review stringlengths 5 16.6k | score stringclasses 5
values | area stringclasses 12
values | text stringlengths 31 5.65k |
|---|---|---|---|
This paper proposes a exploration method, called deep evidential reinforcement learning (DERL) to improve the exploration ability of RLRS. The DERL algorithm uses recurrent neural network to represent the dynamic feature of the user state, and conducts evidential-actor-critic module to enable better exploration. Expe... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a exploration method, called deep evidential reinforcement learning (DERL) to improve the exploration ability of RLRS. The DERL algorithm uses recurrent neural network to represent the dynamic feature of the user state, and conducts evidential-actor-critic module to enable better explorati... |
This paper proposes a class of regularization objectives that are used to improve the effectiveness of parameter-efficient tuning methods (PETs) for pre-trained language models. Specifically, the objective consists of two main components: (i) a (potentially learnable) diffusion bridge defining a target diffusion proces... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a class of regularization objectives that are used to improve the effectiveness of parameter-efficient tuning methods (PETs) for pre-trained language models. Specifically, the objective consists of two main components: (i) a (potentially learnable) diffusion bridge defining a target diffusio... |
The paper proposes Graphair, an automated graph data augmentation method for fair graph representation learning. Graphair is designed to automatically discover fairness-aware augmentations from input graphs in order to circumvent sensitive information while preserving other informative features. Adversarial learning an... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper proposes Graphair, an automated graph data augmentation method for fair graph representation learning. Graphair is designed to automatically discover fairness-aware augmentations from input graphs in order to circumvent sensitive information while preserving other informative features. Adversarial lea... |
This paper extends the existing distributionally robust optimization toward the "unsupervised" case, where group information is not given in advance. Following the basic formulation of G-DRO (minimization of the loss function on the worst group), the authors incorporate a grouper model to infer the underlying (unreveal... | 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 extends the existing distributionally robust optimization toward the "unsupervised" case, where group information is not given in advance. Following the basic formulation of G-DRO (minimization of the loss function on the worst group), the authors incorporate a grouper model to infer the underlying (... |
The paper compares the ReLU activation function to its many smooth variants that converge to it in the limit of low temperatures. The paper proposes to investigate the role of these two sub-classes of activation functions by looking at their gradient propagation at initialization time. In particular, prior work has sho... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper compares the ReLU activation function to its many smooth variants that converge to it in the limit of low temperatures. The paper proposes to investigate the role of these two sub-classes of activation functions by looking at their gradient propagation at initialization time. In particular, prior work... |
This paper focuses on a version of a continual learning problem where for each continual task a learner is given a set of labeled instance, and a set of unlabeled instances in which some are relevant to the task and some are not. The authors make the claim that this version of the continual learning task is most reali... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper focuses on a version of a continual learning problem where for each continual task a learner is given a set of labeled instance, and a set of unlabeled instances in which some are relevant to the task and some are not. The authors make the claim that this version of the continual learning task is mo... |
In this paper, the authors consider the problem of optimal partial transport. In this problem some fraction of a mass distribution over a set of supply nodes, needs to be transported and distributed to a subset of the demand nodes. The amount of mass supplied or demanded by each node as well as the per unit transportat... | Recommendation: 6: marginally above the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
In this paper, the authors consider the problem of optimal partial transport. In this problem some fraction of a mass distribution over a set of supply nodes, needs to be transported and distributed to a subset of the demand nodes. The amount of mass supplied or demanded by each node as well as the per unit tra... |
The authors use the 2-divergence (Chi^2) as the objective function for sampling target distribution. And they parameterize the distribution by a normalizing flow. They use importance sampling ideas to approximate the Chi^2 divergence. By solving the proposed optimization method numerically, they directly simulate the G... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The authors use the 2-divergence (Chi^2) as the objective function for sampling target distribution. And they parameterize the distribution by a normalizing flow. They use importance sampling ideas to approximate the Chi^2 divergence. By solving the proposed optimization method numerically, they directly simula... |
The paper proposes a new semi-parametric learning method called SPIN. It builds upon non-parametric transformers (NPT), which use attention between training points to learn compact and effective predictors at the cost of quadratic runtime in the number of training points considered (context size). SPIN addresses this b... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The paper proposes a new semi-parametric learning method called SPIN. It builds upon non-parametric transformers (NPT), which use attention between training points to learn compact and effective predictors at the cost of quadratic runtime in the number of training points considered (context size). SPIN addresse... |
This paper proposed a new self-supervised learning model for video representation by introducing the idea of saccades/fixations. More specifically, a semantic change aware contrastive learning is proposed such that a positive pair is from the same fixation region of the same video, followed by reorganization of prototy... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper proposed a new self-supervised learning model for video representation by introducing the idea of saccades/fixations. More specifically, a semantic change aware contrastive learning is proposed such that a positive pair is from the same fixation region of the same video, followed by reorganization of... |
This paper presents ROSMO, a model-based offline RL algorithm based on MuZero Unplugged (MZU). ROSMO improves MZU in the offline RL setting in two main ways. First, it only performs one-step rollout for advantage estimation, rather than MCTS with multi-step rollouts. Second, it uses advantage-based policy improvement w... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper presents ROSMO, a model-based offline RL algorithm based on MuZero Unplugged (MZU). ROSMO improves MZU in the offline RL setting in two main ways. First, it only performs one-step rollout for advantage estimation, rather than MCTS with multi-step rollouts. Second, it uses advantage-based policy impro... |
This paper proposes a method for multitask learning of NLP tasks with compositional codes. Given a task id, a sequence of discrete codes are obtained. Code vectors obtained from these codes using an embedding table are fed as input to the language model as a task representation. The codebook and the language model are ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a method for multitask learning of NLP tasks with compositional codes. Given a task id, a sequence of discrete codes are obtained. Code vectors obtained from these codes using an embedding table are fed as input to the language model as a task representation. The codebook and the language mo... |
The paper describes a network that combines a Laplace framework into a reinforcement learning architecture achieve superior performance over GRU and RNN in an evidence accumulation task both in terms of reward gain and generalization. Moreover, the activity of units in the proposed architecture resembles map-like neura... | Recommendation: 5: marginally below the acceptance threshold | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
