review stringlengths 5 16.6k | score stringclasses 5
values | area stringclasses 12
values | text stringlengths 31 5.65k |
|---|---|---|---|
The paper focuses on the poor annotation issue and proposes a distantly supervised iterative training method for knowledge graph construction. The authors leverage distant supervision to automatically label entities and triples from a find-grained domain based on a coarse domain. In specific, they first use distant sup... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper focuses on the poor annotation issue and proposes a distantly supervised iterative training method for knowledge graph construction. The authors leverage distant supervision to automatically label entities and triples from a find-grained domain based on a coarse domain. In specific, they first use dis... |
The authors consider approximation of probability density functions by deep belief networks with binary hidden units. Convergence and rates of approximation are provided with respect to the L_q norm and K-L divergence.
The topic of approximation by deep belief networks is interesting. Rates of convergence are given by... | Recommendation: 3: reject, not good enough | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The authors consider approximation of probability density functions by deep belief networks with binary hidden units. Convergence and rates of approximation are provided with respect to the L_q norm and K-L divergence.
The topic of approximation by deep belief networks is interesting. Rates of convergence are ... |
This paper proposes a pipeline that detects epistemologically biased words (e.g., ripped - negative) and stereotypes (e.g., women should be dressed like brides) in texts. The authors develop an interactive user-interface to show the detected output with a goal of helping journalists and editors. The backend of this int... | Recommendation: 3: reject, not good enough | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper proposes a pipeline that detects epistemologically biased words (e.g., ripped - negative) and stereotypes (e.g., women should be dressed like brides) in texts. The authors develop an interactive user-interface to show the detected output with a goal of helping journalists and editors. The backend of ... |
This paper focuses on the core issue of simultaneous machine translation: time to start translation. The paper proposes HMT inspired by HMM, where the next token generation is regarded as the hidden variables, and target tokens as the observable variables. At each time step, K indicates available length of source for t... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper focuses on the core issue of simultaneous machine translation: time to start translation. The paper proposes HMT inspired by HMM, where the next token generation is regarded as the hidden variables, and target tokens as the observable variables. At each time step, K indicates available length of sour... |
This work solves the problem of graph unlearning, where a sequence of requests arrive to delete graph elements (nodes, edges) from trained graph neural networks (GNN). To unlearn information from a trained GNN, its influence on model weights must be deleted from the model. This work formalizes required properties for g... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This work solves the problem of graph unlearning, where a sequence of requests arrive to delete graph elements (nodes, edges) from trained graph neural networks (GNN). To unlearn information from a trained GNN, its influence on model weights must be deleted from the model. This work formalizes required properti... |
This paper improves Group VIT, a text-supervised segmentation model, by introducing region-wise cross-view consistent regularization. The author benchmarked for zero-shot segmentation capability across multiple datasets and showed improvements over several baselines.
My primary concern with this work is selecting the... | Recommendation: 3: reject, not good enough | Area: Unsupervised and Self-supervised learning | Review:
This paper improves Group VIT, a text-supervised segmentation model, by introducing region-wise cross-view consistent regularization. The author benchmarked for zero-shot segmentation capability across multiple datasets and showed improvements over several baselines.
My primary concern with this work is selec... |
This paper addresses the issue of pre-training loss cannot fully explain downstream performance, they instead claim that the flatness of the model is well-correlated with downstream performance. The author showed that at the same level of pretraining loss, large models have better downstream performance, because of the... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper addresses the issue of pre-training loss cannot fully explain downstream performance, they instead claim that the flatness of the model is well-correlated with downstream performance. The author showed that at the same level of pretraining loss, large models have better downstream performance, becaus... |
This work analyses model-based approaches to offline RL. Instead of considering model ensemble, it suggests using entropy regularization as a way to explore beyond the data support.
The paper is clearly written and easy to follow. However, although I am not familiar with the offline RL literature, I found the propose... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This work analyses model-based approaches to offline RL. Instead of considering model ensemble, it suggests using entropy regularization as a way to explore beyond the data support.
The paper is clearly written and easy to follow. However, although I am not familiar with the offline RL literature, I found the... |
The authors present a spike encoding scheme that they argue reduces the number of spikes from an event based spike stream. They argue this enables the training of larger SNNs with fewer timesteps.
They use this representation to train an SNN and test their trained model on DVS datasets such as N-MNIST and DVS-CIFAR10 ... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The authors present a spike encoding scheme that they argue reduces the number of spikes from an event based spike stream. They argue this enables the training of larger SNNs with fewer timesteps.
They use this representation to train an SNN and test their trained model on DVS datasets such as N-MNIST and DVS-... |
This paper proposes a general framework being able to enhance a wide variety of existing click-through rate (CTR) prediction models. In order to balance memorization and generalization, an instance-wise gating network is utilized to dynamically select the feature embedding which is fused with the deep representation o... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a general framework being able to enhance a wide variety of existing click-through rate (CTR) prediction models. In order to balance memorization and generalization, an instance-wise gating network is utilized to dynamically select the feature embedding which is fused with the deep represen... |
This work presents a simple post-training strategy to balance the worst-group and average-case accuracies of a classifier. In addition, it proposes a new metric to summarize this tradeoff. An extension on the simple method is also proposed to achieve better results. The proposed method outperforms several existing base... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This work presents a simple post-training strategy to balance the worst-group and average-case accuracies of a classifier. In addition, it proposes a new metric to summarize this tradeoff. An extension on the simple method is also proposed to achieve better results. The proposed method outperforms several exist... |
The paper on hand presents an approach for s-t graph cuts, relying on a differentiable quadratic approximation of the original problem. The approach is demonstrated for different applications.
Strength:
The idea of approximating the original linear problem by a special quadratic program, which is differentiable is int... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The paper on hand presents an approach for s-t graph cuts, relying on a differentiable quadratic approximation of the original problem. The approach is demonstrated for different applications.
