Paper deep dive

Editable Neural Networks

Anton Sinitsin, Vsevolod Plokhotnyuk, Dmitriy Pyrkin, Sergei Popov, Artem Babenko

Year: 2020Venue: ICLR 2020Area: Model EditingType: EmpiricalEmbeddings: 42

Models: DenseNet-169, ResNet-18, Transformer (small)

Intelligence

Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 95%

Last extracted: 3/12/2026, 7:56:18 PM

Summary

The paper introduces 'Editable Training', a model-agnostic training technique for deep neural networks that enables efficient, localized patching of model mistakes without degrading performance on other samples. By leveraging meta-learning and gradient-based editing, the method optimizes model parameters to be 'prepared' for future corrections, demonstrating effectiveness on image classification and machine translation tasks.

Entities (5)

CIFAR-10 · dataset · 100%Editable Training · method · 100%ResNet-18 · neural-network-architecture · 98%Gradient Descent Editor · algorithm · 95%Meta-learning · paradigm · 95%

Relation Signals (3)

Editable Training → evaluatedon → CIFAR-10

confidence 100% · First, we experiment on image classification with the small CIFAR-10 dataset

Editable Training → optimizes → Neural Network Parameters

confidence 90% · The core idea behind Editable Training is to enforce the model parameters θ to be 'prepared' for the editor function.

Gradient Descent Editor → implements → Editable Training

confidence 85% · A natural way to implement Edit(θ,l e ) for deep neural networks is using gradient descent.

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:These days deep neural networks are ubiquitously used in a wide range of tasks, from image classification and machine translation to face identification and self-driving cars. In many applications, a single model error can lead to devastating financial, reputational and even life-threatening consequences. Therefore, it is crucially important to correct model mistakes quickly as they appear. In this work, we investigate the problem of neural network editing $-$ how one can efficiently patch a mistake of the model on a particular sample, without influencing the model behavior on other samples. Namely, we propose Editable Training, a model-agnostic training technique that encourages fast editing of the trained model. We empirically demonstrate the effectiveness of this method on large-scale image classification and machine translation tasks.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

41,890 characters extracted from source content.