The paper describes a network that combines a Laplace framework into a reinforcement learning architecture achieve superior performance over GRU and RNN in an evidence accumulation task both in terms of reward gain and generalization. Moreover, the activity of units in the proposed architecture resembles map-li... |
This paper presents a novel predictor-corrector algorithm for time-varying stochastic optimization. It exploits the continuity of the domain shift. This work demonstrates the superiority of the proposed algorithm over stochastic gradient descent both theoretically and empirically. Further, this works shows that a simpl... | Recommendation: 6: marginally above the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper presents a novel predictor-corrector algorithm for time-varying stochastic optimization. It exploits the continuity of the domain shift. This work demonstrates the superiority of the proposed algorithm over stochastic gradient descent both theoretically and empirically. Further, this works shows that... |
This paper provides a holistic theoretical framework to evaluate stability and transferability of graph convolutional networks (GCN). As expected, the various theoretical results provide bounds on the changes in GCN outputs in terms of perturbations to the input signal or the structure. A novel perspective on the trans... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper provides a holistic theoretical framework to evaluate stability and transferability of graph convolutional networks (GCN). As expected, the various theoretical results provide bounds on the changes in GCN outputs in terms of perturbations to the input signal or the structure. A novel perspective on t... |
This paper describes PMxiUp, a new way to add to text data augmentation by replacing parts of speech (POS) and interpolating features. The proposed method swaps out tokens that belong to a particular POS tag and uses feature space interpolation in order. The author discovered through experimentation that nouns are esse... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper describes PMxiUp, a new way to add to text data augmentation by replacing parts of speech (POS) and interpolating features. The proposed method swaps out tokens that belong to a particular POS tag and uses feature space interpolation in order. The author discovered through experimentation that nouns ... |
This paper proposes another variant of Adam by combining the idea of Nesterov momentum with the adaptive algorithms. The authors prove that in the nonconvex optimization setting, the proposed algorithm converges faster than the other adaptive algorithms. Experimental results show that the proposed algorithm performs be... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes another variant of Adam by combining the idea of Nesterov momentum with the adaptive algorithms. The authors prove that in the nonconvex optimization setting, the proposed algorithm converges faster than the other adaptive algorithms. Experimental results show that the proposed algorithm per... |
This paper proposes to combine human input in the form of a set of predicates and objects that they associate with a given problem, and realising the fact that for every deterministic MDPs there is a symbolic representation that they dub optimistic representation that reaches the goal state. They provide the theoretic... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes to combine human input in the form of a set of predicates and objects that they associate with a given problem, and realising the fact that for every deterministic MDPs there is a symbolic representation that they dub optimistic representation that reaches the goal state. They provide the t... |
1) The paper establishes a theorem to explain why polyhedra produced by deep networks are simple and uniform and the linear region dominates. 2) The paper proves that deep learning does not overfit. 3) The paper promotes the theoretical research of ReLU. 4) The paper is submitted with a basic implementation.
Strength:
... | Recommendation: 5: marginally below the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
1) The paper establishes a theorem to explain why polyhedra produced by deep networks are simple and uniform and the linear region dominates. 2) The paper proves that deep learning does not overfit. 3) The paper promotes the theoretical research of ReLU. 4) The paper is submitted with a basic implementation.
St... |
This work considers application scenarios that utilize large language models in a in-context learning paradigm with training data that cannot fit the context due to the input-length limitation.
To alleviate the challenge posed by the maximum length limit of language models, this work proposes the knn prompting method.
... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This work considers application scenarios that utilize large language models in a in-context learning paradigm with training data that cannot fit the context due to the input-length limitation.
To alleviate the challenge posed by the maximum length limit of language models, this work proposes the knn prompting ... |
The paper looks at he notion of _Populated Region Set (PRS)_ which are decision regions with at least one training example in them. The central message of the paper is that networks with lower number of PRS have better robustness.
## Strength
* The concept of PRS is interesting and well motivated.
* The design of the ... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The paper looks at he notion of _Populated Region Set (PRS)_ which are decision regions with at least one training example in them. The central message of the paper is that networks with lower number of PRS have better robustness.
## Strength
* The concept of PRS is interesting and well motivated.
* The design... |
This paper aims to evaluate the robustness of recognition systems to adversarially sampled training sets, and thus get an estimate of the worst case performance. The main contribution is an approach to identify the worst possible training set; this approach is based on maximizing the MMD between the training set and th... | Recommendation: 1: strong reject | Area: Deep Learning and representational learning | Review:
This paper aims to evaluate the robustness of recognition systems to adversarially sampled training sets, and thus get an estimate of the worst case performance. The main contribution is an approach to identify the worst possible training set; this approach is based on maximizing the MMD between the training se... |
This paper proposes a novel hierarchical Bayesian approach to federated learning, and derive ELBO objective function using variational inference techniques. Then the block-coordinate descent algorithm is devised, which fit well in the federated learning regime. Theoretical result on the convergence of the algorithm is... | Recommendation: 6: marginally above the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
This paper proposes a novel hierarchical Bayesian approach to federated learning, and derive ELBO objective function using variational inference techniques. Then the block-coordinate descent algorithm is devised, which fit well in the federated learning regime. Theoretical result on the convergence of the algo... |
This paper proposes to employ a differentiable neural ray tracer for wireless channel modeling. It models the time-angle channel impulse response as a superposition of multiple paths, and the wireless characteristics of each path are a result of multiple evaluations of an implicit neural network. The proposed framework... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes to employ a differentiable neural ray tracer for wireless channel modeling. It models the time-angle channel impulse response as a superposition of multiple paths, and the wireless characteristics of each path are a result of multiple evaluations of an implicit neural network. The proposed f... |
Focus on the consistent of ML interpretations, this paper introduces a new objective of consistency based on a notion called truthful interpretation by applying Fourier analysis of Boolean functions. Experimental results show that the method achieves higher consistency compared with other methods.
Strong Points:
[1]Th... | Recommendation: 5: marginally below the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
Focus on the consistent of ML interpretations, this paper introduces a new objective of consistency based on a notion called truthful interpretation by applying Fourier analysis of Boolean functions. Experimental results show that the method achieves higher consistency compared with other methods.