Strength:
The idea of approximating the original linear problem by a special quadratic program, which is differentiabl... |
In this paper, the authors introduce a learn-to-optimize (L20)-based approach for quasi-Newton algorithms. The main idea is to learn on-the-fly an approximation of the inverse Hessian matrix modeled as the product of random permutation and block diagonal matrices. The proposed architecture called LODO consists of a lin... | Recommendation: 6: marginally above the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
In this paper, the authors introduce a learn-to-optimize (L20)-based approach for quasi-Newton algorithms. The main idea is to learn on-the-fly an approximation of the inverse Hessian matrix modeled as the product of random permutation and block diagonal matrices. The proposed architecture called LODO consists ... |
This paper is targeted at class-imbalanced node classification. The authors propose UNREAL, an over-sampling-based approach which iteratively adds unlabeled nodes into the training set. UNREAL consists of the Dual Pseudo-tag Alignment Mechanism (DPAM), Node-Reordering, and Discarding Geometrically Imbalanced nodes (DGI... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper is targeted at class-imbalanced node classification. The authors propose UNREAL, an over-sampling-based approach which iteratively adds unlabeled nodes into the training set. UNREAL consists of the Dual Pseudo-tag Alignment Mechanism (DPAM), Node-Reordering, and Discarding Geometrically Imbalanced no... |
The paper provides an interpretation of DINO as a mixture model with components with a circular von Mises-Fisher distribution.
Using this interpretation, the authors propose a normalization when computing cluster assignments, which improves stability and flexibility of the mixture model.
The paper proposes a very inter... | Recommendation: 8: accept, good paper | Area: Unsupervised and Self-supervised learning | Review:
The paper provides an interpretation of DINO as a mixture model with components with a circular von Mises-Fisher distribution.
Using this interpretation, the authors propose a normalization when computing cluster assignments, which improves stability and flexibility of the mixture model.
The paper proposes a ve... |
The paper aims to solve constrained monotone variational inequality when the projection step on the constraint is not simple. In particular, it uses the interior point method to handle the constraint, and then reformulate it to a form that can be solved by ADMM. For the convergence result, it provides guarantees for tw... | Recommendation: 6: marginally above the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
The paper aims to solve constrained monotone variational inequality when the projection step on the constraint is not simple. In particular, it uses the interior point method to handle the constraint, and then reformulate it to a form that can be solved by ADMM. For the convergence result, it provides guarantee... |
This article presents, Phenaki, an approach to text-to-video modelling and synthesis based on an autoregresserve transfomer-based encoder-decoder model.
# Stage 1. C-ViViT:
Frames from a given video are first linearly projected into patch-embeddings then processed by a spatial transformer; the resulting output of the... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
This article presents, Phenaki, an approach to text-to-video modelling and synthesis based on an autoregresserve transfomer-based encoder-decoder model.
# Stage 1. C-ViViT:
Frames from a given video are first linearly projected into patch-embeddings then processed by a spatial transformer; the resulting outpu... |
This paper proposes a neural network based approach to preserve and discover first integrals in the underlying target systems. In this paper, authors propose two instances of the approach, namely cFINDE and dFINDE, which works in continuous-time and discrete-time respectively. The proposed FINDE method is a great explo... | Recommendation: 8: accept, good paper | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper proposes a neural network based approach to preserve and discover first integrals in the underlying target systems. In this paper, authors propose two instances of the approach, namely cFINDE and dFINDE, which works in continuous-time and discrete-time respectively. The proposed FINDE method is a gre... |
Authors proposed a mosaic augmentation strategy for better representation learning in a self-supervised learning setup. The proposed MosRep strategy was applied to both MoCo-v2 and BYOL frameworks improving it's performance by a considerable margin on ImageNet-100 and ImageNet-1k datasets. The approach was further eval... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
Authors proposed a mosaic augmentation strategy for better representation learning in a self-supervised learning setup. The proposed MosRep strategy was applied to both MoCo-v2 and BYOL frameworks improving it's performance by a considerable margin on ImageNet-100 and ImageNet-1k datasets. The approach was furt... |
The authors present an alternative approach to image content representation called context clusters. These are compared as alternatives to convolutional networks and vision transformer approaches.
The work is well presented and well written. They argue for their approach in a scientific manner and set it in place for... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The authors present an alternative approach to image content representation called context clusters. These are compared as alternatives to convolutional networks and vision transformer approaches.
The work is well presented and well written. They argue for their approach in a scientific manner and set it in p... |
This paper proposes to apply the formalism of Lie groups to capture continuous transformationsto improve models robustness to distributional shifts. Specifically, it structures the representation of corresponding vector space of the assumed Lie group by learning some basis metrics and then constructs their Lie operator... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper proposes to apply the formalism of Lie groups to capture continuous transformationsto improve models robustness to distributional shifts. Specifically, it structures the representation of corresponding vector space of the assumed Lie group by learning some basis metrics and then constructs their Lie ... |
The authors propose a new task, music-to-text synaesthesia, in which audio features from musical tracks are extracted and interpreted as textual descriptions. The authors collect a dataset composed of classical recordings and a set of manually annotated textual descriptions, and evaluate a multi-modal encoder-decoder ... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The authors propose a new task, music-to-text synaesthesia, in which audio features from musical tracks are extracted and interpreted as textual descriptions. The authors collect a dataset composed of classical recordings and a set of manually annotated textual descriptions, and evaluate a multi-modal encoder-... |
This paper proposes a scheme to improve self-supervised learning for videos. This is done by randomly generating synthetic “saccades” which are expressed as binary masks applied to the input images. This masking and a modified loss allows to better capture semantic changes. Furthermore, a loss for semantic consistency ... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper proposes a scheme to improve self-supervised learning for videos. This is done by randomly generating synthetic “saccades” which are expressed as binary masks applied to the input images. This masking and a modified loss allows to better capture semantic changes. Furthermore, a loss for semantic cons... |
The paper argues that intermediate rewards are also crucial in the learning of Generative Flow Network, and proposes a novel GFlowNet learning framework, dubbed GAFlowNet, to incorporate intermediate rewards. The paper specifies intermediate rewards by intrinsic motivation to deal with the exploration of state space fo... | Recommendation: 6: marginally above the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The paper argues that intermediate rewards are also crucial in the learning of Generative Flow Network, and proposes a novel GFlowNet learning framework, dubbed GAFlowNet, to incorporate intermediate rewards. The paper specifies intermediate rewards by intrinsic motivation to deal with the exploration of state ... |
In this paper, the authors propose an adversarially robust neural Lyapunov control (ARNLC) method to improve the robustness and generalization capabilities for Lyapunov theory-based stability control. They claimed that the main contributions are:
[1] propose a perturbed Lyapunov risk for learning the control policy un... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
In this paper, the authors propose an adversarially robust neural Lyapunov control (ARNLC) method to improve the robustness and generalization capabilities for Lyapunov theory-based stability control. They claimed that the main contributions are:
[1] propose a perturbed Lyapunov risk for learning the control p... |
The paper proposes to view knowledge distillation from the Supervision Complexity angle, or how easy/hard is it for the student model to learn the targets. One reason why distillation has been thought to be helpful is that the soft labels additional provide information about class similarities which the one-hot labels ... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The paper proposes to view knowledge distillation from the Supervision Complexity angle, or how easy/hard is it for the student model to learn the targets. One reason why distillation has been thought to be helpful is that the soft labels additional provide information about class similarities which the one-hot... |
