the rollout used to collect data is nonstochastic, removing the issue of noisiness impacting the data that's collected. They then do an interesting thing where they calculate a weighted mixture of the policy update with a deterministic model, and the update with a stochastic one. The best performing of these that they tested seems to have been a 50/50 split. This is essentially just a knob you can turn on stochasticity, to trade off between the regularizing effect of noise and the varianceincreasingnegative effect of it. https://i.imgur.com/fi0dHgf.png https://i.imgur.com/LLbDaRw.png Based on my read of the experiments in the paper, the most impressive thing here is how well their information bottleneck mechanism works as a way to improve generalization, compared to both the baseline and other regularization approaches. It does look like there's some additional benefit to SNI, particularly in the CoinRun setting, but very little in the MultiRoom setting, and in general the difference is less dramatic than the difference from using the information bottleneck. 
[link]
Domain translation  for example, mapping from a summer to a winter scene, or from a photorealistic image to an object segmentation map  is often performed by GANs through something called cycle consistency loss. This model works by having, for each domain, a generator to map domain A into domain B, and a discriminator to differentiate between real images from domain B, and those that were constructed through the crossdomain generator. With a given image in domain A, training happens by using the A→B generator to map it into domain B, and then then B→ A generator to map it back the original domain. These generators are then trained using two losses: one based on the Bdomain discriminator, to push the generated image to look like it belongs from that domain, and another based on the L2 loss between the original domain A image, and the image you get on the other end when you translate it into B and back again. This paper addresses an effect (identified originally in an earlier paper) where in domains with a many to one mapping between domains (for example, mapping a realistic scene into a domain segmentation map, where information is inherently lost by translating pixels to object outlines), the cycle loss incentivizes the model to operate in a strange, steganographic way, where it saves information about the that would otherwise be lost in the form of lowamplitude random noise in the translated image. This lowamplitude information can't be isolated, but can be detected in a few ways. First, we can simply examine images and notice that information that could not have been captured in the lowerinformation domain is being perfectly reconstructed. Second, if you add noise to the translation in the lowerinformation domain, in such a way as to not perceptibly change the translation to human eyes, this can cause the predicted image off of that translation to deteriorate considerably, suggesting that the model was using information that could be modified by such small additions of noise to do its reconstruction. https://i.imgur.com/08i1j0J.png The authors of this paper ask whether it's possible to train models that don't perform this steganographic informationstoring (which they call "self adversarial examples"). A typical approach to such a problem would be to train generators to perform translations with and without the steganographic information, but even though we can prove the existence of the information, we can't isolate it in a way that would allow us to remove it, and thus create these kinds of training pairs. The two tactics the paper uses are: 1) Simply training the generators to be able to translate a domainmapped image with noise as well as one without noise, in the hope that this would train it not use information that can be interfered with by the application of such noise. 2) In addition to a L2 cycle loss, adding a discriminator to differentiate between the backtranslated image and the original one. I believe the idea here is that if both of the encoders are adding in noise as a kind of secret signal, this would be a way for the discriminator to distinguish between the original and reconstructed image, and would thus be penalized. They find that both of these methods reduce the use of steganographic information, as determined both by sensitivity to noise (where less sensitivity of reconstruction to noise means less use of coded information) and reconstruction honesty (which constrains accuracy of reconstruction in many to one domains to be no greater than the prediction that a supervised predictor could make given the image from the compressed domain 
[link]
A common critique of deep learning is its brittleness offdistribution, combined with its tendency to give confident predictions for offdistribution inputs, as is seen in the case of adversarial examples. In response to this critique, a number of different methods have cropped up in recent years, that try to capture a model's uncertainty as well as its overall prediction. This paper tries to do a broad evaluation of uncertainty methods, and, particularly, to test how they perform on out of distribution data, including both data that is perturbed from its original values, and fully OOD data from groundtruth categories never seen during training. Ideally, we would want an uncertainty method that is less confident in its predictions as data is made more dissimilar from the distribution that the model is trained on. Some metrics the paper uses for capturing this are:  Brier Score (The difference between predicted score and ground truth 0/1 label, averaged over all examples)  Negative Log Likelihood  Expected Calibration Error (Within a given bucket, this is calculated as the