Strong Point... |
This paper presented a new module to improve self-supervised contrastive visual representation learning. Specifically, the proposed module focused on equivariance in the leaned latent space. Experimental analysis on two public datasets showed that when applying the proposed module to existing methods (SimCLR, BYOL and ... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper presented a new module to improve self-supervised contrastive visual representation learning. Specifically, the proposed module focused on equivariance in the leaned latent space. Experimental analysis on two public datasets showed that when applying the proposed module to existing methods (SimCLR, B... |
In this work, they propose a new paradigm based on the task vectors.
This task vector implies a direction of the pre-trained model. Task vectors are defined as a subtraction between a pre-trained model and fine-tuned model from a same pre-trained model with a specific task. This vectors can be used with arithmetic oper... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
In this work, they propose a new paradigm based on the task vectors.
This task vector implies a direction of the pre-trained model. Task vectors are defined as a subtraction between a pre-trained model and fine-tuned model from a same pre-trained model with a specific task. This vectors can be used with arithme... |
The paper proposes an architectural modification of the transformer module, suitable for online predictions. The modification is meant to prevent heavy re-computations when new tokens arrive.
Authors propose two flavors of the continual transformer: one in which previous outputs are updated (revised), and one focused ... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The paper proposes an architectural modification of the transformer module, suitable for online predictions. The modification is meant to prevent heavy re-computations when new tokens arrive.
Authors propose two flavors of the continual transformer: one in which previous outputs are updated (revised), and one ... |
The paper presents a framework to transform a neural network model into an interpretable CBM model that does not require a domain expert to provide concepts and does not require labeled concept data. The authors leveraged GPT3 model by prompting it in a certain way to get concepts related to each output class. They lat... | Recommendation: 6: marginally above the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper presents a framework to transform a neural network model into an interpretable CBM model that does not require a domain expert to provide concepts and does not require labeled concept data. The authors leveraged GPT3 model by prompting it in a certain way to get concepts related to each output class. ... |
This paper considers the problem of reweighting training samples to improve model performance on out-of-distribution test samples. The approach formulates real distribution shifts (covariate and concept related) using the exponential tilt assumption. With this assumption, the problem of improving performance on OOD sam... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper considers the problem of reweighting training samples to improve model performance on out-of-distribution test samples. The approach formulates real distribution shifts (covariate and concept related) using the exponential tilt assumption. With this assumption, the problem of improving performance on... |
The authors describe regularization based strategies for learning feature transformations that result in risk-invariant classifiers. They give two canonical generative DAGs and study domain generalization under these DAGs. The authors prove that any feature transformation must satisfy all relevant (conditional) indepen... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The authors describe regularization based strategies for learning feature transformations that result in risk-invariant classifiers. They give two canonical generative DAGs and study domain generalization under these DAGs. The authors prove that any feature transformation must satisfy all relevant (conditional)... |
This paper presents a method to automatically evaluate the contribution of a data point from some historical distribution to a new task without knowing the exact distributions of both old data and the new task. Experiental results show that the proposed wighting method works consitently better than several baseline met... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
This paper presents a method to automatically evaluate the contribution of a data point from some historical distribution to a new task without knowing the exact distributions of both old data and the new task. Experiental results show that the proposed wighting method works consitently better than several base... |
This paper unrolls the update formula of the LIF model along the time axis and uses the vectorized variables for parallel acceleration. It proposes a simple method to determine when the first spikes are fired. In this manner, both the goal of speeding up training and lowering the spike counts are accomplished.
Strength... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
This paper unrolls the update formula of the LIF model along the time axis and uses the vectorized variables for parallel acceleration. It proposes a simple method to determine when the first spikes are fired. In this manner, both the goal of speeding up training and lowering the spike counts are accomplished.
... |
This paper introduces AutoMoE, an AutoML framework for searching efficient sparsely activated models. Experiments are conducted on three machine translation benchmark datasets, including WMT'14 En-De, WMT'14 En-Fr, and WMT'19 En-De. Compared with previous models (e.g., HAT), AutoMoE achieves a better trade-off between ... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper introduces AutoMoE, an AutoML framework for searching efficient sparsely activated models. Experiments are conducted on three machine translation benchmark datasets, including WMT'14 En-De, WMT'14 En-Fr, and WMT'19 En-De. Compared with previous models (e.g., HAT), AutoMoE achieves a better trade-off ... |
The authors consider approximate inference in Bayesian networks, specifically in the subset of Bayesian networks that can be represented as plate diagrams, which are common for large population studies, the population is composed of subgroups and each group is composed of many individuals. Given global the population p... | Recommendation: 5: marginally below the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The authors consider approximate inference in Bayesian networks, specifically in the subset of Bayesian networks that can be represented as plate diagrams, which are common for large population studies, the population is composed of subgroups and each group is composed of many individuals. Given global the popu... |
The authors proposed a novel backdoor defense, which used first-order gradient to identify bad neurons using clean dataset. Also, a new Adaptive Regularization (AR) mechanism is used to fine-tune the model using detected bad neurons. They performed experiments on ten different backdoor attacks with three benchmark data... | Recommendation: 5: marginally below the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The authors proposed a novel backdoor defense, which used first-order gradient to identify bad neurons using clean dataset. Also, a new Adaptive Regularization (AR) mechanism is used to fine-tune the model using detected bad neurons. They performed experiments on ten different backdoor attacks with three benchm... |
The authors propose to model robotics taxonomy data via a proposed Gaussian process hyperbolic latent variable model (GPHLVM). The GPHLVM maintains the structure of the human-designed taxonomy while embedding the taxonomy node features into a hyperbolic latent space, which naturally encodes a continuous hierarchical st... | Recommendation: 8: accept, good paper | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The authors propose to model robotics taxonomy data via a proposed Gaussian process hyperbolic latent variable model (GPHLVM). The GPHLVM maintains the structure of the human-designed taxonomy while embedding the taxonomy node features into a hyperbolic latent space, which naturally encodes a continuous hierarc... |
The authors study convergence and generalization of implicit models (in particular simple deep equilibrium models).
The submission aims at addressing the following gaps in the literature:
* **Existing convergence results rely on studying the read-out layer:** The authors address this by deriving a convergence result... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The authors study convergence and generalization of implicit models (in particular simple deep equilibrium models).
