Subset (19)

iclr23 · 14k rows

Split (3)

train · 8.4k rows

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.

review stringlengths 5 16.6k	score stringclasses 5 values	area stringclasses 12 values	text stringlengths 31 5.65k
This paper focuses on what it refers to as "selection collider bias", using a causal inference framework to discuss spurious correlations in language data. The paper seems largely focused on formalizing notions of selection bias and spurious correlations through the language of causal inference, and presenting reasonin...	Recommendation: 3: reject, not good enough	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: This paper focuses on what it refers to as "selection collider bias", using a causal inference framework to discuss spurious correlations in language data. The paper seems largely focused on formalizing notions of selection bias and spurious correlations through the language of causal inference, and presenting ...
This paper proposes a meta-learning-based feature selection framework that aims to select effective features for machine learning tasks. Specifically, this wrapper framework consists of two components, i.e., Feature Subset Sampler (FSS) which optimizes the discrete search space in a continuous way, and Meta Feature Est...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a meta-learning-based feature selection framework that aims to select effective features for machine learning tasks. Specifically, this wrapper framework consists of two components, i.e., Feature Subset Sampler (FSS) which optimizes the discrete search space in a continuous way, and Meta Fea...
This paper proposes a self-supervised learning (SSL) framework for tabular data, Masked Encoding for Tabular Data (MET). MET employs masked auto-encoding with Transformer architectures. This paper shows that MET outperforms existing SSL methods for tabular data and classical machine learning approaches. In addition, th...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper proposes a self-supervised learning (SSL) framework for tabular data, Masked Encoding for Tabular Data (MET). MET employs masked auto-encoding with Transformer architectures. This paper shows that MET outperforms existing SSL methods for tabular data and classical machine learning approaches. In addi...
This paper try to show that common state-of-the-art methods in continual learning suffer from substantial forgetting upon starting to learn new tasks, except that this forgetting is temporary and followed by a phase of performance recovery. The main contributions can be summarized as : (1) This work defines a framewor...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper try to show that common state-of-the-art methods in continual learning suffer from substantial forgetting upon starting to learn new tasks, except that this forgetting is temporary and followed by a phase of performance recovery. The main contributions can be summarized as : (1) This work defines a ...
This paper proposes new learned state representations that can improve exploration by introducing time predictive semantics into the state representation. The authors propose a new notion of novelty based on time prediction representations and use this to create a new insertion criterion, a new cell count strategy and ...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes new learned state representations that can improve exploration by introducing time predictive semantics into the state representation. The authors propose a new notion of novelty based on time prediction representations and use this to create a new insertion criterion, a new cell count strat...
In this work a model for sequences with the following properties is proposed: - Modular description of the transition function - Handling of interventions, defined as the modification of a target of the transition function - Learning of the dependency structure - Use of latent variables The authors build on top of t...	Recommendation: 6: marginally above the acceptance threshold	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: In this work a model for sequences with the following properties is proposed: - Modular description of the transition function - Handling of interventions, defined as the modification of a target of the transition function - Learning of the dependency structure - Use of latent variables The authors build on ...
This paper is based on the insight that anomaly detection can be treated as a causal discovery problem, to both improve detection and to facilitate root cause analysis. An anomaly is defined as a datum that does not fit the discovered data-generating process, in short an instance where the process is non-stationary. Th...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper is based on the insight that anomaly detection can be treated as a causal discovery problem, to both improve detection and to facilitate root cause analysis. An anomaly is defined as a datum that does not fit the discovered data-generating process, in short an instance where the process is non-statio...
This paper studied the problem of learning a Lipschitz continuous function on a manifold. The authors used the Lagrangian dual problem and showed that its empirical version can be an approximation of the primal with statistical consistency. The authors further showed that, by using a weighted point cloud Laplacian, the...	Recommendation: 5: marginally below the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper studied the problem of learning a Lipschitz continuous function on a manifold. The authors used the Lagrangian dual problem and showed that its empirical version can be an approximation of the primal with statistical consistency. The authors further showed that, by using a weighted point cloud Laplac...