difference between accuracy to ground truth labels, and the average predicted score in that bucket, capturing that you'd ideally want to have a lower predicted score in cases where you have low accuracy, and vice versa)  Entropy  For labels that are fully out of distribution, and don't map to any of the model's categories, you can't directly calculate ground truth accuracy, but you can ideally ask for a model that has high entropy (close to uniform) probabilities over the classes it knows about when the image is drawn from an entirely different class The authors test over image datasets small (MNIST) and large (ImageNet and CIFAR10), as well as a categorical adclickprediction dataset. They came up with some interesting findings. https://i.imgur.com/EVnjS1R.png 1. More fully principled Bayesian estimation of posteriors over parameters, in the form of Stochastic Variational Inference, works well on MNIST, but quite poorly on either categorical data or higher dimensional image datasets https://i.imgur.com/3emTYNP.png 2. Temperature scaling, which basically performs a second supervised calibration using a holdout set to push your probabilities towards true probabilities, performs well indistribution but collapses fairly quickly offdistribution (which sort of makes sense given that it too is just another supervised method that can do poorly when offdistribution) 3. In general, ensemble methods, where you train different models on different subsets of the data and take their variance as uncertainty, perform the best across the bigger image models as well as the ad click model, likely because SVI (along with many other Bayesian methods) is too computationally intensive to get to work well on higherdimensional data 4. Overall, none of the methods worked particularly well, and even the bestperforming ones were often confidently wrong offdistribution I think it's fair to say that we're far from where we wish we were when it comes to models that "know when they don't know," and this paper does a good job of highlighting that in specific fashion.
1 Comments

[link]
This paper combines imitation learning algorithm GAIL with recent advances in goalconditioned reinforcement learning, to create a combined approach that can make efficient use of demonstrations, but can also learn information about a reward that can allow the agent to outperform the demonstrator. Goalconditioned learning is a form of rewarddriven reinforcement learning where the reward is a defined to be 1 when an agent reaches a particular state, and 0 otherwise. This can be a particularly useful form of learning for navigation tasks, where, instead of only training your agent to reach a single hardcoded goal (as you would with a reward function) you teach it to reach arbitrary goals when information about the goal is passed in as input. A typical difficulty with this kind of learning is that its reward is sparse: for any given goal, if an agent never reaches it, it won't ever get reward signal it can use to learn to find it again. A clever solution to this, proposed by earlier method HER (Hindsight Experience Replay), is to perform rollouts of the agent trajectory, and then train your model to reach all the states it actually reached along that trajectory. Said another way, even if your agent did a random, useless thing with respect to one goal, if you retroactively decided that the goal was where it ended up, then it'd be able to receive reward signal after all. In a learning scenario with a fixed reward, this trick wouldn't make any sense, since you don't want to train your model to only go wherever it happened to initially end up. But because the policy here is goalconditioned, we're not giving our policy wrong information about how to go to the place we want, we're incentivizing it to remember ways it got to where it ended up, in the hopes that it can learn generalizable things about how to reach new places. The other technique being combined in this paper is imitation learning, or learning from demonstrations. Demonstrations can be highly useful for showing the agent how to get to regions of state space it might not find on its own. The authors of this paper advocate creating a goalconditioned version of one particular imitation learning algorithm (Generative Adversarial Imitation Learning, or GAIL), and combining that with an offpolicy version of Hindsight Experience Replay. In their model, a discriminator tries to tell the behavior of the demonstrator from that of the agent, given some input goal, and uses that as loss, combined with the loss of a more normal Q learning loss with a reward set to 1 when a goal is achieved. Importantly, they amplify both of these methods using the relabeling trick mentioned before: for both the demonstrators and the actual agent trajectories, they take tuples of (state, next state, goal) and replace the intended goal with another state reached later in the trajectory. For the Q learner, this performs its normal role as a way to get reward in otherwise sparse settings, and for the imitation learner, it is a form of data amplification, where a single trajectory + goal can be turned into multiple trajectories "successfully" reaching all of the intermediate points along the observed trajectory. The authors show that their method learns more quickly (as a result of the demonstrations), but also is able to outperform demonstrators, which it wouldn't generally be able to do without an independent, nondemonstrator reward signal 
[link]