The submission aims at addressing the following gaps in the literature:
* **Existing convergence results rely on studying the read-out layer:** The authors address this by deriving a convergenc... |
This paper focuses on anomaly detection models and proposes an adversarially robust anomaly detector based on the diffusion model. In detail, they first add gaussian to the image through the forward process, and then reconstruct it through the reverse process; the Multiscale Reconstruction Error Map is computed as the... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper focuses on anomaly detection models and proposes an adversarially robust anomaly detector based on the diffusion model. In detail, they first add gaussian to the image through the forward process, and then reconstruct it through the reverse process; the Multiscale Reconstruction Error Map is compute... |
The paper proposes a novel method for open-set 3D detection using image-level class supervision. The core idea is leveraging each of the image and point-cloud modalities to generate pseudo labels for unseen classes. To improve the positive and negative sample matching, the authors propose debiased cross-modal contrasti... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes a novel method for open-set 3D detection using image-level class supervision. The core idea is leveraging each of the image and point-cloud modalities to generate pseudo labels for unseen classes. To improve the positive and negative sample matching, the authors propose debiased cross-modal c... |
This paper investigates a problem I would not have thought about -- CNNs overfitting to input size due to the pooling arithmetic involved -- this in itself is an extremely exotic thought! The authors show that due to the unconsumed padding, often the networks trained on specific input sizes will not generalize or extra... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper investigates a problem I would not have thought about -- CNNs overfitting to input size due to the pooling arithmetic involved -- this in itself is an extremely exotic thought! The authors show that due to the unconsumed padding, often the networks trained on specific input sizes will not generalize ... |
PLOT [Pacchiano 2021] leverages a adversarial-training-like scheme to train NN for online binary classification problem: at each time step, retrain the model M while pretending that the new data are of the positive class, and give prediction on the new data using the retrained model M.
This paper proposes to add a sec... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
PLOT [Pacchiano 2021] leverages a adversarial-training-like scheme to train NN for online binary classification problem: at each time step, retrain the model M while pretending that the new data are of the positive class, and give prediction on the new data using the retrained model M.
This paper proposes to a... |
This paper introduces a first-of-its-kind suite of benchmarking tasks for antibody-specific language models and provides some interesting observations about the behavior of general protein models and antibody-specific models on these tasks. It also introduces a new antibody-specific pretraining objective based on the u... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper introduces a first-of-its-kind suite of benchmarking tasks for antibody-specific language models and provides some interesting observations about the behavior of general protein models and antibody-specific models on these tasks. It also introduces a new antibody-specific pretraining objective based ... |
The submission studies attention patterns in vision transformer models, finding significant sparsity across layers. Manipulations that enforce greater sparsity do not severely reduce the performance of these models. A comparison between attentional sparsity in transformers and activation sparsity in convolutional neura... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
The submission studies attention patterns in vision transformer models, finding significant sparsity across layers. Manipulations that enforce greater sparsity do not severely reduce the performance of these models. A comparison between attentional sparsity in transformers and activation sparsity in convolution... |
The paper attempts to propose a solution to clean moire which includes patterns of colored stripes with varied frequencies captured by a camera degrading the visual quality of images. Instead of using a large CNN to clean a whole input image, it targets image patches with lighter CNNs. In order to determine the width o... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper attempts to propose a solution to clean moire which includes patterns of colored stripes with varied frequencies captured by a camera degrading the visual quality of images. Instead of using a large CNN to clean a whole input image, it targets image patches with lighter CNNs. In order to determine the... |
This paper improves upon previous oblivious sketching and turnstile streaming results for $\ell_1$ and logistic regression, achieving constant factor approximation, and an efficient algorithm in the sketched space. They demonstrate:
1. Based on a modification of Munteanu et al. (2021) (with significant change of the a... | Recommendation: 8: accept, good paper | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper improves upon previous oblivious sketching and turnstile streaming results for $\ell_1$ and logistic regression, achieving constant factor approximation, and an efficient algorithm in the sketched space. They demonstrate:
1. Based on a modification of Munteanu et al. (2021) (with significant change ... |
The paper proposes a general design improvement over models from the DETR family. Specifically, it introduces an additional set of *modulated queries* which are convex combinations of the regular queries. The weights used for the combination are predicted by a small MLP operating on the pooled representation of the ima... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper proposes a general design improvement over models from the DETR family. Specifically, it introduces an additional set of *modulated queries* which are convex combinations of the regular queries. The weights used for the combination are predicted by a small MLP operating on the pooled representation of... |
Providing informative rewards is crucial for effective reinforcement learning. Prior IRL methods overfit to demonstrations and fail to learn generalizable rewards. To combat this issue, this paper proposes BC-IRL, which uses gradient-based bi-level optimization to learn the reward. The authors evaluate on two continuou... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
Providing informative rewards is crucial for effective reinforcement learning. Prior IRL methods overfit to demonstrations and fail to learn generalizable rewards. To combat this issue, this paper proposes BC-IRL, which uses gradient-based bi-level optimization to learn the reward. The authors evaluate on two c... |
The paper proposes PatchDCT for high-quality instance segmentation. Different from DCT-Mask, the whole image mask is divided into different patches. Each patch is refined individually by the classifier and regressor. The refinement is performed in a multi-stage. Improvements on mask quality are observed on COCO, Citysc... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper proposes PatchDCT for high-quality instance segmentation. Different from DCT-Mask, the whole image mask is divided into different patches. Each patch is refined individually by the classifier and regressor. The refinement is performed in a multi-stage. Improvements on mask quality are observed on COCO... |