This paper proposes a self-supervised model for video object discovery. It only takes consecutive RGB frames as input instead of optical flows as in previous works. The model is trained to reconstruct the optical flow between any paired frames that generated through an off-the-shelf optical flow estimator. A temporal c...	Recommendation: 5: marginally below the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper proposes a self-supervised model for video object discovery. It only takes consecutive RGB frames as input instead of optical flows as in previous works. The model is trained to reconstruct the optical flow between any paired frames that generated through an off-the-shelf optical flow estimator. A te...
The paper presents a new training methodology, HRAT, for neural networks to be deployed on Memristor (RRAM) based hardware. The presented methodology accounts for non-idealities such as Quantization error for weights and limited output swing of operational amplifier circuits. Additionally, the methodology also covers t...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: The paper presents a new training methodology, HRAT, for neural networks to be deployed on Memristor (RRAM) based hardware. The presented methodology accounts for non-idealities such as Quantization error for weights and limited output swing of operational amplifier circuits. Additionally, the methodology also ...
This paper address the problem of GNNs being vulnerable to imperceptible adversarial attacks and also has some issues in generalizing out-of-distribution data. The proposed solution, graph predictive coding network (GPCN), uses a novel message-passing scheme developed based on the theory of predictive coding. The autho...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper address the problem of GNNs being vulnerable to imperceptible adversarial attacks and also has some issues in generalizing out-of-distribution data. The proposed solution, graph predictive coding network (GPCN), uses a novel message-passing scheme developed based on the theory of predictive coding. T...
The paper presents simplified closed-form estimates for the test risk and other generalization metrics of kernel ridge regression. Compared to prior work, the authors claim their derivations are greatly simplified and lead to higher interpretability. Test risk and other objects of interest are expressed in a transpar...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper presents simplified closed-form estimates for the test risk and other generalization metrics of kernel ridge regression. Compared to prior work, the authors claim their derivations are greatly simplified and lead to higher interpretability. Test risk and other objects of interest are expressed in a ...
A method for posterior inference with a variant of GBDT is proposed. The weak learner used in this paper is a variant of the standard decision tree, where decision rules are oblivious (same split used at all nodes of the tree at the same level), the splitting procedure is randomized by use of the max-Gumbel trick to sa...	Recommendation: 6: marginally above the acceptance threshold	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: A method for posterior inference with a variant of GBDT is proposed. The weak learner used in this paper is a variant of the standard decision tree, where decision rules are oblivious (same split used at all nodes of the tree at the same level), the splitting procedure is randomized by use of the max-Gumbel tri...
Similar to as is done in done in diffusion models for U-net layers, this paper applies time-conditioning to FNO layers and demonstrates it applicability on a wide variety of tasks. The paper also proposes a method to normalize the Fourier kernel so that it's more perturbation resistant. Strengths Time-conditioning...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: Similar to as is done in done in diffusion models for U-net layers, this paper applies time-conditioning to FNO layers and demonstrates it applicability on a wide variety of tasks. The paper also proposes a method to normalize the Fourier kernel so that it's more perturbation resistant. Strengths Time-cond...
This paper proposes a distribution-dependent stage-aware ranking score (DDSAR-Score) for efficient face detection (FD), which can explicitly characterize the stage-level expressivity and identify the individual importance of each stage, thus satisfying the model structure design criterion of the FD backbone. As an accu...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a distribution-dependent stage-aware ranking score (DDSAR-Score) for efficient face detection (FD), which can explicitly characterize the stage-level expressivity and identify the individual importance of each stage, thus satisfying the model structure design criterion of the FD backbone. As...
This paper proposes to integrate an SSL auxiliary objective with standard model-free MARL methods (specifically MAT in this paper) to reap the same benefits as CV and single-agent RL has seen with these types of approaches. Because each agent in a typical MARL setup receives a partial observation of the world, in order...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes to integrate an SSL auxiliary objective with standard model-free MARL methods (specifically MAT in this paper) to reap the same benefits as CV and single-agent RL has seen with these types of approaches. Because each agent in a typical MARL setup receives a partial observation of the world, ...