Expand or collapse full text

Published as a conference paper at ICLR 2020 EDITABLENEURALNETWORKS Anton Sinitsin 1∗ ant.sinitsin@gmail.com Vsevolod Plokhotnyuk 2∗ vsevolod-pl@yandex.ru Dmitriy Pyrkin 2∗ alagaster@yandex.ru Sergei Popov 1,2 popovsergey95@gmail.com Artem Babenko 1,2 artem.babenko@phystech.edu 1 Yandex 2 National Research University Higher School of Economics ABSTRACT These days deep neural networks are ubiquitously used in a wide range of tasks, from image classification and machine translation to face identification and self- driving cars. In many applications, a single model error can lead to devastating financial, reputational and even life-threatening consequences. Therefore, it is cru- cially important to correct model mistakes quickly as they appear. In this work, we investigate the problem of neural network editing — how one can efficiently patch a mistake of the model on a particular sample, without influencing the model be- havior on other samples. Namely, we propose Editable Training, a model-agnostic training technique that encourages fast editing of the trained model. We empiri- cally demonstrate the effectiveness of this method on large-scale image classifica- tion and machine translation tasks. 1INTRODUCTION Deep neural networks match and often surpass human performance on a wide range of tasks includ- ing visual recognition (Krizhevsky et al. (2012); D. C. Ciresan (2011)), machine translation (Hassan et al. (2018)) and others (Silver et al. (2016)). However, just like humans, artificial neural networks sometimes make mistakes. As we trust them with more and more important decisions, the cost of such mistakes grows ever higher. A single misclassified image can be negligible in academic re- search but can be fatal for a pedestrian in front of a self-driving vehicle. A poor automatic translation for a single sentence can get a person arrested (Hern (2018)) or ruin company’s reputation. Since mistakes are inevitable, deep learning practitioners should be able to adjust model behavior by correcting errors as they appear. However, this is often difficult due to the nature of deep neural networks. In most network architectures, a prediction for a single input depends on all model param- eters. Therefore, updating a neural network to change its predictions on a single input can decrease performance across other inputs. Currently, there are two workarounds commonly used by practitioners. First, one can re-train the model on the original dataset augmented with samples that account for the mistake. However, this is computationally expensive as it requires to perform the training from scratch. Another solution is to use a manual cache (e.g. lookup table) that overrules model predictions on problematic samples. While being simple, this approach is not robust to minor changes in the input. For instance, it will not generalize to a different viewpoint of the same object or paraphrasing in natural language processing tasks. In this work, we present an alternative approach that we call Editable Training. This approach involves training neural networks in such a way that the trained parameters can be easily edited af- terwards. Editable Training employs modern meta-learning techniques (Finn et al. (2017)) to ensure ∗ Equal contribution 1 arXiv:2004.00345v2 [cs.LG] 22 Jul 2020 Published as a conference paper at ICLR 2020 that model’s mistakes can be corrected without harming its overall performance. With thorough experimental evaluation, we demonstrate that our method works on both small academical datasets and industry-scale machine learning tasks. We summarize the contributions of this study as follows: •We address a new problem of fast editing of neural network models. We argue that this problem is extremely important in practice but, to the best of our knowledge, receives little attention from the academic community. •We propose Editable Training — a model-agnostic method of neural network training that learns models, whose errors can then be efficiently corrected. 1 •We extensively evaluate Editable Training on large-scale image classification and machine translation tasks, confirming its advantage over existing baselines. 2RELATED WORK In this section, we aim to position our approach with respect to existing literature. Namely, we explain the connections of Editable Neural Networks with ideas from prior works. Meta-learningis a family of methods that aim to produce learning algorithms, appropriate for a particular machine learning setup. These methods were shown to be extremely successful in a large number of problems, such as few-shot learning (Finn et al. (2017); Nichol et al. (2018)), learnable optimization (Andrychowicz et al. (2016)) and reinforcement learning (Houthooft et al. (2018)). In- deed, Editable Neural Networks also belong to the meta-learning paradigm, as they basically ”learn to allow effective patching”. While neural network correction has significant practical importance, we are not aware of published meta-learning works, addressing this problem. Catastrophic forgettingis a well-known phenomenon arising in the problem of lifelong/continual learning (Ratcliff (1990)). For a sequence of learning tasks, it turns out that after deep neural net- works learn on newer tasks, their performance on older tasks deteriorates. Several lines of research address overcoming catastrophic forgetting. The methods based on Elastic Weight Consolidation (Kirkpatrick et al. (2016)) update model parameters based on their importance to the previous learn- ing tasks. The rehearsal-based methods (Robins (1995)) occasionally repeat learning on samples from earlier tasks to ”remind” the model about old data. Finally, a line of work (Garnelo et al. (2018); Lin et al. (2019)) develops specific neural network architectures that reduce the effect of