This paper proposes a simple variation of batch normalization with a bounded scaling parameter to estimate the importance of each channel. To only suppress unimportant units while preserving important units, the authors design a regularization loss on these bounded scaling factors, with additional hyperparameters. The ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a simple variation of batch normalization with a bounded scaling parameter to estimate the importance of each channel. To only suppress unimportant units while preserving important units, the authors design a regularization loss on these bounded scaling factors, with additional hyperparamete... |
The manuscript describes a Transformer-based neural network aimed at predicting the energies and the forces of molecules. The network is trained on the prediction of the energy of ~130,000 small molecules. The loss includes the square deviation between the ground truth energy and the predicted energy, and also terms ai... | Recommendation: 3: reject, not good enough | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The manuscript describes a Transformer-based neural network aimed at predicting the energies and the forces of molecules. The network is trained on the prediction of the energy of ~130,000 small molecules. The loss includes the square deviation between the ground truth energy and the predicted energy, and also ... |
This paper extends on the power NeRF, by introducing an additional vector to present the object code for each 3D point. The authors designed a few loss functions and showed that with these loss functions the object code for the 3D points can be directly optimized from 2D image annotations. The authors then used the lea... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper extends on the power NeRF, by introducing an additional vector to present the object code for each 3D point. The authors designed a few loss functions and showed that with these loss functions the object code for the 3D points can be directly optimized from 2D image annotations. The authors then used... |
The author proposed a few-shot retrieval evaluation setting, which addresses the difference in the search intent and query distribution for different retrieval tasks. They also proposed a simple recipe for few-shot retrieval by prompting an LLM to generate synthetic task-specific training data and then train a dual enc... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The author proposed a few-shot retrieval evaluation setting, which addresses the difference in the search intent and query distribution for different retrieval tasks. They also proposed a simple recipe for few-shot retrieval by prompting an LLM to generate synthetic task-specific training data and then train a ... |
The authors of the given manuscript extend a bio-plausible backpropagation alternative, “SoftHebb,” to apply to training multiple layers. In biological plausibility, this algorithm bows to the constraints of no weight transport, local plasticity, and no time-locked updates. In addition, the learning is unsupervised. In... | Recommendation: 6: marginally above the acceptance threshold | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
The authors of the given manuscript extend a bio-plausible backpropagation alternative, “SoftHebb,” to apply to training multiple layers. In biological plausibility, this algorithm bows to the constraints of no weight transport, local plasticity, and no time-locked updates. In addition, the learning is unsuperv... |
This paper proposes a new method DifFace, being able to cope with unseen and complex degradations more gracefully without complicated loss designs. Comprehensive experiments show that DifFace is better than current state-of-the-art methods, especially in cases with severe degradations.
Pros:
1. The motivation is clear... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a new method DifFace, being able to cope with unseen and complex degradations more gracefully without complicated loss designs. Comprehensive experiments show that DifFace is better than current state-of-the-art methods, especially in cases with severe degradations.
Pros:
1. The motivation ... |
This work highlights an intriguing problem regarding the tuning of soft prompts in vision and language models. This research asserts that current soft-prompt tuning papers have a tendency to overfit on training data while losing generalization ability on test data. To avoid this, they strive to optimize soft prompts in... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This work highlights an intriguing problem regarding the tuning of soft prompts in vision and language models. This research asserts that current soft-prompt tuning papers have a tendency to overfit on training data while losing generalization ability on test data. To avoid this, they strive to optimize soft pr... |
The paper presents a training framework (from scratch) for one- time-step SNNs that uses a novel variant of the recently proposed Hoyer regularizer. By estimating the threshold of each SNN layer as the Hoyer extremum of a clipped version of its activation map, the approach not only downscales the value of the trainable... | Recommendation: 5: marginally below the acceptance threshold | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
The paper presents a training framework (from scratch) for one- time-step SNNs that uses a novel variant of the recently proposed Hoyer regularizer. By estimating the threshold of each SNN layer as the Hoyer extremum of a clipped version of its activation map, the approach not only downscales the value of the t... |
The present work explores the relationship between input noise and network sparseness. Such sparseness is known to exist in biological networks (brains), and has been hypothesized to arise as a solution to increase the signal-to-noise ratio of the network and to keep energy consumption low. Here, the authors find that... | Recommendation: 3: reject, not good enough | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
The present work explores the relationship between input noise and network sparseness. Such sparseness is known to exist in biological networks (brains), and has been hypothesized to arise as a solution to increase the signal-to-noise ratio of the network and to keep energy consumption low. Here, the authors f... |
The main contribution of this paper is Cy2C-GNN, a model that can distinguish pairs of isomorphic graphs that 1-, 2- and 3-WL tests cannot. The proposed approach identifies the cycle basis of the graph and then constructs complete subgraphs consisting of the nodes of each basis element. The clique adjacency matrix is c... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The main contribution of this paper is Cy2C-GNN, a model that can distinguish pairs of isomorphic graphs that 1-, 2- and 3-WL tests cannot. The proposed approach identifies the cycle basis of the graph and then constructs complete subgraphs consisting of the nodes of each basis element. The clique adjacency mat... |
This paper proposes a strategy to augment flows with annealed importance sampling (AIS) with the objective to minimize importance weight variance. AIS is responsible to generate samples in regions where the flow struggles. The proposed FAB was shown to produce
better results than training via maximum likelihood on MD s... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper proposes a strategy to augment flows with annealed importance sampling (AIS) with the objective to minimize importance weight variance. AIS is responsible to generate samples in regions where the flow struggles. The proposed FAB was shown to produce
better results than training via maximum likelihood... |
This paper proposes the energy transformer, a transformer architecture that uses a recurrent energy transformer block. The energy transformer block updates its input in accordance to minimizing two energy functions. Experiments on image reconstruction and graph anomaly detection demonstrate the effectiveness of the pro... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes the energy transformer, a transformer architecture that uses a recurrent energy transformer block. The energy transformer block updates its input in accordance to minimizing two energy functions. Experiments on image reconstruction and graph anomaly detection demonstrate the effectiveness of... |