This paper presents a framework for deriving multi-lingual data from an execution-based dataset (e.g., MBPP, HumanEval, MathQA), which entails applying rule-based transformations based on static analysis to the prompt and test cases. The key idea is that the canonical solution does not need to be translated since evalu... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper presents a framework for deriving multi-lingual data from an execution-based dataset (e.g., MBPP, HumanEval, MathQA), which entails applying rule-based transformations based on static analysis to the prompt and test cases. The key idea is that the canonical solution does not need to be translated sin... |
The paper proposes a new graph neural network O-GNN for reasoning over molecular graphs that specifically encodes ring structures and claims and shows that this improves upon regular GNNs in terms of number of layers required. This is achieved by extending regular message passing GNNs by updating ring encodings (using... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper proposes a new graph neural network O-GNN for reasoning over molecular graphs that specifically encodes ring structures and claims and shows that this improves upon regular GNNs in terms of number of layers required. This is achieved by extending regular message passing GNNs by updating ring encoding... |
This paper proposed to incorporate Coulomb potential, London dispersion potential and Pauli repulsion potential in the input features of deep learning models, in order to deal with the long range interactions in PBC systems like crystal materials. It is proved that the proposed model outperforms SOTA in two material da... | Recommendation: 3: reject, not good enough | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper proposed to incorporate Coulomb potential, London dispersion potential and Pauli repulsion potential in the input features of deep learning models, in order to deal with the long range interactions in PBC systems like crystal materials. It is proved that the proposed model outperforms SOTA in two mat... |
This paper employs the tensorized embedding proposed in word2ket to the network embedding. It is similar with the second order part in network embedding method LINE, which is motivated from word2vec. Its superiority is verified on the tasks of topology reconstruction and link prediction.
Strength
- It is interesting t... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper employs the tensorized embedding proposed in word2ket to the network embedding. It is similar with the second order part in network embedding method LINE, which is motivated from word2vec. Its superiority is verified on the tasks of topology reconstruction and link prediction.
Strength
- It is inter... |
The authors study the curvature profile of latent representations through the layers of deep convolutional neural networks, similar to prior works on the intrinsic dimensions pioneered by Ansuini et al. (2019). The authors propose a plausible approach to densely sample the neighborhood around each input image, as neede... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The authors study the curvature profile of latent representations through the layers of deep convolutional neural networks, similar to prior works on the intrinsic dimensions pioneered by Ansuini et al. (2019). The authors propose a plausible approach to densely sample the neighborhood around each input image, ... |
This paper aims to find a good 3D CNN architecture for video recognition. This study proposes to use a novel proxy task, maximizing a designed entropy value, to search for the effective architecture. The experiments show the effectiveness of the proposed algorithm.
Strength
1. The experiments are solid, showing a good... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper aims to find a good 3D CNN architecture for video recognition. This study proposes to use a novel proxy task, maximizing a designed entropy value, to search for the effective architecture. The experiments show the effectiveness of the proposed algorithm.
Strength
1. The experiments are solid, showin... |
This paper proposes a regularization framework for training graph attention networks. The proposed method is based on the idea of performing edge interventions to determine the relative importance of an edge to specific prediction tasks. Extensive experiments show that the new regularization framework helps improve gen... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a regularization framework for training graph attention networks. The proposed method is based on the idea of performing edge interventions to determine the relative importance of an edge to specific prediction tasks. Extensive experiments show that the new regularization framework helps imp... |
This paper provides a theory for identifying latent causal structure using observational data when the weight-variant linear gaussian model assumption holds.
## Strength
1. This paper is very well and claerly written.
2. A detailed and interesting motivation in Section 3 helps a lot in understanding the problem.
3. Thi... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper provides a theory for identifying latent causal structure using observational data when the weight-variant linear gaussian model assumption holds.
## Strength
1. This paper is very well and claerly written.
2. A detailed and interesting motivation in Section 3 helps a lot in understanding the problem... |
The authors propose a new Federated Averaging algorithm, called z-SignFedAvg, that compresses the model updates to their signs after perturbing them with noise. They show both theoretically and empirically that z-SignFedAvg enjoys a faster convergence rate than existing sign-based compression methods. The authors furth... | Recommendation: 6: marginally above the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
The authors propose a new Federated Averaging algorithm, called z-SignFedAvg, that compresses the model updates to their signs after perturbing them with noise. They show both theoretically and empirically that z-SignFedAvg enjoys a faster convergence rate than existing sign-based compression methods. The autho... |
This paper presents a new state space model (SSM) that achieves on par with transformer based language models while inferring up to 2 times faster. The authors first highlight the drawback of previous SSMs compared to attention -- SSMs can't recall earlier tokens in the sequence and can't compare tokens across the sequ... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper presents a new state space model (SSM) that achieves on par with transformer based language models while inferring up to 2 times faster. The authors first highlight the drawback of previous SSMs compared to attention -- SSMs can't recall earlier tokens in the sequence and can't compare tokens across ... |
A Lipschitz-continuous generator of GAN cannot map a unimodal distribution to disconnected distributions. Therefore, in order to transform a unimodal distribution, e.g., Gaussian, of the latent variable to cover all modes of the disconnected target distribution, the generator must map a subset of the latent space outsi... | Recommendation: 3: reject, not good enough | Area: Generative models | Review:
A Lipschitz-continuous generator of GAN cannot map a unimodal distribution to disconnected distributions. Therefore, in order to transform a unimodal distribution, e.g., Gaussian, of the latent variable to cover all modes of the disconnected target distribution, the generator must map a subset of the latent spa... |
*I am a vision person and an emergency reviewer.*
This paper proposes to combine VICReg and a temporal verification loss for video representation learning, and inspect its impact on the data-efficiency of down-streaming reinforcement learning tasks. It is shown that the proposed method out-performs VICReg, MOCO and DI... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
*I am a vision person and an emergency reviewer.*
This paper proposes to combine VICReg and a temporal verification loss for video representation learning, and inspect its impact on the data-efficiency of down-streaming reinforcement learning tasks. It is shown that the proposed method out-performs VICReg, MOC... |
This paper tackles the task of domain shift in Visual Document Understanding (VDU) involving entity recognition, key-value extraction, and document visual question answering. DocTTA (document test time adaption) method is introduced which takes a pre-trained model from the source domain and adapts it to the target doma... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper tackles the task of domain shift in Visual Document Understanding (VDU) involving entity recognition, key-value extraction, and document visual question answering. DocTTA (document test time adaption) method is introduced which takes a pre-trained model from the source domain and adapts it to the tar... |
The paper provides a method to learn an embedding from sparse coding and dictionary learning through the sparse manifold transform. The authors show that this “white-box model” performs well compared to Deep Learning methods on classical image datasets like MNIST and CIFAR.