catastrophic forgetting. The problem of efficient neural network patching differs from continual learning, as our setup is not sequential in nature. However, correction of model for mislabeled samples must not affect its behavior on other samples, which is close to overcoming catastrophic forgetting task. Adversarial training.The proposed Editable Training also bears some resemblance to the ad- versarial training (Goodfellow et al. (2015)), which is the dominant approach of adversarial attack defense. The important difference here is that Editable Training aims to learn models, whose behav- ior on some samples can be efficiently corrected. Meanwhile, adversarial training produces models, which are robust to certain input perturbations. However, in practice one can use Editable Train- ing to efficiently cover model vulnerabilities against both synthetic (Szegedy et al. (2013); Yuan et al. (2017); Ebrahimi et al. (2017); Wallace et al. (2019)) and natural (Hendrycks et al. (2019)) adversarial examples. 3EDITINGNEURALNETWORKS In order to measure and optimize the model’s ability for editing, we first formally define the oper- ation of editing a neural network. Letf(x,θ)be a neural network, withxdenoting its input and θbeing a set of network parameters. The parametersθare learned by minimizing a task-specific objective functionL base (θ), e.g. cross-entropy for multi-class classification problems. Then, if we discover mistakes in the model’s behavior, we can patch the model by changing its parametersθ. Here we aim to change model’s predictions on a subset of inputs, corresponding to misclassified objects, without affecting other inputs. We formalize this goal using theeditor 1 The source code is available online athttps://github.com/xtinkt/editable 2 Published as a conference paper at ICLR 2020 function: ˆ θ=Edit(θ,l e ). Informally, this is a function that adjustsθto satisfy a given constraint l e ( ˆ θ)≤0, whose role is to enforce desired changes in the model’s behavior. For instance, in the case of multi-class classification,l e can guarantee that the model assigns input xto the desired labely ref :l e ( ˆ θ) = max y i logp(y i |x, ˆ θ)−logp(y ref |x, ˆ θ). Under such definition ofl e , the constraintl e ( ˆ θ)≤0is satisfied iffarg max y i logp(y i |x, ˆ θ) =y ref . To be practically feasible, the editor function must meet three natural requirements: •Reliability:the editor must guaranteel e ( ˆ θ)≤0for the chosen family ofl e (·); •Locality:the editor should minimize influence onf(·, ˆ θ)outside of satisfyingl e ( ˆ θ)≤0; •Efficiency:the editor should be efficient in terms of runtime and memory; Intuitively, the editor locality aims to minimize changes in model’s predictions for inputs unrelated tol e . For classification problem, this requirement can be formalized as minimizing the difference between model’s predictions over the ”control” setX c :E x∈X c #[f(x, ˆ θ)6=f(x,θ)]→min. 3.1GRADIENTDESCENTEDITOR A natural way to implementEdit(θ,l e )for deep neural networks is using gradient descent. Param- etersθare shifted against the gradient direction−α∇ θ l e (θ)for several iterations until the constraint l e (θ)≤0is satisfied. We formulate the SGD editor with up toksteps and learning rateαas: Edit k α (θ,l e ,k) = θ,ifl e (θ)≤0ork= 0 Edit k−1 α (θ−α·∇ θ l e (θ),l e ),otherwise (1) The standard gradient descent editor can be further augmented with momentum, adaptive learn- ing rates (Duchi et al. (2010); Zeiler (2012)) and other popular deep learning tricks (Kingma & Ba (2014); Smith & Topin (2017)). One technique that we found practically useful is Resilient Backpropagation: RProp, SignSGD by Bernstein et al. (2018) or RMSProp by Tieleman & Hinton (2012). We observed that these methods produce more robust weight updates that improve locality. 3.2EDITABLETRAINING The core idea behind Editable Training is to enforce the model parametersθto be ”prepared” for the editor function. More formally, we want to learn such parametersθ, that the editorEdit(θ,l e ) is reliable, local and efficient, as defined in above. Our training procedure employs the fact that Gradient Descent Editor (1) is differentiable w.r.t.θ. This well-known observation (Finn et al. (2017)) allows us to optimize through the editor function directly via backpropagation (see Figure 1). l e (⋅) Model: Edit: Batch: θ X,y Edit θ L edit L loc L base Obj(θ) l e (⋅) X,y θ L edit L loc L base θ Edit Obj(θ) Figure 1: A high-level scheme of editable training: (left) forward pass, (right) backward pass. Editable Training is performed on minibatches of constraintsl e ∼p(l e )(e.g. images and target labels). First, we compute the edited parameters ˆ θ=Edit(θ,l e )by applying up toksteps of gradient descent (1). Second, we compute the objective that measures locality and efficiency of the editor function: 3 Published as a conference paper at ICLR 2020 Obj(θ,l e ) =L base (θ) +c edit ·L edit (θ) +c loc ·L loc (θ)(2) L edit (θ) =max(0,l e (Edit k α (θ,l e ))(3) L loc (θ) =E x∼p(x) D KL (p(y|x,θ)||p(y|x,Edit k α (θ,l e )))(4) Intuitively,L edit (θ)encourages reliability and efficiency of the editing procedure by making sure the constraint is satisfied in underkgradient steps. The final termL loc (θ)is responsible for locality by minimizing the KL divergence between the predictions of original and edited models. We use hyperparametersc edit ,c loc to balance between the original task-specific objective, editor ef- ficiency and locality. Setting both of them to large positive values would cause the model to sacrifice some of its performance for a better edit. On the other hand, sufficiently smallc edit ,c loc will not cause any deterioration of the main training objective while still improving the editor function in all our experiments (see Section 4). We attribute this to the fact that neural networks are typically overparameterized. Most neural networks can accommodate the edit-related properties and still have enough capacity to optimizeObj(θ,l e ). The learning stepαand other optimizer parameters (e.g. βfor RMSProp) are trainable parameters of Editable Training and we optimize them explicitly via gradient descent. 