This paper focuses on the adversarial training of robust graph neural networks. The motivation is attractive to me, namely, the divergence of distribution between clean graphs and attacked graphs. This motivation leads to the GAME model, a set of effective all-round designs: leveraging MoE module to construct a more po... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper focuses on the adversarial training of robust graph neural networks. The motivation is attractive to me, namely, the divergence of distribution between clean graphs and attacked graphs. This motivation leads to the GAME model, a set of effective all-round designs: leveraging MoE module to construct a... |
With the aim of type inference for untyped code, the paper feeds, in addition to tokenized source code, contextual information gained from static analysis to the seq2seq-based code completion model called CodeT5. CodeT5 itself can infer types of elements in untyped code (including user-defined and parametric types) suc... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
With the aim of type inference for untyped code, the paper feeds, in addition to tokenized source code, contextual information gained from static analysis to the seq2seq-based code completion model called CodeT5. CodeT5 itself can infer types of elements in untyped code (including user-defined and parametric ty... |
This paper explores a unconditional learning generation model using Wasserstein Gradient Flow directly on the data space. To make the training feasible, the authors propose to use a deep density ratio estimator for learning in Wasserstein space.Authors demonstrate the proposed method on Cifar10 and CelebA, and the resu... | Recommendation: 5: marginally below the acceptance threshold | Area: Generative models | Review:
This paper explores a unconditional learning generation model using Wasserstein Gradient Flow directly on the data space. To make the training feasible, the authors propose to use a deep density ratio estimator for learning in Wasserstein space.Authors demonstrate the proposed method on Cifar10 and CelebA, and ... |
The authors propose an approach for the visual room rearrangement challenge that creates a voxelized map of the scene during the walkthrough and unshuffle phases, finds the differences between the maps, and trains an agent to rearrange the objects based on the differences. The approach improves substantially over the p... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The authors propose an approach for the visual room rearrangement challenge that creates a voxelized map of the scene during the walkthrough and unshuffle phases, finds the differences between the maps, and trains an agent to rearrange the objects based on the differences. The approach improves substantially ov... |
The paper introduces a framework for measuring and optimizing for quantile-based risk functions, which go beyond the standard practice of measuring and optimizing for the average (expected) loss function. The framework allows one to optimize for any quantity expressed as an integral over the quantile function of the lo... | Recommendation: 6: marginally above the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper introduces a framework for measuring and optimizing for quantile-based risk functions, which go beyond the standard practice of measuring and optimizing for the average (expected) loss function. The framework allows one to optimize for any quantity expressed as an integral over the quantile function o... |
Adversarial transferability is often studied in the image-model-to-image-model setting. This paper studies how to transfer adversarial examples from image models to video (or multi-view) models. They note that simply transferring adversarial examples from image models to video models are suboptimal, and propose a promp... | Recommendation: 6: marginally above the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
Adversarial transferability is often studied in the image-model-to-image-model setting. This paper studies how to transfer adversarial examples from image models to video (or multi-view) models. They note that simply transferring adversarial examples from image models to video models are suboptimal, and propose... |
This paper implements existing DP optimizers with 1.24x training speed than the most efficient competitor. Previous methods accelerate the DP training speed by using two rounds of back-propagation or sacrificing memory. This paper uses one back-propagation and never instantiates per-sample gradients so that it reduces ... | Recommendation: 5: marginally below the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper implements existing DP optimizers with 1.24x training speed than the most efficient competitor. Previous methods accelerate the DP training speed by using two rounds of back-propagation or sacrificing memory. This paper uses one back-propagation and never instantiates per-sample gradients so that it ... |
This paper introduces an approach for rendering novel views from a single image. The approach is based on a pose-conditional image-to-image diffusion model. The paper shows the proposed approach can be generalized to the shapes that are not seen during the test time. The results of the proposed approach are superior th... | Recommendation: 5: marginally below the acceptance threshold | Area: Generative models | Review:
This paper introduces an approach for rendering novel views from a single image. The approach is based on a pose-conditional image-to-image diffusion model. The paper shows the proposed approach can be generalized to the shapes that are not seen during the test time. The results of the proposed approach are sup... |
This paper focuses on the task of predicting the effects of mutations in protein sequences. To that end it leverages a graph neural network with : 1) Biochemical (eg., AA type, SAS, B-factor) and geometric (eg., 3D coordinates of alpha-carbons, dihedral angles) features of amino acids (nodes in the graphs) and edge fea... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper focuses on the task of predicting the effects of mutations in protein sequences. To that end it leverages a graph neural network with : 1) Biochemical (eg., AA type, SAS, B-factor) and geometric (eg., 3D coordinates of alpha-carbons, dihedral angles) features of amino acids (nodes in the graphs) and ... |
This paper studies the generalization property of federated learning under the two-level distribution framework: client D_i is sampled from a meta-distribution P, and sample Z_i^j is sampled from D_i. Excess risk for unseen clients and semi-excess risk for seen data are analyzed under various conditions in high probabi... | Recommendation: 8: accept, good paper | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper studies the generalization property of federated learning under the two-level distribution framework: client D_i is sampled from a meta-distribution P, and sample Z_i^j is sampled from D_i. Excess risk for unseen clients and semi-excess risk for seen data are analyzed under various conditions in high... |
This paper combines two tasks - masked image modeling and masked language modeling - in a single framework, and trains a vision-language encoder to solve both tasks. The resulting model shows some capability for image-to-text generation, as well as reasonable representation learning performance.
Strengths:
- This paper... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper combines two tasks - masked image modeling and masked language modeling - in a single framework, and trains a vision-language encoder to solve both tasks. The resulting model shows some capability for image-to-text generation, as well as reasonable representation learning performance.