**Strengths**
The method proposed in the pap... | Recommendation: 8: accept, good paper | Area: Unsupervised and Self-supervised learning | Review:
The paper provides a method to learn an embedding from sparse coding and dictionary learning through the sparse manifold transform. The authors show that this “white-box model” performs well compared to Deep Learning methods on classical image datasets like MNIST and CIFAR.
**Strengths**
The method proposed in... |
This paper provides a method of combining the minimum description length principle to the finding of optimal policies for multitask RL problems. This allows the model to trade off between adapting to new information, and maintaining simplicity, thus adapting to epistemic uncertainty naturally.
After introducing the me... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper provides a method of combining the minimum description length principle to the finding of optimal policies for multitask RL problems. This allows the model to trade off between adapting to new information, and maintaining simplicity, thus adapting to epistemic uncertainty naturally.
After introducin... |
This paper addresses the many-domain generalization problem by treating each patient as a domain in healthcare applications. The paper proposes a model MANYDG to capture the domain and domain-invariant label representation by combining mutual reconstruction and orthogonal projection to explicitly remove the patient cov... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper addresses the many-domain generalization problem by treating each patient as a domain in healthcare applications. The paper proposes a model MANYDG to capture the domain and domain-invariant label representation by combining mutual reconstruction and orthogonal projection to explicitly remove the pat... |
The paper explores an approach to using existing (learned) skills in a new task in the same domain. The approach keeps the existing skills intact (i.e., does not modify them); learns a policy over them (called the "scheduler") that maximises current task reward; and, at the same time, learns another policy (called the ... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper explores an approach to using existing (learned) skills in a new task in the same domain. The approach keeps the existing skills intact (i.e., does not modify them); learns a policy over them (called the "scheduler") that maximises current task reward; and, at the same time, learns another policy (cal... |
CLARE uses a learned dynamics model and conservative policy objective in order to tackle the problem of learning a reward function from a dataset of expert trajectories, i.e. "offline IRL". Their contribution tackles both the reward extrapolation problem in offline IRL as well as improves generalization of the learned ... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
CLARE uses a learned dynamics model and conservative policy objective in order to tackle the problem of learning a reward function from a dataset of expert trajectories, i.e. "offline IRL". Their contribution tackles both the reward extrapolation problem in offline IRL as well as improves generalization of the ... |
This work studies long-tailed learning in the semantic level, which goes deeper beyond the conventional quantity bias. Specially, the authors explore to explain multiple phenomena by defining the semantic level imbalance and propose a measure by volume of manifold to reweight the learning objective, which has been demo... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This work studies long-tailed learning in the semantic level, which goes deeper beyond the conventional quantity bias. Specially, the authors explore to explain multiple phenomena by defining the semantic level imbalance and propose a measure by volume of manifold to reweight the learning objective, which has b... |
A purely theoretical contribution in which the authors present a series of approximation results describing which densities can be well-approximated by DBNs. Additionally, they provide quantitative error bounds for both Lp based norms and the KL divergence.
Strengths: Fills gaps in the theoretical literature on what ... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
A purely theoretical contribution in which the authors present a series of approximation results describing which densities can be well-approximated by DBNs. Additionally, they provide quantitative error bounds for both Lp based norms and the KL divergence.
Strengths: Fills gaps in the theoretical literature ... |
In this paper, the authors observed that RvS, their main baseline method, does not do well on Antmaze when not given privileged information such as the goal location. They hypothesized that by modifying the reward to do bootstrapping, similar to other methods based on temporal difference learning(e.g. CQL, BCQ), that w... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
In this paper, the authors observed that RvS, their main baseline method, does not do well on Antmaze when not given privileged information such as the goal location. They hypothesized that by modifying the reward to do bootstrapping, similar to other methods based on temporal difference learning(e.g. CQL, BCQ)... |
The paper proposes to leverage a pool of pretrained models without expensive fine-tuning. They empirically demonstrate the proposed method achieves SOTA results on several benchmarks without significant loss in inference speed.
**Strengths:**
The authors propose a creative way how to use a pool of pretrained models ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes to leverage a pool of pretrained models without expensive fine-tuning. They empirically demonstrate the proposed method achieves SOTA results on several benchmarks without significant loss in inference speed.