An interesting category of machine learning papers  to which this paper belongs  are papers which use learning systems as a way to explore the incentive structures of problems that are difficult to intuitively reason about the equilibrium properties of. In this paper, the authors are trying to better understand how different dynamics of a cooperative communication game between agents, where the speaking agent is trying to describe an object such that the listening agent picks the one the speaker is being shown, influence the communication protocol (or, to slightly anthropomorphize, the language) that the agents end up using. In particular, the authors experiment with what happens when the listening agent is frequently replaced during training with a untrained listener who has no prior experience with the agent. The idea of this experiment is that if the speaker is in a scenario where listeners need to frequently "relearn" the mapping between communication symbols and objects, this will provide an incentive for that mapping to be easier to quickly learn. https://i.imgur.com/8csqWsY.png The metric of ease of learning that the paper focuses on is "topographic similarity", which is a measure of how compositional the communication protocol is. The objects they're working with have two properties, and the agents use a pair of two discrete symbols (two letters) to communicate about them. A perfectly compositional language would use one of the symbols to represent each of the properties. To mathematically measure this property, the authors calculate (cosine) similarity between the two objects property vectors, and the (edit) distance between the two objects descriptions under the emergent language, and calculate the correlation between these quantities. In this experimental setup, if a language is perfectly compositional, the correlation will be perfect, because every time a property is the same, the same symbol will be used, so two objects that share that property will always share that symbol in their linguistic representation. https://i.imgur.com/t5VxEoX.png The premise and the experimental setup of this paper are interesting, but I found the experimental results difficult to gain intuition and confidence from. The authors do show that, in a regime where listeners are reset, topographic similarity rises from a beginningoftraining value of .54 to an end of training value of .59, whereas in the baseline, noreset regime, the value drops to .51. So there definitely is some amount of support for their claim that listener resets lead to higher compositionality. But given that their central quality is just a correlation between similarities, it's hard to gain intuition for whether the difference is a meaningful. It doesn't naively seem particularly dramatic, and it's hard to tell otherwise without more references for how topographic similarity would change under a wider range of different training scenarios. 
[link]
Reinforcement Learning is often broadly separated into two categories of approaches: modelfree and modelbased. In the former category, networks simply take observations and input and produce predicted bestactions (or predicted values of available actions) as output. In order to perform well, the model obviously needs to gain an understanding of how its actions influence the world, but it doesn't explicitly make predictions about what the state of the world will be after an action is taken. In modelbased approaches, the agent explicitly builds a dynamics model, or a model in which it takes in (past state, action) and predicts next state. In theory, learning such a model can lead to both interpretability (because you can "see" what the model thinks the world is like) and robustness to different reward functions (because you're learning about the world in a way not explicitly tied up with the reward). This paper proposes an interesting melding of these two paradigms, where an agent learns a model of the world as part of an endtoend policy learning. This works through something the authors call "observational dropout": the internal model predicts the next state of the world given the prior one and the action, and then with some probability, the state of the world that both the policy and the next iteration of the dynamics model sees is replaced with the model's prediction. This incentivizes the network to learn an effective dynamics model, because the farther the predictions of the model are from the true state of the world, the worse the performance of the learned policy will be on the iterations where the only observation it can see is the predicted one. So, this architecture is modelfree in the sense that the gradient used to train the system is based on applying policy gradients to the reward, but modelbased in the sense that it does have an internal world representation. https://i.imgur.com/H0TNfTh.png The authors find that, at a simple task, Swing Up Cartpole, very low probabilities of seeing the true world (and thus very high probabilities of the policy only seeing the dynamics model output) lead to world models good enough that a policy trained only on trajectories sampled from that model can perform relatively well. This suggests that at higher probabilities of the true world, there was less value in the dynamics model being accurate, and consequently less training signal for it. (Of course, policies that often could only see the predicted world performed worse during their original training iteration compared to policies that could see the real world more frequently). On a more complex task of CarRacing, the authors looked at how well a policy trained using the representations of the world model as input could perform, to examine whether it was learning useful things about the world. https://i.imgur.com/v9etll0.png They found an interesting tradeoff, where at high probabilities (like before) the dynamics model had little incentive to be good, but at low probabilities it didn't have enough contact with the real dynamics of the world to learn a sensible policy. 