This work aims to improve the efficiency and robustness of knowledge augmented LM. The intuition is for LM to decide when an external knowledge source is needed. A metric Thrust is developed for this decision, based on the relationship between the query embedding and the clusters of instance embeddings. 1) representa...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This work aims to improve the efficiency and robustness of knowledge augmented LM. The intuition is for LM to decide when an external knowledge source is needed. A metric Thrust is developed for this decision, based on the relationship between the query embedding and the clusters of instance embeddings. 1) re...
The paper studies phase transition for detecting communities under the degree-corrected block model (DCBM). The motivation is that chi-square test is not good for DCBM because of artifacts, so they use the the Signed-Quadrilateral (SgnQ) instead. The authors provide the SgnQ efficient regime, the computationally infeas...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper studies phase transition for detecting communities under the degree-corrected block model (DCBM). The motivation is that chi-square test is not good for DCBM because of artifacts, so they use the the Signed-Quadrilateral (SgnQ) instead. The authors provide the SgnQ efficient regime, the computationall...
Paper proposes a method, MUST, for unsupervised adaptation of a zero shot classifier. A pre-trained open world model CLIP is taken to produce classification embeddings for a vocabulary of words. These are then use to kick-start a training method that adapts a zero-shot classification model by a combination of three obj...	Recommendation: 8: accept, good paper	Area: Unsupervised and Self-supervised learning	Review: Paper proposes a method, MUST, for unsupervised adaptation of a zero shot classifier. A pre-trained open world model CLIP is taken to produce classification embeddings for a vocabulary of words. These are then use to kick-start a training method that adapts a zero-shot classification model by a combination of t...
The authors propose a zero-cost evaluation metric to improve zero-shot neural architecture search efficiency in this work. This work uses gradient distribution and network information at initialization for scoring candidate models and achieves a strong rank consistency with the performance of candidate models. Extensiv...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The authors propose a zero-cost evaluation metric to improve zero-shot neural architecture search efficiency in this work. This work uses gradient distribution and network information at initialization for scoring candidate models and achieves a strong rank consistency with the performance of candidate models. ...
The paper under consideration proposes a “Boomerang” approach for image variability-related applications. The idea is quite simple. Given a standard diffusion model: - run forward diffusion for several times resulting in noised image - run backward diffusion for several times resulting in an image which is controllably...	Recommendation: 3: reject, not good enough	Area: Generative models	Review: The paper under consideration proposes a “Boomerang” approach for image variability-related applications. The idea is quite simple. Given a standard diffusion model: - run forward diffusion for several times resulting in noised image - run backward diffusion for several times resulting in an image which is cont...
This paper proposes a new lossy image compression paradigm based on unconditional diffusion models. The noise corrupted image is compressed and sent to the decoder via reverse channel coding, and the reconstruction of the input image is done through the denoising process applied to the received noise corrupted image. E...	Recommendation: 5: marginally below the acceptance threshold	Area: Generative models	Review: This paper proposes a new lossy image compression paradigm based on unconditional diffusion models. The noise corrupted image is compressed and sent to the decoder via reverse channel coding, and the reconstruction of the input image is done through the denoising process applied to the received noise corrupted ...
First, the authors built a dataset consisting of neural network checkpoints that perform specific tasks, such as MNIST classification, including its loss and error. The proposed data-driven optimizer, G.pt, is a neural network model trained using this dataset. Specifically, if the parameters of the initial checkpoint, ...	Recommendation: 5: marginally below the acceptance threshold	Area: Generative models	Review: First, the authors built a dataset consisting of neural network checkpoints that perform specific tasks, such as MNIST classification, including its loss and error. The proposed data-driven optimizer, G.pt, is a neural network model trained using this dataset. Specifically, if the parameters of the initial chec...
Training deep learning models is becoming increasingly challenging: there is simply not enough memory to store the activations generated by the forward pass until the corresponding gradients are computed. The paper proposes to split each activations in two tensors, $o$ and $a$. The $o$ tensor can be discarded quickly, ...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: Training deep learning models is becoming increasingly challenging: there is simply not enough memory to store the activations generated by the forward pass until the corresponding gradients are computed. The paper proposes to split each activations in two tensors, $o$ and $a$. The $o$ tensor can be discarded q...