**Strengths:**
The authors propose a creative way how to use a pool of pretrained... |
This paper proposes gradient annealing method in dynamic sparse training to improve the performance. Based on gradient annealing, the author of the paper proposes AutoSparse algorithm that uses a learnable threshold to find better sparse topology and sparsity distribution during DST. Multiple experiments are conducted ... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper proposes gradient annealing method in dynamic sparse training to improve the performance. Based on gradient annealing, the author of the paper proposes AutoSparse algorithm that uses a learnable threshold to find better sparse topology and sparsity distribution during DST. Multiple experiments are co... |
The paper proposes a vision-language-action grounding framework for robot manipulation, with the guidance of a domain specific language Combinatory Categorial Grammar (CCG). CCG first parse the natural language to more structured program, and associate the objects and targets via the grounding module. Then it produces ... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper proposes a vision-language-action grounding framework for robot manipulation, with the guidance of a domain specific language Combinatory Categorial Grammar (CCG). CCG first parse the natural language to more structured program, and associate the objects and targets via the grounding module. Then it p... |
The paper is proposing a forward feature selection algorithm with the use of attention mechanism for assessing the feature relevance of currently not selected features. Since attention mechanism consider all unselected features at once, there is computational saving, compared to the common greedy forward selection algo... | Recommendation: 8: accept, good paper | Area: General Machine Learning | Review:
The paper is proposing a forward feature selection algorithm with the use of attention mechanism for assessing the feature relevance of currently not selected features. Since attention mechanism consider all unselected features at once, there is computational saving, compared to the common greedy forward select... |
The authors study the approximation of Nash equilibrium in very large extensive-form games with prefect recall. One line of research approximates with the use of neural networks the tabular algorithm: counterfactual regret minimization (CFR). One issue with these methods is that they train a neural network with target... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The authors study the approximation of Nash equilibrium in very large extensive-form games with prefect recall. One line of research approximates with the use of neural networks the tabular algorithm: counterfactual regret minimization (CFR). One issue with these methods is that they train a neural network wit... |
This paper proposes an aggregation-aware mixed-precision quantization method for GNNs, which quantizes different nodes features with different learnable quantization parameters, including bit-width and step-size. The paper also proposes a Nearest Neighbor Strategy to deal with the generalization on unseen graphs. Empir... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper proposes an aggregation-aware mixed-precision quantization method for GNNs, which quantizes different nodes features with different learnable quantization parameters, including bit-width and step-size. The paper also proposes a Nearest Neighbor Strategy to deal with the generalization on unseen graph... |
In the development of social sciences to investigate human societies, measurement models in the social sciences did not keep in mind the deep societal reach of algorithms. This calls for new methods to measure feasible information. However, this paper shows that innovative approaches to obtaining such fair and trustwo... | Recommendation: 5: marginally below the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
In the development of social sciences to investigate human societies, measurement models in the social sciences did not keep in mind the deep societal reach of algorithms. This calls for new methods to measure feasible information. However, this paper shows that innovative approaches to obtaining such fair and... |
Graph transformers have drawn growing attention in representation learning on graph-structured data. However, their success is limited to small graphs due to the quadratic complexity of the dot-product attention module in transformers. This paper proposes the Deformable Graph Transformer (DGT) to address this issue, by... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
Graph transformers have drawn growing attention in representation learning on graph-structured data. However, their success is limited to small graphs due to the quadratic complexity of the dot-product attention module in transformers. This paper proposes the Deformable Graph Transformer (DGT) to address this i... |
This paper considers a concatenation of all intermediate variables through a forward diffusion as a latent code.
The authors consider the forward process of diffusion models as an encoder.
Using two independently trained diffusion models, the proposed method encodes an image with a fixed random seed for one model and... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper considers a concatenation of all intermediate variables through a forward diffusion as a latent code.
The authors consider the forward process of diffusion models as an encoder.
Using two independently trained diffusion models, the proposed method encodes an image with a fixed random seed for one m... |
This paper proposes a diffusion model-based data augmentation method. The proposed method combines the contrastive loss with the diffusion model generation process to generate data with a large diversity between different classes. It shows that using the proposed method generates data and trains a classifier, the train... | Recommendation: 3: reject, not good enough | Area: Generative models | Review:
This paper proposes a diffusion model-based data augmentation method. The proposed method combines the contrastive loss with the diffusion model generation process to generate data with a large diversity between different classes. It shows that using the proposed method generates data and trains a classifier, t... |
In this paper, the authors present SUBSELNET - a non- adaptive subset selection framework for solving a particular aspect of subset selection problem - improving generalizability of the subset selection approach; with existing methods, the algorithm has to be executed from the beginning for each new model.
The authors ... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
In this paper, the authors present SUBSELNET - a non- adaptive subset selection framework for solving a particular aspect of subset selection problem - improving generalizability of the subset selection approach; with existing methods, the algorithm has to be executed from the beginning for each new model.
The ... |
The paper introduces a DiffPure inspired method (Diffusion models for adversarial robustness) in the context of speech-reco acoustic models. The adaptation of DiffPure, called AudioPure, operates in time domain with a diffusion process before the stft/mel feature extraction. The paper focuses on measuring the performan... | 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 DiffPure inspired method (Diffusion models for adversarial robustness) in the context of speech-reco acoustic models. The adaptation of DiffPure, called AudioPure, operates in time domain with a diffusion process before the stft/mel feature extraction. The paper focuses on measuring the p... |
This paper performs a benchmark of active learning for deep nets (aka DAL).
The authors make the distinction between _supervised_ methods and _semi-supervised_ methods (respectively SAL and SSAL) by the ability to harness unlabeled data.
The proposed benchmark is comprehensive (as in the amount of methods tested) and ... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
This paper performs a benchmark of active learning for deep nets (aka DAL).
The authors make the distinction between _supervised_ methods and _semi-supervised_ methods (respectively SAL and SSAL) by the ability to harness unlabeled data.