4EXPERIMENTS In this section, we extensively evaluate Editable Training on several deep learning problems and compare it to existing alternatives for efficient model patching. 4.1TOYEXPERIMENT: CIFAR-10 First, we experiment on image classification with the small CIFAR-10 dataset with standard train/test splits (Krizhevsky et al.). The training dataset is further augmented with random crops and random horizontal flips. All models trained on this dataset follow the ResNet-18 (He et al. (2015)) architec- ture and use the Adam optimizer (Kingma & Ba (2014)) with default hyperparameters. Our baseline is ResNet-18 (He et al. (2015)) neural network trained to minimize the standard cross- entropy loss without Editable Training. This model provides6.3%test error rate at convergence. Comparing editor functions.As a preliminary experiment, we compare several variations of edi- tor functions for the baseline model without Editable Training. We evaluate each editor by applying N=1000editsl e . Each edit consists of an image from the test set assigned with a random (likely in- correct) label uniformly chosen from 0 to 9. AfterNindependent edits, we compute three following metrics over the entire test set: •Drawdown— mean absolute difference of classification error before and after performing an edit. Smaller drawdown indicates better editor locality. •Success Rate— a rate of edits, for which editor succeeds in underk=10gradient steps. •Num Steps— an average number of gradient steps needed to perform a single edit. Editor FunctionGDScaled GDRPropRMSPropMomentumAdam Drawdown3.8%2.81%1.99%1.77%2.42%19.4% Success Rate98.8%99.1%100%100%96.0%100% Num Steps3.543.912.993.115.603.86 Table 1: Comparison of different editor functions on the CIFAR10 dataset with the baseline ResNet18 model trained without Editable Training. •Gradient Descent (GD)— standard gradient descent. •Scaled GD— like GD, but the learning rate is divided by the global gradient norm from the first gradient step. 4 Published as a conference paper at ICLR 2020 •RProp— like GD, but the algorithm only uses the sign of gradients:θ−α·sign(∇ θ l e (θ)). •RMSProp— like GD, but the learning rate for each individual parameter is divided by √ rms t +whererms 0 = [∇ θ l e (θ 0 )] 2 andrms t+1 =β·rms t + (1−β)·[∇ θ l e (θ)] 2 . •Momentum GD— like GD, but the update follows the accumulated gradient directionν: ν 0 = 0;ν t+1 =α·∇ θ l e (θ 0 ) +μ·ν t . •Adam— adaptive momentum algorithm as described in Kingma & Ba (2014) with tunable α,β 1 ,β 2 . To prevent Adam from replicating RMSProp, we restrictβ 1 to[0.1,1.0]range. For each optimizer, we tune all hyperparameters (e.g. learning rate) to optimize locality while ensuring that editor succeeds in underk= 10steps for at least95%of edits. We also tune the editor function by limiting the subset of parameters it is allowed to edit. The ResNet-18 model consists of six parts: initial convolutional layer, followed by four ”chains” of residual blocks and a final linear layer that predicts class logits. We experimented with editing the whole model as well as editing each individual ”chain”, leaving parameters from other layers fixed. For each editor Table 1 reports the numbers, obtained for the subset of editable parameters, corresponding to the smallest drawdown. For completeness, we also report the drawdown of Gradient Descent and RMSProp for different subsets of editable parameters in Table 2. Editable LayersWhole ModelChain 1Chain 2Chain 3Chain 4 Gradient Descent3.8%18.3%7.7%5.3%4.76% RMSProp2.29%22.8%1.85%1.77%1.99% Table 2: Mean Test Error Drawdown when editing different ResNet18 layers on CIFAR10. Table 1 and Table 2 demonstrate that the editor function locality is heavily affected by the choice of editing function even for models trained without Editable Training. Both RProp and RMSProp significantly outperform the standard Gradient Descent while Momentum and Adam show smaller gains. In fact, without the constraintβ 1 >0.1the tuning procedure returnsβ 1 = 0, which makes Adam equivalent to RMSProp. We attribute the poor performance of Adam and Momentum to the fact that most methods only make a few gradient steps till convergence and the momentum term cannot accumulate the necessary statistics. Editable Training.Finally, we report results obtained with Editable Training. On each training batch, we use a single constraintl e ( ˆ θ) = max y i logp(y i |x, ˆ θ)−logp(y ref |x, ˆ θ), wherexis sampled from the train set andy ref is a random class label (from 0 to 9). The model is then trained by directly minimizing objective (2) withk=10editor steps and all other parameters optimized by backpropagation. We compare our Editable Training against three baselines, which also allow efficient model cor- rection. The first natural baseline is Elastic Weight Consolidation (Kirkpatrick et al. (2016)): a technique that penalizes the edited model with the squared difference in parameter space, weighted by the importance of each parameter. Our second baseline is a semi-parametric Deep k-Nearest Neighbors (DkNN) model (Papernot & McDaniel (2018)) that makes predictions by usingknearest neighbors in the space of embeddings, produced by different CNN layers. For this approach, we edit the model by flipping labels of nearest neighbors until the model predicts the correct class. Finally we compare to alternative editor function inspired by Conditional Neural Processes (CNP) (Garnelo et al. (2018)) that we refer to asEditable+CNP. For this baseline, we train a specialized CNP model architecture