[link]
In the last two years, the Transformer architecture has taken over the worlds of language modeling and machine translation. The central idea of Transformers is to use selfattention to aggregate information from variablelength sequences, a task for which Recurrent Neural Networks had previously been the most common choice. Beyond that central structural change, one more nuanced change was from having a single attention mechanism on a given layer (with a single set of query, key, and value weights) to having multiple attention heads, each with their own set of weights. The change was framed as being conceptually analogous to the value of having multiple feature dimensions, each of which focuses on a different aspect of input; these multiple heads could now specialize and perform different weighted sums over input based on their specialized function. This paper performs an experimental probe into the value of the various attention heads at test time, and tries a number of different pruning tests across both machine translation and language modeling architectures to see their impact on performance. In their first ablation experiment, they test the effect of removing (that is, zeromasking the contribution of) a single head from a single attention layer, and find that in almost all cases (88 out of 96) there's no statistically significant drop in performance. Pushing beyond this, they ask what happens if, in a given layer, they remove all heads but the one that was seen to be most important in the single head tests (the head that, if masked, caused the largest performance drop). This definitely leads to more performance degradation than the removal of single heads, but the degradation is less than might be intuitively expected, and is often also not statistically significant. https://i.imgur.com/Qqh9fFG.png This also shows an interesting distribution over where performance drops: in machine translation, it seems like decoderdecoder attention is the least sensitive to heads being pruned, and encoderdecoder attention is the most sensitive, with a very dramatic performance dropoff observed if particularly the last layer of encoderdecoder attention is stripped to a single head. This is interesting to me insofar as it shows the intuitive roots of attention in these architectures; attention was originally used in encoderdecoder parts of models to solve problems of pulling out information in a source sentence at the time it's needed in the target sentence, and this result suggests that a lot of the value of multiple heads in translation came from making that mechanism more expressive. Finally, the authors performed an iterative pruning test, where they ordered all the heads in the network according to their singlehead importance, and pruned starting with the least important. Similar to the results above, they find that drops in performance at high rates of pruning happen eventually to all parts of the model, but that encoderdecoder attention suffers more quickly and more dramatically if heads are removed. https://i.imgur.com/oS5H1BU.png Overall, this is a clean and straightforward empirical paper that asks a fairly narrow question and generates some interesting findings through that question. These results seem reminiscent to me of the Lottery Ticket Hypothesis line of work, where it seems that having a network with a lot of weights is useful for training insofar as it gives you more chances at an initialization that allows for learning, but that at test time, only a small percentage of the weights have ultimately become important, and the rest can be pruned. In order to make the comparison more robust, I'd be interested to see work that does more specific testing of the number of heads required for good performance during training and also during testing, divided out by different areas of the network. (Also, possibly this work exists and I haven't found it!) 