The proposed benchmark is comprehensive (as in the amount of methods test... |
The paper debiases models by resampling with probabilities proportional to gradient norms. It uses the generalized cross-entropy loss from LfF to train a biased model and debiases the main model by sampling probabilities from gradient norms. It leverages the observation that rarer samples tend to have higher gradient n... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The paper debiases models by resampling with probabilities proportional to gradient norms. It uses the generalized cross-entropy loss from LfF to train a biased model and debiases the main model by sampling probabilities from gradient norms. It leverages the observation that rarer samples tend to have higher gr... |
The paper proposes a new method to improve calibration of multi-class classifiers. It is similar to the existing methods of MMCE and SB-ECE in the sense that it adds a trainable auxiliary loss function (Expected Squared Difference, ESD) to the cross-entropy loss (NLL) during training of the classifier. Similarly to SB-... | 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 a new method to improve calibration of multi-class classifiers. It is similar to the existing methods of MMCE and SB-ECE in the sense that it adds a trainable auxiliary loss function (Expected Squared Difference, ESD) to the cross-entropy loss (NLL) during training of the classifier. Similarl... |
This paper conducts empirical investigation regarding the best practice of building cascade models. Specifically, the paper investigated the impact of early-exist condition and the choice of model calibration on the efficiency-accuracy performance of the cascade, and investigated the performance of both vision and text... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper conducts empirical investigation regarding the best practice of building cascade models. Specifically, the paper investigated the impact of early-exist condition and the choice of model calibration on the efficiency-accuracy performance of the cascade, and investigated the performance of both vision ... |
In this paper, the authors show how an appropriate basis function decomposition can be used to provide a much simpler convergence analysis for gradient-based algorithms on several representative learning problems, from simple kernel regression to complex DNNs. 1) The authors prove that GD learns the coefficients of an ... | Recommendation: 6: marginally above the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
In this paper, the authors show how an appropriate basis function decomposition can be used to provide a much simpler convergence analysis for gradient-based algorithms on several representative learning problems, from simple kernel regression to complex DNNs. 1) The authors prove that GD learns the coefficient... |
The paper proposes what they call "the best possible operator" for decentralized Q-learning when there are multiple agents. The operator updates the policies of each agent individually at each step and can be shown to converge to the optimal joint policy. However, the operator is too complex and impractical. What the p... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper proposes what they call "the best possible operator" for decentralized Q-learning when there are multiple agents. The operator updates the policies of each agent individually at each step and can be shown to converge to the optimal joint policy. However, the operator is too complex and impractical. Wh... |
The authors propose an unsupervised denoising method for situations in the high noise regime with side information. They frame the problem for time-series applications, though the proposed method appears general. Their method is a variant of the Noise2Noise method (Lehtinen et al 2018), which itself builds off of Denoi... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The authors propose an unsupervised denoising method for situations in the high noise regime with side information. They frame the problem for time-series applications, though the proposed method appears general. Their method is a variant of the Noise2Noise method (Lehtinen et al 2018), which itself builds off ... |
This paper studies off-policy learning when there is distribution shift between the data distribution and target distribution. The paper points out that the vanilla inverse propensity score (IPS) estimator suffers from large variance and bias when the estimated data distribution is inaccurate. To this end, this paper p... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper studies off-policy learning when there is distribution shift between the data distribution and target distribution. The paper points out that the vanilla inverse propensity score (IPS) estimator suffers from large variance and bias when the estimated data distribution is inaccurate. To this end, this... |
The authors propose a new estimator for debiasing Recommender Systems called Targeted Doubly Robust (TDR) which addresses some limitations of the Doubly Robust (DR) and Error Imputation Based (EIB) estimators.
They show first that DR has a large variance and is sensitive to small propensities. The variance of DR, whi... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The authors propose a new estimator for debiasing Recommender Systems called Targeted Doubly Robust (TDR) which addresses some limitations of the Doubly Robust (DR) and Error Imputation Based (EIB) estimators.
They show first that DR has a large variance and is sensitive to small propensities. The variance of... |
The paper first highlights the observation that the policy space size can be prohibitive for policy search and off-policy evaluation when the state-action space is large (e.g., high-dimensional continuous). Then, the paper aims to propose a unified theory of policy abstractions for determining equivalence between give... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper first highlights the observation that the policy space size can be prohibitive for policy search and off-policy evaluation when the state-action space is large (e.g., high-dimensional continuous). Then, the paper aims to propose a unified theory of policy abstractions for determining equivalence betw... |
This paper defines three types of equivariance: correct, incorrect and extrinsic. The authors prove an upper bound on the accuracy when incorrect equivariance is used. Then the authors show empirically that using incorrect equivariance (e.g. using an equivariant neural network with incorrect labels) is worse than using... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper defines three types of equivariance: correct, incorrect and extrinsic. The authors prove an upper bound on the accuracy when incorrect equivariance is used. Then the authors show empirically that using incorrect equivariance (e.g. using an equivariant neural network with incorrect labels) is worse th... |
The authors propose a Vision language model that can use the external knowledge graph ConceptNet. The model consists of a transformer model with a graph-based model to embed graph information. The authors claim the proposed model can outperform state-of-the-art models in image captioning.
Strengths:
- The model design... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The authors propose a Vision language model that can use the external knowledge graph ConceptNet. The model consists of a transformer model with a graph-based model to embed graph information. The authors claim the proposed model can outperform state-of-the-art models in image captioning.
Strengths:
- The mode... |
In this paper, the authors study a class of contextual multi-armed bandit (MAB) problems where the input space is high-dimensional and the action space is continuous. This is a common setting in medical diagnostics where, say, a drug dosing regime must be chosen baed on medical diagnostics/imaging. The continuous actio... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
In this paper, the authors study a class of contextual multi-armed bandit (MAB) problems where the input space is high-dimensional and the action space is continuous. This is a common setting in medical diagnostics where, say, a drug dosing regime must be chosen baed on medical diagnostics/imaging. The continuo... |
Training deep models to correctly detect and outline rotated objects is difficult due to the nature of the metric of reference (SkewIoU). The authors propose KFIoU, an easily computable and differentiable loss that relies on transforming bounding boxes into Gaussians and estimating the overlap between the Gaussians ana... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
Training deep models to correctly detect and outline rotated objects is difficult due to the nature of the metric of reference (SkewIoU). The authors propose KFIoU, an easily computable and differentiable loss that relies on transforming bounding boxes into Gaussians and estimating the overlap between the Gauss... |
This paper proposes a representation learning framework for identifying agent and object representations from observations in the scenario of an agent interacting with an object. In particular, the work aims to learn an isometric representation (i.e., capture underlying geometric states in a loss-less fashion) with obj... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a representation learning framework for identifying agent and object representations from observations in the scenario of an agent interacting with an object. In particular, the work aims to learn an isometric representation (i.e., capture underlying geometric states in a loss-less fashion) ... |
The submission introduces an agent called MEME, which is built on Agent57, which in turn is a super inefficient atari player that requires a ridiculous number of training samples, not even remotely being sensible. Numerous ad-hoc or semi ad-hoc techniques are used (most of which with little to no formal ground) to miti... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The submission introduces an agent called MEME, which is built on Agent57, which in turn is a super inefficient atari player that requires a ridiculous number of training samples, not even remotely being sensible. Numerous ad-hoc or semi ad-hoc techniques are used (most of which with little to no formal ground)... |
This work addresses three issues that current MLPs on graph methods have: 1) the misalignment between content feature and label spaces, 2) the strict hard matching to the teacher’s output, and 3) the sensitivity to node feature noises. In particular, this paper presents NOSMOG, a novel method to learn MLPs with remarka... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This work addresses three issues that current MLPs on graph methods have: 1) the misalignment between content feature and label spaces, 2) the strict hard matching to the teacher’s output, and 3) the sensitivity to node feature noises. In particular, this paper presents NOSMOG, a novel method to learn MLPs with... |
The paper proposes a method to poison the generative models used in DGR in an effort to degrade continual learning methods using DGR.