that performs edits by adding a specialcondition vectorto intermediate activations. This vector is generated by an additional ”encoder” layer. We train the CNP model to solve the original classification problem when the condition vector is zero (hence, the model behaves as standard ResNet18) and minimizeL edit andL loc when the condition vector is applied. After tuning the CNP architecture, we obtained the best performance when the condition vector is computed with a single ResNet block that receives the image representation via activations from the third residual chain of the main ResNet-18 model. This ”encoder” also conditions on the target classy ref with an embedding layer (lookup table) that is added to the third chain activations. The resulting procedure becomes the following: first, apply encoder to the edited sample and compute the condition vector, then add this vector to the third layer chain activations for all subsequent inputs. 5 Published as a conference paper at ICLR 2020 TrainingEditorEditableTest ErrorTest ErrorSuccessNum ProcedureFunctionLayersRateDrawdownRateSteps Baseline Training GDAll6.3%3.8%98.8%3.54 RMSPropChain 36.3%1.77%100%3.11 GDAll6.34%1.42%100%3.39 Editablec loc = 0.01GDChain 36.28%1.44%100%2.82 RMSPropChain 36.31%0.86%100%4.13 Editablec loc = 0.1RMSPropChain 37.19%0.65%100%4.76 Editable+CNP (best)Cond. vectorChain 36.33%1.06%98.9%n/a Baseline TrainingGD+EWCChain 36.3%1.92%100%3.88 Baseline TrainingRMSProp+EWCChain 36.3%1.24%98.1%4.03 DkNNk= 10Flip Labelsn/a6.36%1.76%100%n/a DkNNk= 100Flip Labelsn/a7.04%1.05%100%n/a Table 3: Editable Training of ResNet18 on CIFAR10 dataset with different editor functions. Table 3 demonstrates two advantages of Editable Training. First, withc loc =0.01it is able to reduce drawdown (compared to models trained without Editable Training) while having no significant effect on test error rate. Second, editing Chain 3 alone is almost as effective as editing the whole model. This is important because it allows us to reduce training time, making Editable Training≈2.5times slower than baseline training. Note, Editable+CNP turned out to be almost as effective as models trained with gradient-based editors while being simpler to implement. 4.2ANALYZINGEDITEDMODELS In this section, we aim to interpret the differences between the models learned with and without Editable Training. First, we investigate which inputs are most affected when the model is edited on a sample that belongs to each particular class. Based on Figure 2 (left), we conclude that edits of baseline model cause most drawdown on samples that belong to the same class as the edited input (prior to edit). However, this visualization loses information by reducing edits to their class labels. In Figure 2 (middle) we apply t-SNE (van der Maaten & Hinton (2008)) to analyze the structure of the ”edit space”. Intuitively, two edited versions of the same model are considered close if they make similar predictions. We quantify this by computing KL-divergence between the model’s pre- dictions before and after edit for each of10.000test samples. These KL divergences effectively form a10.000-dimensional model descriptor. We compute these descriptors for4.500edits applied to models trained with and without Editable Training. These vectors are then embedded in two- dimensional space with the t-SNE algorithm. We plot the obtained charts on Figure 2 (middle), with point colors denoting original class labels of edited images. As expected, the baseline edits for images of the same class are mapped to close points. In turn, Editable Training does not always follow this pattern: the edit clusters are formed based on both original and target labels with a highly interlinked region in the middle. Combined with the fact that Editable Training has a significantly lower drawdown, this lets us hypothesize that with Editable Training neural networks learn representations where edits affect objects of the same original class to a smaller extent. Conversely, the t-SNE visualization lacks information about the true dimensionality of the data manifold. To capture this property, we also perform truncated SVD decomposition of the same matrix of descriptors. Our main interest is the number of SVD components required to explain a given percentage of data variance. In Figure 2 (right) we report the explained variance ratio for models obtained with and without Editable Training. These results present evidence that Editable Training learns representations that exploit the neural network capacity to a greater extent. 6 Published as a conference paper at ICLR 2020 0123456789 Drawdown per class 0 1 2 3 4 5 6 7 8 9 Original labels Confusion matrix Baseline training Editable training 01000200030004000 1.0 0.9 0.8 0.7 5441329 Explained variance ratios baseline editable Figure 2: Edited model visualizations(Left)Confusion matrix of baseline model: rows correspond to editing images belonging to each of 10 classes; columns represent drawdowns per individual class. (Middle)t-SNE visualizations. Point color represents original class labels; brightness encodes edit targets(Right)The proportion of explained variance versus the number of components. 