[link]
SelfSupervised Learning is a broad category of approaches whose goal is to learn useful representations by asking networks to perform constructed tasks that only use the content of a dataset itself, and not external labels. The idea with these tasks is to design tasks such that solving them requires the network to have learned useful Some examples of this approach include predicting the rotation of rotated images, reconstructing color from greyscale, and, the topic of this paper, maximizing mutual information between different areas of the image. The hope behind this last approach is that if two areas of an image are generated by the same set of underlying factors (in the case of a human face: they're parts of the same person's face), then a representation that correctly captures those factors for one area will give you a lot of information about the representation of the other area. Historically, this conceptual desire for representations that are mutually informative has been captured by mutual information. If we define the representation distribution over the data of area 1 as p(x) and area 2 as q(x), the mutual information is the KL divergence between the joint distribution of these two distributions and the product of their marginals. This is an old statistical intuition: the closer the joint is to the product of marginals, the closer the variables are to independent; the farther away, the closer they are to informationally identical. https://i.imgur.com/2SzD5d5.png This paper argues that the presence of the KL divergence in this mutual information formulation impedes the ability of networks to learn useful representations. This argument is theoretically based on a result from a recent paper (which for the moment I'll just take as foundation, without reading it myself) that empirical lowerbound measurements of mutual information, of the kind used in these settings, are upper bounded by log(n) where n is datapoints. Our hope in maximizing a lower bound to any quantity is that the bound is fairly tight, since that means that optimizing a network to push upward a lower bound actually has the effect of pushing the actual value up as well. If the lower bound we can estimate is constrained to be far below the actual lower bound in the data, then pushing it upward doesn't actually require the value to move upward. The authors identify this as a particular problem in areas where the underlying mutual information of the data is high, such as in videos where one frame is very predictive of the next, since in those cases the constraint imposed by the dataset size will be small relative to the actual possible maximum mutual information you could push your network to achieve. https://i.imgur.com/wm39mQ8.png Taking a leaf out of the GAN literature, the authors suggest keeping replacing the KL divergence component of mutual information and replacing it with the Wasserstein Distance; otherwise known as the "earthmover distance", the Wasserstein distance measures the cost of the least costly way to move probability mass from one distribution to another, assuming you're moving that mass along some metric space. A nice property of the Wasserstein distance, in both GANs and in this application) is that they don't saturate quite as quickly: the value of a KL divergence can shoot up if the distributions are even somewhat different, making it unable to differentiate between distributions that are somewhat and very far away, whereas a Wasserstein distance continues to have more meaningful signal in that regime. In the context of the swap for mutual information, the authors come up with the "Wasserstein Dependency Measure", which is just the Wasserstein Distance between the joint distributions and the product of the marginals. https://i.imgur.com/3s2QRRz.png In practice, they use the dual formulation of the Wasserstein distance, which amounts to applying a (neural network) function f(x) to values from both distributions, optimizing f(x) so that the values are far apart, and using that distance as your training signal. Crucially, this function has to be relatively smooth in order for the dual formulation to work: in particular it has to have a small Lipschitz value (meaning its derivatives are bounded by some value). Intuitively, this has the effect of restricting the capacity of the network, which is hoped to incentivize it to use its limited capacity to represent true factors of variation, which are assumed to be the most compact way to represent the data. Empirically, the authors found that their proposed Wasserstein Dependency Measure (with a slight variation applied to reduce variance) does have the predicted property of performing better for situations where the native mutual information between two areas is high. I found the theoretical points of this paper interesting, and liked the generalization of the idea of Wasserstein distances from GANs to a new area. That said, I wish I had a better mechanical sense for how it ground out in actual neural network losses: this is partially just my own lack of familiarity with how e.g. mutual information losses are actually formulated as network objectives, but I would have appreciated an appendix that did a bit more of that mapping between mathematical intuition and practical network reality. 