### Strengths:
1. I agree with the authors that the poisoning attacks on continual learning systems are understudied (especially with the overlap with FL here).
2. The high-level idea ... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
The paper proposes a method to poison the generative models used in DGR in an effort to degrade continual learning methods using DGR.
### Strengths:
1. I agree with the authors that the poisoning attacks on continual learning systems are understudied (especially with the overlap with FL here).
2. The high-lev... |
This paper applies stream-based active learning to datasets with time stamps. The main idea is to select data points with a large loss change $\left\lVert \frac{d}{dt} \hat{\mathcal{L}} \right\rVert$, where $\hat{\mathcal{L}}$ is an estimated loss. The authors propose three methods by following this policy, and validat... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper applies stream-based active learning to datasets with time stamps. The main idea is to select data points with a large loss change $\left\lVert \frac{d}{dt} \hat{\mathcal{L}} \right\rVert$, where $\hat{\mathcal{L}}$ is an estimated loss. The authors propose three methods by following this policy, and... |
The authors propose a method called "Soft Sampling" which is a data selection method for training of deep learning models. They also show some theoretical convergence guarantees and experimentally show that their method achieves similar accuracy as GradMatch (a SoTA method) while using less resources (as measured by wa... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The authors propose a method called "Soft Sampling" which is a data selection method for training of deep learning models. They also show some theoretical convergence guarantees and experimentally show that their method achieves similar accuracy as GradMatch (a SoTA method) while using less resources (as measur... |
This paper proposed a technique to perform federated learning in the presence of non-IID data across clients. The main claim of the paper is to adjust the learning rate at the server side depending upon the similarity of the skewed gradients from non-iid data clients. The authors claim to achieve higher accuracy as com... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposed a technique to perform federated learning in the presence of non-IID data across clients. The main claim of the paper is to adjust the learning rate at the server side depending upon the similarity of the skewed gradients from non-iid data clients. The authors claim to achieve higher accurac... |
With the increase in the complexity of machine learning models, it becomes crucial to explain model decisions and increase transparency between human-model interactions. To this end, several explanation methods have been proposed in recent literature for explaining model predictions, but few works discuss necessary des... | Recommendation: 5: marginally below the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
With the increase in the complexity of machine learning models, it becomes crucial to explain model decisions and increase transparency between human-model interactions. To this end, several explanation methods have been proposed in recent literature for explaining model predictions, but few works discuss neces... |
This paper considers a setting of contextual bandits where the space of actions is continuous and contexts are large. Different from existing approaches in contextual bandits, the paper proposes to modify the DDPG algorithm, a popular algorithm in reinforcement learning into a context bandit algorithm with continuous a... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper considers a setting of contextual bandits where the space of actions is continuous and contexts are large. Different from existing approaches in contextual bandits, the paper proposes to modify the DDPG algorithm, a popular algorithm in reinforcement learning into a context bandit algorithm with cont... |
This paper considers the problem of MaxEnt reinforcement learning, with the goal of increasing the expressiveness of the policy while still allowing for practical entropy maximization. The proposed method relies on latent variables to avoid complexities with using EBMs.
The challenge with directly using latent variab... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper considers the problem of MaxEnt reinforcement learning, with the goal of increasing the expressiveness of the policy while still allowing for practical entropy maximization. The proposed method relies on latent variables to avoid complexities with using EBMs.
The challenge with directly using laten... |
The authors introduce an unsupervised probe for knowledge in large language models to tackle a perceived challenge in verifying truthfulness of predictions/generated text. The authors' proposed method involves classifying/ranking model activations between binary labels/framings of a yes-no question. The authors demonst... | 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 authors introduce an unsupervised probe for knowledge in large language models to tackle a perceived challenge in verifying truthfulness of predictions/generated text. The authors' proposed method involves classifying/ranking model activations between binary labels/framings of a yes-no question. The authors... |
This paper proposes a metric for comparing learned representations, and furthermore doing it so that models can be switched based on data. That sounds useful for a variety of purposes. That said:
This paper was assigned to a completely wrong reviewer; an indication of some sort of a failure in the ICLR process. I don'... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper proposes a metric for comparing learned representations, and furthermore doing it so that models can be switched based on data. That sounds useful for a variety of purposes. That said:
This paper was assigned to a completely wrong reviewer; an indication of some sort of a failure in the ICLR process... |
The paper studies the application of denoising diffusion probabilistic models to seq2seq text generation. Authors propose DiffuSeq - a diffusion-based model for seq2seq task using a technique they call "partially noising". Authors also show the theoretical connection between DiffuSeq and standard AR / NAR language mod... | Recommendation: 5: marginally below the acceptance threshold | Area: Generative models | Review:
The paper studies the application of denoising diffusion probabilistic models to seq2seq text generation. Authors propose DiffuSeq - a diffusion-based model for seq2seq task using a technique they call "partially noising". Authors also show the theoretical connection between DiffuSeq and standard AR / NAR lang... |
This paper proposes an attention-based model called the mesh-independent operator learner (MIOL) to provide proper treatments of input functions and query coordinates for the solution functions by detaching the dependence on input and output meshes. The proposed models pre-trained with benchmark datasets of operator le... | 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 an attention-based model called the mesh-independent operator learner (MIOL) to provide proper treatments of input functions and query coordinates for the solution functions by detaching the dependence on input and output meshes. The proposed models pre-trained with benchmark datasets of ope... |
This paper proposes a lossy image compression algorithm. The framework is similar to previous work of transform coding built on VAEs, and the proposed method replaced the the decoder with a conditional diffusion model. The paper provided comprehensive evaluations on various datasets and a number of metrics, and demonst... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