4.3EDITABLEFINE-TUNING FOR LARGE SCALE IMAGE CLASSIFICATION Section 4.1 demonstrates the success of Editable Training on the small CIFAR-10 dataset. However, many practical applications require training for many weeks on huge datasets. Re-training such model for the sake of better edits may be impractical. In contrast, it would be more efficient to start from a pre-trained model and fine-tune it with Editable Training. Here we experiment with the ILSVRC image classification task (Deng et al. (2009)) and consider two pre-trained architectures: smaller ResNet-18 and deeper DenseNet-169 (Huang et al. (2016)) networks. For each architecture, we start with pre-trained model weights 2 and fine-tune them on the same dataset with Editable Training. More specifically, we choose the training objectiveL base (θ) as KL-divergence between the predictions of the original network and its fine-tuned counterpart. Intuitively, this objective encourages the network to preserve its original classification behavior, while being trained to allow local edits. Similar to Section 4.1, the editor functions are only allowed to modify a subset of neural network layers. We experiment with two choices of such subsets. First, we try to edit a pre-existing layer in the network. Namely, we select the third out of four ”chains” in both architectures. In the second experiment, we augment each architecture with an extra trainable layer after the last convolutional layer. We set an extra layer to be a residual block with a 4096-unit dense layer, followed by ELU activation (Clevert et al. (2015)) and another 1024-unit dense layer. The evaluation is performed onN=1000edits with random target class. We measure the drawdown on the full ILSVRC validation set of50.000images. We use the SGD optimizer with momentum μ=0.9. We set the learning rate to10 −5 for the pre-existing layers and10 −3 for the extra block. The ImageNet training data is augmented with random resized crops and random horizontal flips. Our baselines for this task are the pre-trained architectures without Editable Fine-Tuning. However, during experiments, we noticed that minimizing the KL-divergenceL(θ)has a side-effect of im- proving validation error. We attribute this improvement to the self-distillation phenomenon (Hinton et al. (2015); Furlanello et al. (2018)). To disentangle these two effects, we consider an additional baseline where the model is trained to minimize the KL-divergence without Editable Training terms. For fair comparison, we also include baselines that edit an extra layer. This layer is initialized at random for the pre-trained models and fine-tuned for the models trained with distillation. The results in Table 4 show that Editable Training can be effectively applied in the fine-tuning scenario, achieving the best results with an extra trainable layer. In all cases Editable Fine-Tuning took under48hours on a single GeForce 1080 Ti GPU while a single edit requires less than150ms. 4.3.1REALISTICEDITTASKS WITHNATURALADVERSARIALEXAMPLES In all previous experiments, we considered edits with randomly chosen target class. However, in many practical scenarios, most of these edits will never occur. For instance, it is far more likely that an image previously classified as ”plane” would require editing into ”bird” than into ”truck” 2 We use publicly available pre-trained models fromhttps://github.com/pytorch/vision. 7 Published as a conference paper at ICLR 2020 ModelTrainingEditableTest ErrorMeanSuccessNum ArchitectureProcedureLayersRateDrawdownRateSteps ResNet18 Pre-trainedChain 330.95%3.89%99.8%3.582 Pre-trainedExtra layer30.95%9.18%100%4.272 DistillationExtra layer30.75%2.80%100%2.63 EditableChain 330.53%3.78%99.8%3.616 EditableExtra layer30.61%0.57%100%3.388 DenseNet169 Pre-trainedChain 325.49%5.20%100%2.551 Pre-trainedExtra layer25.47%9.05%100%3.874 DistillationExtra layer24.33%1.67%100%2.822 EditableChain 324.32%4.47%100%2.556 EditableExtra layer24.38%0.96%100%2.970 Table 4: Editable Training on the ImageNet dataset with RMSProp editor function. or ”ship”. To address this consideration, we employ the Natural Adversarial Examples (NAE) data set by Hendrycks et al. (2019). This data set contains7.500natural images that are particularly hard to classify with neural networks. Without edits, a pre-trained model can correctly predict less than1%of NAEs, but the correct answer is likely to be within top-100 classes ordered by predicted probabilities (see Figure 5 left). The next set of experiments quantifies Editable Training in this more realistic setting. All models are evaluated on a sample of1.000edits, each corresponding to one Natural Adversarial Example and its reference class. We measure the drawdown from each edit on50.000ILSVRC test images. We evaluate best techniques from Section 4.3 and their modifications that account for NAEs: •Editable Training: Random— model trained to edit on random targets from the uniform distribution, same as in Table 4. Compared to the same pre-trained and distilled baselines. •Editable Training: Match Ranks— model trained to edit ImageNet training images with targets sampled based on their rank under NAE rank distribution (see 5, left). •Editable Training: Train on NAE— model trained to edit6.500natural adversarial examples. These NAEs do not overlap with1.000NAE examples used for evaluation. TrainingTest Drawdown SuccessNum ProcedureErrorRateSteps Baseline Training Pre-trained30.99%4.54%100%3.822 Distillation30.75%1.62%100%2.192 Editable Training Random edits30.79%0.314%100%2.594 Match ranks30.76%0.146%100%2.149 Train on NAE30.86%0.167%100%2.236 Table 5: Editing Natural Adversarial Examples for ResNet18: (Top-Left)Editor effectiveness when editingN= 1000NAEs; (Top-Right)Reference class rank distribution for baseline model, (Bottom-Right)Error rate for edit sequences, ResNet18 baseline and Match Ranks. Pale areas indicate std. deviation over 10 runs. 02004006008001000 Rank of correct answer 0% 5% 10% Answer rank distribution on NAE 0510152025 Number of sequential edits 0.4 0.6 0.8 1.0 Test error with standard deviation pre-trained match rank The results in Table 5 (top-left) show that Editable Training significantly reduces drawdown for NAEs even when trained with random targets. However, accounting for the distribution of target 8 Published as a conference paper at ICLR 2020 classes improves locality even further. Surprisingly enough, training on6.500actual NAEs fares no better than simply matching the distribution of