This paper aims at achieving robustness in uncertain environments. The proposed approach, a mean-standard deviation formulation, seemingly reduces the desired level of robustness to one parameter tuning. Experiments are conducted on discrete and continuous control domains. The idea of relating distributional RL for th...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper aims at achieving robustness in uncertain environments. The proposed approach, a mean-standard deviation formulation, seemingly reduces the desired level of robustness to one parameter tuning. Experiments are conducted on discrete and continuous control domains. The idea of relating distributional R...
The paper uses the NTK to discuss the effect of momentum on the training success of PINNs. It is shown that momentum (or adaptive momentum) can improve the convergence speed compared to standard SGD-based training, using both theoretical and empirical arguments. The paper is well-written and covers an interesting and t...	Recommendation: 6: marginally above the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The paper uses the NTK to discuss the effect of momentum on the training success of PINNs. It is shown that momentum (or adaptive momentum) can improve the convergence speed compared to standard SGD-based training, using both theoretical and empirical arguments. The paper is well-written and covers an interesti...
This paper introduces zero-label prompt selection (ZPS), which selects the optimal prompt for an unseen task without any labeled data or gradient update. ZPS obtains pseudo labels of an unseen task through prompt ensemble, calculates pseudo accuracy by comparing pseudo labels with the prediction of each prompt, and sel...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper introduces zero-label prompt selection (ZPS), which selects the optimal prompt for an unseen task without any labeled data or gradient update. ZPS obtains pseudo labels of an unseen task through prompt ensemble, calculates pseudo accuracy by comparing pseudo labels with the prediction of each prompt,...
In this paper, the authors proposed a mixture-of-memory augmentation framework for a zero-shot dense retrieval (MoMA-DR). The key idea is to enrich the query representation with external resources such as wiki, knowledge graphs, etc. The framework consists of two components: a retrieval model and an augmentation model....	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: In this paper, the authors proposed a mixture-of-memory augmentation framework for a zero-shot dense retrieval (MoMA-DR). The key idea is to enrich the query representation with external resources such as wiki, knowledge graphs, etc. The framework consists of two components: a retrieval model and an augmentatio...
Extreme classification (or deep information retrieval) has been a popular research field that matches an input text (query) to a label that often is also a text. This paper focuses on a sub-field where both the input text and the label text are short. The proposed training algorithm is a data augmentation algorithm tha...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: Extreme classification (or deep information retrieval) has been a popular research field that matches an input text (query) to a label that often is also a text. This paper focuses on a sub-field where both the input text and the label text are short. The proposed training algorithm is a data augmentation algor...
This paper discovers an interesting phenomenon that the SR network is powerful in discriminating the image degradations instead of image contents, especially for well-trained deep networks with global residual and generative adversarial training. To validate the claim, the authors give some empirical evidence by visual...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper discovers an interesting phenomenon that the SR network is powerful in discriminating the image degradations instead of image contents, especially for well-trained deep networks with global residual and generative adversarial training. To validate the claim, the authors give some empirical evidence b...
This paper proposes multi-agent alternate Q-learning (MA2QL) to tackle the non-stationarity problem in decentralized MARL. The main difference between MA2QL and existing method IQL is that MA2QL agents take turns to update their Q-functions while IQL agents simultaneously update their Q-functions. The authors prove tha...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper proposes multi-agent alternate Q-learning (MA2QL) to tackle the non-stationarity problem in decentralized MARL. The main difference between MA2QL and existing method IQL is that MA2QL agents take turns to update their Q-functions while IQL agents simultaneously update their Q-functions. The authors p...
The authors introduce the notion of proportional multicalibration (PMC): a criteria that measures whether a model is both multicalibrated (MC) and fair (in the sense of differentiable calibration (DC) which is an extension of sufficiency). As a reminder, multicalibration refers to a notion of calibration by group a...	Recommendation: 3: reject, not good enough	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The authors introduce the notion of proportional multicalibration (PMC): a criteria that measures whether a model is both multicalibrated (MC) and fair (in the sense of differentiable calibration (DC) which is an extension of sufficiency). As a reminder, multicalibration refers to a notion of calibration by...