Strengths:
- Th... |
This paper studies VCREG, a particular regularization function used for contrastive representation learning. The purpose of this term is to prevent embedding collapse, where all inputs map to the same embedding (a trivial way to satisfy that embeddings corresponding to different views of the same object should map to t... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper studies VCREG, a particular regularization function used for contrastive representation learning. The purpose of this term is to prevent embedding collapse, where all inputs map to the same embedding (a trivial way to satisfy that embeddings corresponding to different views of the same object should ... |
The authors propose a theoretical analysis for the Stochastic Gradient Descent optimizer for normalized deep neural networks including weight decay. In particular, they study the behavior of the loss during training, taking into account that for such models the optimization converges to an equilibrium related to the we... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
The authors propose a theoretical analysis for the Stochastic Gradient Descent optimizer for normalized deep neural networks including weight decay. In particular, they study the behavior of the loss during training, taking into account that for such models the optimization converges to an equilibrium related t... |
The paper describes a new method for AutoML (FALCON) that comprises a online search in the space of designs that is guided by a GNN model ("meta-GNN") that is trained along the way. The GNN comprises a task-agnostic component that tries to capture similarity between designs using various features and relations, as well... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper describes a new method for AutoML (FALCON) that comprises a online search in the space of designs that is guided by a GNN model ("meta-GNN") that is trained along the way. The GNN comprises a task-agnostic component that tries to capture similarity between designs using various features and relations,... |
This work proposed an NFlow-based noise-robust UDA method. Specifically, they trained an NFlow (Durkan et al. 2019) generative model of the feature distribution for each class by using the noisy pseudo labels generated from the source. With the classwise NFlow models, they introduced a modified feature mix-up scheme (... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This work proposed an NFlow-based noise-robust UDA method. Specifically, they trained an NFlow (Durkan et al. 2019) generative model of the feature distribution for each class by using the noisy pseudo labels generated from the source. With the classwise NFlow models, they introduced a modified feature mix-up ... |
In many situations an RL agent incurs costs whenever it acts. Simply applying standard RL algorithms in such situations often doesn’t work, because it is not easy for the algorithm to choose “no action” instead of “small action” (particularly in continuous action spaces). To deal with this, the authors propose Learnabl... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
In many situations an RL agent incurs costs whenever it acts. Simply applying standard RL algorithms in such situations often doesn’t work, because it is not easy for the algorithm to choose “no action” instead of “small action” (particularly in continuous action spaces). To deal with this, the authors propose ... |
This paper formulates several self supervised learning methods (VICReg, SimCLR, MSN, SwAV) with the volume maximization principle as variants of k-means, analyzes their difference on class-balanced and class-imbalanced data, and indicates that the uniform prior w.r.t. prototype vectors is the curse of performance degen... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper formulates several self supervised learning methods (VICReg, SimCLR, MSN, SwAV) with the volume maximization principle as variants of k-means, analyzes their difference on class-balanced and class-imbalanced data, and indicates that the uniform prior w.r.t. prototype vectors is the curse of performan... |
In this paper, the authors proposed a ProtoKNN model, which uses k-nearest neighbor (kNN) on prototype similarities to classify an input image. In particular, the proposed model uses a ProtoPNet backbone (with cosine similarity), and classifies an input image by comparing its prototype similarities with those of the tr... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
In this paper, the authors proposed a ProtoKNN model, which uses k-nearest neighbor (kNN) on prototype similarities to classify an input image. In particular, the proposed model uses a ProtoPNet backbone (with cosine similarity), and classifies an input image by comparing its prototype similarities with those o... |
This paper presents a simple but effective approach to zero-shot learning and explainability for image classification. The title of the paper says it all, but for completeness here is a summary. The main idea is to query a large language model, in particular GPT-3, to obtain multiple short descriptions about a class. T... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper presents a simple but effective approach to zero-shot learning and explainability for image classification. The title of the paper says it all, but for completeness here is a summary. The main idea is to query a large language model, in particular GPT-3, to obtain multiple short descriptions about a ... |
The paper proposes an approach to identify the null space of vision transformers (ViT). They divide this into two parts – (a) null space of patch embeddings, and, (b) null space for the self-attention module(s). The paper uses the above to study the robustness of ViT to null space noise, compare the impact on different... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
The paper proposes an approach to identify the null space of vision transformers (ViT). They divide this into two parts – (a) null space of patch embeddings, and, (b) null space for the self-attention module(s). The paper uses the above to study the robustness of ViT to null space noise, compare the impact on d... |
This paper is about quality of confidence predictions made by a bunch of pretrained models on ImageNet 1K, in terms of selective prediction (measured by the accuracy vs confidence plot), ranking (measured by AUROC), and calibration (using ECE).
The authors claim to evaluate 523 pretrained models from PyTorch and timm ... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper is about quality of confidence predictions made by a bunch of pretrained models on ImageNet 1K, in terms of selective prediction (measured by the accuracy vs confidence plot), ranking (measured by AUROC), and calibration (using ECE).
The authors claim to evaluate 523 pretrained models from PyTorch a... |
This paper proposes a factorization method for clinical questionnaires by by idea of matrix completion. Given an original data matrix which stacks answers of a questionnaire from all respondents. The method searches by optimizing a regularized loss function under interpretability constraints for a pair of factor and lo... | Recommendation: 8: accept, good paper | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper proposes a factorization method for clinical questionnaires by by idea of matrix completion. Given an original data matrix which stacks answers of a questionnaire from all respondents. The method searches by optimizing a regularized loss function under interpretability constraints for a pair of facto... |
This work proposes the Dataset Lottery Ticket Hypothesis(DLTH), a novel problem that studies the possibility of identifying the subset which can reflect the performance consistency with the full data. By Empirical Risk Trend, this work demonstrates the existence of dataset-winning tickets. And extensive experiments are... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This work proposes the Dataset Lottery Ticket Hypothesis(DLTH), a novel problem that studies the possibility of identifying the subset which can reflect the performance consistency with the full data. By Empirical Risk Trend, this work demonstrates the existence of dataset-winning tickets. And extensive experim... |
This paper studies multi-agent problems with a very large number of agents. In the past, such problems have been tackled using a combination of reinforcement learning and mean field approximations. This approach is valid provided the interactions between the players are homogeneous in the sense that every player intera... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper studies multi-agent problems with a very large number of agents. In the past, such problems have been tackled using a combination of reinforcement learning and mean field approximations. This approach is valid provided the interactions between the players are homogeneous in the sense that every playe... |
Traditional representation learning methods (for example, contrastive learning), while making significant progress, have potential fairness issues. This paper proposes training the image attribute editor to generate contrastive sample pairs for each sample in the original dataset, which share the same visual informatio... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
Traditional representation learning methods (for example, contrastive learning), while making significant progress, have potential fairness issues. This paper proposes training the image attribute editor to generate contrastive sample pairs for each sample in the original dataset, which share the same visual in... |