target ranks. For the final set of evaluations, we consider two realistic scenarios that are not covered by our previous experiments. First, we evaluate whether edits performed by our method generalize to substantially similar inputs. This behavior is highly desirable since we want to avoid the edited model repeating old mistakes in a slightly changed context. For each of1,000NAEs, we find the most similar image from test set based on InceptionV3 embeddings (Szegedy et al., 2015). For each such pair, we edit 5 augmentations of the first image and measure how often the model predicts the edited class on 5 augmentations of the second image. A model trained with random edits has an accuracy of 86.7% while ”Editable + Match Ranks” scores 85.9% accuracy. Finally, we evaluate if our technique can perform multiple edits in a sequence. Figure 5 (bottom-left) demonstrates that our approach can cope with sequential edits without ever being trained that way. 4.4EDITABLETRAINING FORMACHINETRANSLATION The previous experiments focused on multi-class classification problems. However, Editable Train- ing can be applied to any task where the model is trained by minimizing a differentiable objective. Our final set of experiments demonstrates the applicability of Editable Training for machine transla- tion. We consider the IWSLT 2014 German-English translation task with the standard training/test splits (Cettolo et al. (2015)). The data is preprocessed with Moses Tokenizer (Koehn et al. (2007)) and converted to lowercase. We further apply the Byte-Pair Encoding with10.000BPE rules learned jointly from German and English training data. Finally, we train the Transformer (Vaswani et al. (2017)) model similar to transformer-base configuration, optimized for IWSLT De-En task 3 . Typical machine translation models use beam search to find the most likely translation. Hence we consider an edit to be successful if and only if the log-probability of target translation is greater than log-probability of any alternative translation. So,l e ( ˆ θ) = max y i logp(y i |s, ˆ θ)−logp(y 0 |s, ˆ θ), wheresis a source sentence,y 0 denotes target translation andy i k i=1 are alternative translations. During training, we approximate this by findingk=32most likely translations with beam search using the Transformer model trained normally on the same data. The edit targets are sampled from the same model by sampling with temperatureτ=1.2. The resulting edit consists of three parts: a source sentence, a target translation and a set of alternative translations. We defineL loc as KL-divergence between the predictions of the original and edited model averaged over target tokens,L loc =E x,y∈D 1 |y| ∑ t D KL (p(y t |x,y 0:t ,θ)||p(y t |x,y 0:t Edit k α (θ,l e ))), whereD is a data batch,xandyare the source and translation phrases respectively,y 0:t denotes a translation prefix. TheEditfunction optimizes the final decoder layer using RMSProp with hyperparameters tuned as in Section 4.1. The results in Table 6 show that Editable Training produces a model that matches the baseline translation quality but has less than half of its drawdown. Training ProcedureTest BLEUBLEU DrawdownSuccess rateNum Steps Baseline training,α=10 −3 34.770.76100%2.35 Editable,c loc =100,α=10 −3 34.800.35100%3.07 Editable,c loc =100,α=3·10 −4 34.810.17100%5.5 Table 6: Evaluation of editable Transformer models on IWSLT14 German-English translation task. 5CONCLUSION In this paper we have addressed the efficient correction of neural network mistakes, a highly im- portant task for deep learning practitioners. We have proposed several evaluation measures for comparison of different means of model correction. Then we have introduced Editable Training, a training procedure that produces models that allow gradient-based editing to address corrections of the model behaviour. We demonstrate the advantage of Editable Training against reasonable baselines on large-scale image classification and machine translation tasks. 3 We use Transformer configuration ”transformer iwsltdeen” from Fairseq v0.8.0 (Ott et al. (2019)) 9 Published as a conference paper at ICLR 2020 ACKNOWLEDGMENTS We would like to thank Andrey Voynov for many useful discussions which helped inspire the idea for this study. We also wish to express our sincere appreciation to Pavel Bogomolov for constructive criticism and for his diligent proofreading of this paper. REFERENCES Marcin Andrychowicz, Misha Denil, Sergio Gomez Colmenarejo, Matthew W. Hoffman, David Pfau, Tom Schaul, and Nando de Freitas. Learning to learn by gradient descent by gradient descent.ArXiv, abs/1606.04474, 2016. Jeremy Bernstein, Yu-Xiang Wang, Kamyar Azizzadenesheli, and Animashree Anandkumar. Signsgd: Compressed optimisation for non-convex problems. In Jennifer G. Dy and Andreas Krause (eds.),ICML, volume 80 ofProceedings of Machine Learning Research, p. 559– 568. PMLR, 2018. URLhttp://dblp.uni-trier.de/db/conf/icml/icml2018. html#BernsteinWAA18. Mauro Cettolo, Jan Niehues, Sebastian St ̈ uker, Luisa Bentivogli, and Marcello Federico. Report on the 11 th iwslt evaluation campaign , iwslt 2014. 2015. Djork-Arn ́ e Clevert, Thomas Unterthiner, and Sepp Hochreiter. Fast and accurate deep network learning by exponential linear units (elus).CoRR, abs/1511.07289, 2015. J. Schmidhuber D. C. Ciresan, U. Meier. First superhuman visual pattern recognition.IJCNN, 2011. J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. InCVPR09, 2009. John C. Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization.J. Mach. Learn. Res., 12:2121–2159, 2010. Javid Ebrahimi, Anyi Rao, Daniel Lowd, and Dejing Dou. Hotflip: White-box adversarial examples for text classification. InACL, 2017. Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. InICML, 2017. Tommaso Furlanello, Zachary Chase Lipton, Michael Tschannen, Laurent Itti, and Anima Anand- kumar. Born-again neural networks. InICML, 2018. Marta Garnelo, Dan Rosenbaum, Chris J. Maddison, Tiago Ramalho, David Saxton, Murray Shana- han, Yee Whye Teh, Danilo Jimenez Rezende, and S. M. Ali Eslami. Conditional neural pro- cesses.ArXiv, abs/1807.01613, 2018. Ian Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. InInternational Conference on Learning Representations, 2015. URLhttp:// arxiv.org/abs/1412.6572. Hany Hassan, Anthony Aue, Chang Chen, Vishal Chowdhary, Jonathan R. Clark, Christian Fed- ermann, Xuedong Huang, Marcin Junczys-Dowmunt, William Lewis, Mu Li, Shujie Liu, T. M. Liu, Renqian Luo, Arul Menezes, Tao Qin, Frank Seide, Xu Tan, Fei Tian, Lijun Wu, Shuangzhi Wu, Yingce Xia, Dongdong Zhang, Zhirui Zhang, and Ming Zhou. Achieving human parity on automatic chinese to english news translation.ArXiv, abs/1803.05567, 2018. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog- nition.2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), p. 770– 778, 2015. Dan Hendrycks, Kevin Keliang Zhao, Steven Basart, Jacob Steinhardt, and Dawn Xiaodong Song. Natural adversarial examples.ArXiv, abs/1907.07174, 2019. 10 Published as a conference paper at ICLR 2020 Alex Hern. Facebook translates “good morning” into “attack them”, leading to arrest.The Guardian, 2018. Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the knowledge in a neural network. ArXiv, abs/1503.02531, 2015. Rein Houthooft, Yuhua Chen, Phillip Isola, Bradly C. Stadie, Filip Wolski, Jonathan Ho, and Pieter Abbeel. Evolved policy gradients.ArXiv, abs/1802.04821, 2018. Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks.2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), p. 2261–2269, 2016. Diederik P. Kingma and Jimmy Ba.Adam: A method for stochastic optimization.CoRR, abs/1412.6980, 2014. James Kirkpatrick, Razvan Pascanu, Neil C. Rabinowitz, Joel Veness, Guillaume Desjardins, An- drei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic for- getting in neural networks.Proceedings of the National Academy of Sciences of the United States of America, 114 13:3521–3526, 2016. Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, Chris Dyer, Ond ˇ rej Bo- jar, Alexandra Constantin, and Evan Herbst. Moses: Open source toolkit for statistical ma- chine translation. InProceedings of the 45th Annual Meeting of the Association for Com- putational Linguistics Companion Volume Proceedings of the Demo and Poster Sessions, p. 177–180, Prague, Czech Republic, June 2007. Association for Computational Linguistics. URL https://w.aclweb.org/anthology/P07-2045. Alex Krizhevsky, Vinod Nair, and Geoffrey Hinton. Cifar-10 (canadian institute for advanced re- search). URLhttp://w.cs.toronto.edu/ ̃ kriz/cifar.html. Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton.Imagenet classification with deep convolutional neural networks.In F. Pereira, C. J. C. Burges, L. Bottou, and K. Q. Weinberger (eds.),Advances in Neural Information Processing Systems 25, p. 1097– 1105. Curran Associates, Inc., 2012.URLhttp://papers.nips.c/paper/ 4824-imagenet-classification-with-deep-convolutional-neural-networks. pdf. Min Lin, Jie Fu, and Yoshua Bengio. Conditional computation for continual learning.ArXiv, abs/1906.06635, 2019. Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms.ArXiv, abs/1803.02999, 2018. Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier, and Michael Auli. fairseq: A fast, extensible toolkit for sequence modeling. InNAACL-HLT, 2019. Nicolas Papernot and Patrick D. McDaniel. Deep k-nearest neighbors: Towards confident, inter- pretable and robust deep learning.ArXiv, abs/1803.04765, 2018. Roger Ratcliff. Connectionist models of recognition memory: constraints imposed by learning and forgetting functions.Psychological review, 97(2):285, 1990. Anthony V. Robins. Catastrophic forgetting, rehearsal and pseudorehearsal.Connect. Sci., 7:123– 146, 1995. David Silver, Aja Huang, Chris J. Maddison, Arthur Guez, Laurent Sifre, George van den Driess- che, Julian Schrittwieser, Ioannis Antonoglou, Vedavyas Panneershelvam, Marc Lanctot, Sander Dieleman, Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya Sutskever, Timothy P. Lillicrap, Madeleine Leach, Koray Kavukcuoglu, Thore Graepel, and Demis Hassabis. Mastering the game of go with deep neural networks and tree search.Nature, 529:484–489, 2016. 11 Published as a conference paper at ICLR 2020 Leslie N. Smith and Nicholay Topin. Super-convergence: Very fast training of residual networks using large learning rates.ArXiv, abs/1708.07120, 2017. Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfel- low, and Rob Fergus. Intriguing properties of neural networks.CoRR, abs/1312.6199, 2013. Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision.2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), p. 2818–2826, 2015. T. Tieleman and G. Hinton. Lecture 6.5—RmsProp: Divide the gradient by a running average of its recent magnitude. COURSERA: Neural Networks for Machine Learning, 2012. Laurens van der Maaten and Geoffrey E. Hinton. Visualizing data using t-sne. 2008. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. Attention is all you need. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett (eds.),Advances in Neu- ral Information Processing Systems 30, p. 5998–6008. Curran Associates, Inc., 2017. URL http://papers.nips.c/paper/7181-attention-is-all-you-need.pdf. Eric Wallace, Feng Shi, Nikhil Kandpal, Matt Gardner, and Sameer Singh. Universal adversarial triggers for attacking and analyzing nlp. 2019. Xiaoyong Yuan, Pan He, Qile Zhu, and Xiaolin Li. Adversarial examples: Attacks and defenses for deep learning.IEEE Transactions on Neural Networks and Learning Systems, 30:2805–2824, 2017. Matthew D. Zeiler. Adadelta: An adaptive learning rate method.ArXiv, abs/1212.5701, 2012. 12

Instant research discovery