This work proposes a differentiable logic programming framework that aims to learn first-order logic structures and weights. This framework further supports probabilistic reasoning by performing probabilistic forward chaining. Some theoretical guarantees on error rate and solution optimality are shown. Empirical evalua...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: This work proposes a differentiable logic programming framework that aims to learn first-order logic structures and weights. This framework further supports probabilistic reasoning by performing probabilistic forward chaining. Some theoretical guarantees on error rate and solution optimality are shown. Empirica...
This paper studies the adversarial robustness of MARL, with a focus on observation attacks. The proposed attacker can decide both which agent to attack and how to perturb its observation, via a hybrid-action attacker. Then the paper proposes a robust training approach by alternately training the attacker and the agents...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper studies the adversarial robustness of MARL, with a focus on observation attacks. The proposed attacker can decide both which agent to attack and how to perturb its observation, via a hybrid-action attacker. Then the paper proposes a robust training approach by alternately training the attacker and th...
This paper proposes a network dissection method without using any labelled image data. This is achieved by utilizing a pretrained CLIP model to compute an image-concept similarity matrix. With the help of the matrix, the concept label for a neuron unit will be the one that maximizes the similarity between the neuron qu...	Recommendation: 6: marginally above the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper proposes a network dissection method without using any labelled image data. This is achieved by utilizing a pretrained CLIP model to compute an image-concept similarity matrix. With the help of the matrix, the concept label for a neuron unit will be the one that maximizes the similarity between the n...
The paper introduces ROSMO, a simpler and better version of Muzero Unplugged for offline policy optimization. In the offline setting, Muzero Unplugged can suffer from limited data-coverage, inaccurate models and an expensive compute budget. Instead of relying on an expensive MCTS and suffering from out of distribution ...	Recommendation: 8: accept, good paper	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper introduces ROSMO, a simpler and better version of Muzero Unplugged for offline policy optimization. In the offline setting, Muzero Unplugged can suffer from limited data-coverage, inaccurate models and an expensive compute budget. Instead of relying on an expensive MCTS and suffering from out of distr...
This paper proposes a new prompting method called algorithmic prompting that aims to teach algorithms to LLMs via in-context learning. Specifically, the method focuses on using very details execution traces and explanations as part of the in-context demonstration to remove any ambiguity. Using this prompting method, in...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper proposes a new prompting method called algorithmic prompting that aims to teach algorithms to LLMs via in-context learning. Specifically, the method focuses on using very details execution traces and explanations as part of the in-context demonstration to remove any ambiguity. Using this prompting me...
This paper studies the problem of zero-shot adversarial robustness: adapting pretrained large-scale vision-language models to unseen target tasks with high robust accuracies. With the conjecture that language encoder plays an important role in achieving good zero-shot generalization ability, a contrastive based adversa...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper studies the problem of zero-shot adversarial robustness: adapting pretrained large-scale vision-language models to unseen target tasks with high robust accuracies. With the conjecture that language encoder plays an important role in achieving good zero-shot generalization ability, a contrastive based...
The paper tries to shed light on the cause and nature of simplicity bias and proposes a regularization strategy to mitigate the same. While we know that deep neural networks (DNNs) are prone to learning simple features (simplicity bias), recent work provides evidence that the penultimate features in the DNNs contain ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper tries to shed light on the cause and nature of simplicity bias and proposes a regularization strategy to mitigate the same. While we know that deep neural networks (DNNs) are prone to learning simple features (simplicity bias), recent work provides evidence that the penultimate features in the DNNs ...
The authors propose training feature extractor for Few show segmentation. To suppress the sample level, region levle and patch level heterogeneity between samples, the author propose Multi-level Heterogeneity Suppressing extractor (MuHS). MuHS uses attention to suppress the heterogeneities by reinforcing the attention...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The authors propose training feature extractor for Few show segmentation. To suppress the sample level, region levle and patch level heterogeneity between samples, the author propose Multi-level Heterogeneity Suppressing extractor (MuHS). MuHS uses attention to suppress the heterogeneities by reinforcing the a...