[link]
In response to increasing calls for ways to explain and interpret the predictions of neural networks, one major genre of explanation has been the construction of salience maps for imagebased tasks. These maps assign a relevance or saliency score to every pixel in the image, according to various criteria by which the value of a pixel can be said to have influenced the final prediction of the network. This paper is an interesting blend of ideas from the saliency mapping literature with ones from adversarial examples: it essentially shows that you can create adversarial examples whose goal isn't to change the output of a classifier, but instead to keep the output of the classifier fixed, but radically change the explanation (their term for the previouslydescribed pixel saliency map that results from various explanationfinding methods) to resemble some desired target explanation. This is basically a targeted adversarial example, but targeting a different property of the network (the calculated explanation) while keeping an additional one fixed (keeping the output of the original network close to the original output, as well as keeping the input image itself in a norm ball around the original image. This is done in a pretty standard way: by just defining a loss function incentivizing closeness to the original output and also closeness of the explanation to the desired target, and performing gradient descent to modify pixels until this loss was low. https://i.imgur.com/N9uReoJ.png The authors do a decent job of showing such targeted perturbations are possible: by my assessment of their results their strongest successes at inducing an actual targeted explanation are with Layerwise Relevance Propogation and Pattern Attribution (two of the 6 tested explanation methods). With the other methods, I definitely buy that they're able to induce an explanation that's very unlike the true/original explanation, but it's not as clear they can reach an arbitrary target. This is a bit of squinting, but it seems like they have more success in influencing propogation methods (where the effect size of the output is propogated backwards through the network, accounting for ReLus) than they do with gradient ones (where you're simply looking at the gradient of the output class w.r.t each pixel. In the theory section of the paper, the authors do a bit of differential geometry that I'll be up front and say I did not have the niche knowledge to follow, but which essentially argues that the manipulability of an explanation has to do with the curvature of the output manifold for a constant output. That is to say: how much can you induce a large change in the gradient of the output, while moving a small distance along the manifold of a constant output value. They then go on to argue that ReLU activations, because they are by definition discontinuous, induce sharp changes in gradient for points nearby one another, and this increase the ability for networks to be manipulated. They propose a softplus activation instead, where instead of a sharp discontinuity, the ReLU shape becomes more curved, and show relatively convincingly that at low values of Beta (more curved) you can mostly eliminate the ability of a perturbation to induce an adversarially targeted explanation. https://i.imgur.com/Fwu3PXi.png For all that I didn't have a completely solid grasp of some of the theory sections here, I think this is a neat proof of concept paper in showing that neural networks can be smallperturbation fragile on a lot of different axes: we've known this for a while in the area of adversarial examples, but this is a neat generalization of that fact to a new area. 
[link]
If your goal is to interpret the predictions of neural networks on images, there are a few different ways you can focus your attention. One approach is to try to understand and attach conceptual tags to learnt features, to form a vocabulary with which models can be understood. However, techniques in this family have to content with a number of challenges, from the difficulty in attaching clear concepts to the sheer number of neurons to interpret. An alternate approach, and the one pursued by this paper, is to frame interpretability as a matter of introspecting on *where in an image* the model is pulling information from to make its decision. This is the question for which hard attention provides an answer: identify where in an image a model is making a decision by learning a metamodel that selects small patches of an image, and then makes a classification decision by applying a network to only those patches which were selected. By definition, if only a discrete set of patches were used for prediction, those were the ones that could be driving the model's decision. This central fact of the model only choosing a discrete set of patches is a key complexity, since the choice to use a patch or not is a binary, discontinuous action, and not something through which one can backpropogate gradients. Saccader, the approach put forward by this paper, proposes an architecture which extracts features from locations within an image, and uses those spatially located features to inform a stochastic policy that selects each patch with some probability. Because reinforcement learning by construction is structured to allow discrete actions, the system as a whole can be trained via policy gradient methods. https://i.imgur.com/SPK0SLI.png Diving into a bit more detail: while I don't have a deep familiarity with prior work in this area, my impression is that the notion of using policy gradient to learn a hard attention policy isn't a novel contribution of this work, but rather than its novelty comes from clever engineering done to make that policy easier to learn. The authors cite the problem of sparse reward in learning the policy, which I presume to mean that if you start in more traditional RL fashion by just sampling random patches, most patches will be unclear or useless in providing classification signal, so it will be hard to train well. The Saccader architecture works by extracting localized features in an architecture inspired by the 2019 BagNet paper, which essentially applies very tall and narrow convolutional stacks to spatially small areas of the image. This makes it the case that feature vectors for different overlapping patches can be computed efficiently: instead of rerunning the network again for each patch, it just combined the features from the "tops" of all of the small column networks inside the patch, and used that aggregation as a patchlevel feature. These features from the "representation network" were then used in an "attention network," which uses larger receptive field convolutions to create patchlevel features that integrated the context of things around them. Once these two sets of features were created, they were fed into the "Saccader cell", which uses them to calculate a distribution over patches which the policy then samples over. The Saccader cell is a simplified memory cell, which sets a value to 1 when a patch has been sampled, and applies a very strong penalization on that patch being sampled on future "glimpses" from the policy (in general, classification is performed by making a number of draws and averaging the logits produced for each patch). https://i.imgur.com/5pSL0oc.png I found this paper fairly conceptually clever  I hadn't thought much about using a reinforcement learning setup for classification before  though a bit difficult to follow in its terminology and notation. It's able to perform relatively well on ImageNet, though I'm not steeped enough in that as a benchmark to have an intuitive sense for the paper's claim that their accuracy is meaningfully in the same ballpark as fullimage models. One interesting point the paper made was that their system, while limited to small receptive fields for the patch features, can use an entirely different model for mapping patches to logits once the patches are selected, and so can benefit from more powerful generic classification models being tacked onto the end. 