This work builds on the prior work of curriculum learning to make the curriculum learning process neutral to the training schedule. They first form initial pre-representation-training of the network with the help of NCC/ridge regression. The network is divided into 2 components: ϕ (classifier) and ψ (backbone). The ψ f... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This work builds on the prior work of curriculum learning to make the curriculum learning process neutral to the training schedule. They first form initial pre-representation-training of the network with the help of NCC/ridge regression. The network is divided into 2 components: ϕ (classifier) and ψ (backbone).... |
This paper contributes to the study of gradient-based inverse reinforcement learning, in particular, learning a reward function by gradient ascent, under which an observed policy is optimal. The main problem of interest is the estimation of the gradient of the policy return with respect to the parameters of the reward ... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper contributes to the study of gradient-based inverse reinforcement learning, in particular, learning a reward function by gradient ascent, under which an observed policy is optimal. The main problem of interest is the estimation of the gradient of the policy return with respect to the parameters of the... |
This work proposes a spike calibration algorithm for reducing the conversion errors in ANN-SNN conversion models. The authors provide a new measurement for the conversion error by the offset spike and develop an optimization strategy that can shift initial membrane potential to offset the conversion errors iteratively.... | Recommendation: 8: accept, good paper | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
This work proposes a spike calibration algorithm for reducing the conversion errors in ANN-SNN conversion models. The authors provide a new measurement for the conversion error by the offset spike and develop an optimization strategy that can shift initial membrane potential to offset the conversion errors iter... |
Deep learning-based forecasting techniques for multi-variate time series forecasting have largely replaced more traditional techniques in recent years. However, there is an overall lack of consensus on the effectiveness of such methods overall, and the value of using more complex transformer-based methods has become in... | Recommendation: 3: reject, not good enough | Area: Unsupervised and Self-supervised learning | Review:
Deep learning-based forecasting techniques for multi-variate time series forecasting have largely replaced more traditional techniques in recent years. However, there is an overall lack of consensus on the effectiveness of such methods overall, and the value of using more complex transformer-based methods has b... |
The paper presents a framework for training RL agents to collaborate with human players in 2-player Hanabi games. The key challenge of training such a policy is that human data is finite and fixed, which means that there needs to be a human-like behavior policy for the RL to learn to collaborate with. Prior work shows ... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper presents a framework for training RL agents to collaborate with human players in 2-player Hanabi games. The key challenge of training such a policy is that human data is finite and fixed, which means that there needs to be a human-like behavior policy for the RL to learn to collaborate with. Prior wor... |
The paper deals with building robust graph learning methods not susceptible to adversarial attacks. Existing methods learn the structure of graph either by preprocessing the graph structure or by parametrically learning the graph adjacency. In this paper, the authors propose to simultaneously learn the graph structure ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper deals with building robust graph learning methods not susceptible to adversarial attacks. Existing methods learn the structure of graph either by preprocessing the graph structure or by parametrically learning the graph adjacency. In this paper, the authors propose to simultaneously learn the graph st... |
This paper proposes a DySR method that maintains QoS while maximizing the model performance. The proposed method is mainly based on NAS. Experimental results show the effect of the DySR.
Strength
This paper aims to develops an image SR method on the QoS. The main goal is to propose an efficient method that can work... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a DySR method that maintains QoS while maximizing the model performance. The proposed method is mainly based on NAS. Experimental results show the effect of the DySR.
Strength
This paper aims to develops an image SR method on the QoS. The main goal is to propose an efficient method that ... |
The authors study regularized extensive-form games (EFGs). They show that various algorithms achieve last-iterate convergence guarantees on such games. Then, the authors use the fact that for small amounts of regularization the solution to the regularized problem is also an approximate solution to the original problem.... | Recommendation: 8: accept, good paper | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The authors study regularized extensive-form games (EFGs). They show that various algorithms achieve last-iterate convergence guarantees on such games. Then, the authors use the fact that for small amounts of regularization the solution to the regularized problem is also an approximate solution to the original ... |
The paper proposes three algorithms for finding optimal policies in environment with high-dimensional continuous controls based on decomposing the policy into independent, but cooperating policies for each action dimension, discretizing uniformly each action, and finding its optimal policy efficiently. Empirical evalua... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper proposes three algorithms for finding optimal policies in environment with high-dimensional continuous controls based on decomposing the policy into independent, but cooperating policies for each action dimension, discretizing uniformly each action, and finding its optimal policy efficiently. Empirica... |
The authors proposes a distribution on SO(3) to estimate 3 degrees of freedom rotations in RGB images. The distribution is inspired from the multi-dimensional Laplace distribution which will allow to have estimates that are robust to outliers. The authors also provide an equivalence between the matrix distribution on S... | Recommendation: 6: marginally above the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The authors proposes a distribution on SO(3) to estimate 3 degrees of freedom rotations in RGB images. The distribution is inspired from the multi-dimensional Laplace distribution which will allow to have estimates that are robust to outliers. The authors also provide an equivalence between the matrix distribut... |
This paper proposes an approach for overcoming some of the deficiencies of random-walk graph exploration in learning a low-dimensional shortest path (SP) representation of graphs. After identifying the drawbacks of generating SP representations via random walks, they propose an alternative representation based on betw... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes an approach for overcoming some of the deficiencies of random-walk graph exploration in learning a low-dimensional shortest path (SP) representation of graphs. After identifying the drawbacks of generating SP representations via random walks, they propose an alternative representation based... |
In this paper, the authors propose a method for domain adaptation when you have labelled data from source P and unlabelled data from target Q. Their method learns a reweighting of the source data sample through a learned exponential tilting, where a neural network learns the sufficient statistics of the tilting. They s... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
In this paper, the authors propose a method for domain adaptation when you have labelled data from source P and unlabelled data from target Q. Their method learns a reweighting of the source data sample through a learned exponential tilting, where a neural network learns the sufficient statistics of the tilting... |
This paper aimed to discover high-entropy alloys with high yield strength, by first predicting the yield strength. This work presented a dataset called X-Yield for this task and proposed a bi-level optimization method called Bi-RPT as their corss-quality few shot transfer workflow.
This work is very well motivated. Dis... | Recommendation: 3: reject, not good enough | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper aimed to discover high-entropy alloys with high yield strength, by first predicting the yield strength. This work presented a dataset called X-Yield for this task and proposed a bi-level optimization method called Bi-RPT as their corss-quality few shot transfer workflow.
This work is very well motiva... |
This paper proposes a post-processing method for neural image compression using a diffusion probabilistic model to improve the perceptual quality of the reconstruction results . With a variable-rate model that has been fully trained, no more bits need to be coded. The fidelity and perceptual quality are determined by a... | Recommendation: 3: reject, not good enough | Area: Generative models | Review:
This paper proposes a post-processing method for neural image compression using a diffusion probabilistic model to improve the perceptual quality of the reconstruction results . With a variable-rate model that has been fully trained, no more bits need to be coded. The fidelity and perceptual quality are determi... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.