[link]
Ilyas et al. present a followup work to their paper on the tradeoff between accuracy and robustness. Specifically, given a feature $f(x)$ computed from input $x$, the feature is considered predictive if $\mathbb{E}_{(x,y) \sim \mathcal{D}}[y f(x)] \geq \rho$; similarly, a predictive feature is robust if $\mathbb{E}_{(x,y) \sim \mathcal{D}}\left[\inf_{\delta \in \Delta(x)} yf(x + \delta)\right] \geq \gamma$. This means, a feature is considered robust if the worstcase correlation with the label exceeds some threshold $\gamma$; here the worstcase is considered within a predefined set of allowed perturbations $\Delta(x)$ relative to the input $x$. Obviously, there also exist predictive features, which are however not robust according to the above definition. In the paper, Ilyas et al. present two simple algorithms for obtaining adapted datasets which contain only robust or only nonrobust features. The main idea of these algorithms is that an adversarially trained model only utilizes robust features, while a standard model utilizes both robust and nonrobust features. Based on these datasets, they show that nonrobust, predictive features are sufficient to obtain high accuracy; similarly training a normal model on a robust dataset also leads to reasonable accuracy but also increases robustness. Experiments were done on Cifar10. These observations are supported by a theoretical toy dataset consisting of two overlapping Gaussians; I refer to the paper for details. Also find this summary at [davidstutz.de](https://davidstutz.de/category/reading/). 
[link]
The Lottery Ticket Hypothesis is the idea that you can train a deep network, set all but a small percentage of its highmagnitude weights to zero, and retrain the network using the connection topology of the remaining weights, but only if you reinitialize the unpruned weights to the the values they had at the beginning of the first training. This suggests that part of the value of training such big networks is not that we need that many parameters to use their expressive capacity, but that we need many “draws” from the weight and topology distribution to find initial weight patterns that are welldisposed for learning. This paper out of Uber is a refreshingly exploratory experimental work that tries to understand the contours and contingencies of this effect. Their findings included:  The pruning criteria used in the original paper, where weights are kept according to which have highest final magnitude, works well. However, an alternate criteria, where you keep the weights that have increased the most in magnitude, works just as well and sometimes better. This makes a decent amount of sense, since it seems like we’re using magnitude as a signal of “did this weight come to play a meaningful role during training,” and so weights whose influence increased during training fall in that category, regardless of their starting point https://i.imgur.com/wTkNBod.png  The authors’ next question was: other than just reinitialize weights to their initial values, are there other things we can do that can capture all or part of the performance effect? The answer seems to be yes; they found that the most important thing seems to be keeping the sign of the weights aligned with what it was at its starting point. As long as you do that, redrawing initial weights (but giving them the right sign), or resetting weights to a correctly signed constant value, both work nearly as well as the actual starting values https://i.imgur.com/JeujUr3.png  Turning instead to the weights on the pruning chopping block, the authors find that, instead of just zeroing out all pruned weights, they can get even better performance if they zero the weights that moved towards zero during training, and reinitialize (but freeze) the weights that moved away from zero during training. The logic of the paper is “if the weight was trying to move to zero, bring it to zero, otherwise reinitialize it”. This performance remains high at even lower levels of training than does the initial zeromasking result  Finally, the authors found that just by performing the masking (i.e. keeping only weights with large final values), bringing those back to their values, and zeroing out the rest, *and not training at all*, they were able to get 40% test accuracy on MNIST, much better than chance. If they masked according to “large weights that kept the same sign during training,” they could get a pretty incredible 80% test accuracy on MNIST. Way below even simple trained models, but, again, this model wasn’t *trained*, and the only information about the data came in the form of a binary weight mask This paper doesn’t really try to come up with explanations that wrap all of these results up neatly with a bow, and I really respect that. I think it’s good for ML research culture for people to feel an affordance to just run a lot of targeted experiments aimed at explanation, and publish the results even if they don’t quite make sense yet. I feel like on this problem (and to some extent in machine learning generally), we’re the blind men each grabbing at one part of an elephant, trying to describe the whole. Hopefully, papers like this can bring us closer to understanding strange quirks of optimization like this one 
[link]
As per the “holistic” in the paper title, the goal of this work is to take a suite of existing work within semisupervised learning, and combine many of its ideas into one training pipeline that can (with really impressive empirical success) leverage the advantages of those different ideas. The core premise of supervised learning is that, given truelabel training signal from a small number of labels, you can leverage large amounts of unsupervised data to improve your model. A central intuition of many of these methods is that, even if you don’t know the class of a given sample, you know it *has* a class, and you can develop a loss by pushing your model to predict the class for an example and a modified or perturbed version of that example, since, if you have a prior belief that that modification should not change your true class label, then your unlabeled data point should have the same class prediction both times. Entropy minimization is built off similar notions: although we don’t know a point’s class, we know it must have one, and so we’d like our model to make a prediction that puts more of its weight on a single class, rather than be spread out, since we know the “correct model” will be a very confident prediction of one class, though we don’t know which it is. These methods will give context and a frame of mind for understanding the techniques merged together into the MixMatch approach. At its very highest level, MixMatch’s goal is to take in a dataset of both labeled and unlabeled data, and produce a training set of inputs, predictions, and (occasionally constructed or modified labels) to calculate a model update loss from. https://i.imgur.com/6lHQqMD.png  First, for each unlabeled example in the dataset, we produce K different augmented versions of that image (by cropping it, rotating it, flipping it, etc). This is in the spirit of the consistency loss literature, where you want your model to make the same prediction across augmentations  Do the same augmentation for each labeled example, but only once per input, rather than k times  Run all of your augmented examples through your model, and take the average of their predictions. This is based on the idea that the average of the predictions will be a lower variance, more stable pseudotarget to pull each of the individual predictions towards. Also in the spirit of making something more shaped like a real label, they undertake a sharpening step, turning down the temperature of the averaged distribution. This seems like it would have the effect of more confidently pulling the original predictions towards a single “best guess” label  At this point, we have a set of augmented labeled data, with a true label, and also a set of augmented unlabeled data, with a label based off of an averaged and sharpened best guess from the model over different modifications. At this point, the pipeline uses something called “MixUp” (on which there is a previous paper, so I won’t dive into it too much here), which takes pairs of data points, calculates a convex combination of the inputs, runs it through the model, and uses as the lossfunction target a convex combination of the outputs. So, in the simple binary case, if you have a positive and negatively labeled image and sample a combination parameter of 0.75, you have an image that is 0.75 positive, 0.25 negative, and the new label that you’re calculating cross entropy loss against is 0.75.  MixMatch generates pairs for its MixUp calculation by mixing (heh) labeled and unlabeled data together, and pairing each labeled and unlabeled pair with one observation from the merged set. At this point, we have combined inputs, and we have combined labels, and we can calculate loss between them With all of these methods combined, this method takes the previous benchmark of 38% error, for a CIFAR dataset with only 250 labels, and drops that to 11%, which is a pretty astonishing improvement in error rate. After performing an ablation study, they find that MixUp itself, temperature sharpening, and calculating K>1 augmentations of unlabeled data rather than K=1 are the strongest valueadds; it doesn’t appear like there’s that much difference that comes from mixing between unlabeled and labeled for the MixUp pairs. 