This paper proposes a lossy image compression algorithm. The framework is similar to previous work of transform coding built on VAEs, and the proposed method replaced the the decoder with a conditional diffusion model. The paper provided comprehensive evaluations on various datasets and a number of metrics, and... |
This paper presents a unified framework for human voice synthesis based on neural analysis and synthesis modules. The analysis part extracts disentangled and controllable voice features based on domain knowledge (pitch, periodicity/aperiodicity and timbre). Each of them is trained independently with a dedicated neural ... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper presents a unified framework for human voice synthesis based on neural analysis and synthesis modules. The analysis part extracts disentangled and controllable voice features based on domain knowledge (pitch, periodicity/aperiodicity and timbre). Each of them is trained independently with a dedicated... |
This paper studies the Position-information Pattern from Padding (PPP) encoded in neural networks. PPP is defined as difference between algorithmically and optimally padded activations, then SNR and MAE are computed on the difference to quantify how much PPP exists. The result shows a pre-trained CNN model contains a l... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper studies the Position-information Pattern from Padding (PPP) encoded in neural networks. PPP is defined as difference between algorithmically and optimally padded activations, then SNR and MAE are computed on the difference to quantify how much PPP exists. The result shows a pre-trained CNN model cont... |
This paper propses a method for prompt tuning that is less sensitive to overfitting compared to the CoOp approach. The main idea is simple and effective, and is based on the observation that CoOp starts overftting when the training is not stoped early, outpefroming Zero-shot CLIP. The authors propose to requalarize t... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper propses a method for prompt tuning that is less sensitive to overfitting compared to the CoOp approach. The main idea is simple and effective, and is based on the observation that CoOp starts overftting when the training is not stoped early, outpefroming Zero-shot CLIP. The authors propose to requa... |
The authors introduce a method to learn the parametrization of an Evolution Strategy through the meta-training of a self-attention-based architecture. They demonstrate the generalization of their meta-training procedure to test tasks that are different from the meta-training tasks. They further analyze the limitations ... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
The authors introduce a method to learn the parametrization of an Evolution Strategy through the meta-training of a self-attention-based architecture. They demonstrate the generalization of their meta-training procedure to test tasks that are different from the meta-training tasks. They further analyze the limi... |
This paper uses a classical tool in numerical analysis, namely modified equation, to account for the effect of small but finite learning rate used in (stochastic) gradient descent with momentum. As a result of this analysis and as a standard treatment, the continuous time limit of the optimizer, which is an ODE, will h... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper uses a classical tool in numerical analysis, namely modified equation, to account for the effect of small but finite learning rate used in (stochastic) gradient descent with momentum. As a result of this analysis and as a standard treatment, the continuous time limit of the optimizer, which is an ODE... |
This paper proposes a method that allows replacing the text encoder in a text-to-image model without retraining the entire model. It uses a Model Translation Network (MTN) that is trained with text corpus data (or parallel text data for multilingual text encoders) to align the representation spaces of two text encoders... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a method that allows replacing the text encoder in a text-to-image model without retraining the entire model. It uses a Model Translation Network (MTN) that is trained with text corpus data (or parallel text data for multilingual text encoders) to align the representation spaces of two text ... |
The authors propose an equivariant graph contrastive learning framework that adopts two principles: invariance to intra-graph
augmentations and equivariance to cross-graph augmentations. They use Mixup as the method for cross-graph augmentation and argue that the cross-graph augmentation captures global semantic shifts... | Recommendation: 3: reject, not good enough | Area: Unsupervised and Self-supervised learning | Review:
The authors propose an equivariant graph contrastive learning framework that adopts two principles: invariance to intra-graph
augmentations and equivariance to cross-graph augmentations. They use Mixup as the method for cross-graph augmentation and argue that the cross-graph augmentation captures global semanti... |
This paper initiates the study for a buyer with budgets to learn online bidding strategies in repeated first-price auctions. They propose an RL-based bidding algorithm against the optimal non-anticipating strategy under stationary competition. The algorithm obtains T^{1/2}-regret if the bids are all revealed at the end... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper initiates the study for a buyer with budgets to learn online bidding strategies in repeated first-price auctions. They propose an RL-based bidding algorithm against the optimal non-anticipating strategy under stationary competition. The algorithm obtains T^{1/2}-regret if the bids are all revealed at... |
This paper proposes a neural network based machine learning method for deciding binary variables at each search step. Some design decisions such as layers of the neural network, the loss functions and etc are made. Some numerical results to demonstrate the performance of the proposed method are given.
Strengths:
The r... | Recommendation: 3: reject, not good enough | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper proposes a neural network based machine learning method for deciding binary variables at each search step. Some design decisions such as layers of the neural network, the loss functions and etc are made. Some numerical results to demonstrate the performance of the proposed method are given.
Strengths... |
This paper provides a practical framework for heterogeneous federated learning (HFL). Existing HFL frameworks suffer from privacy leakage through query data communication. Compared to prior work in HFL or adaptation with privacy-preserving feature, the proposed framework 1) is lightweight and efficient, 2) provides for... | 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 provides a practical framework for heterogeneous federated learning (HFL). Existing HFL frameworks suffer from privacy leakage through query data communication. Compared to prior work in HFL or adaptation with privacy-preserving feature, the proposed framework 1) is lightweight and efficient, 2) prov... |
This paper talks about ZSL. The authors ground their motivation on the fact that in-class samples could be still very different from each other, and they propose to tackle this problem in ZSL. The main contribution yields the concept of Metric Learning with Implicit Semantics (MLIS) and it is plugged into the training ... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper talks about ZSL. The authors ground their motivation on the fact that in-class samples could be still very different from each other, and they propose to tackle this problem in ZSL. The main contribution yields the concept of Metric Learning with Implicit Semantics (MLIS) and it is plugged into the t... |
This work empirically investigates various types of image misclassification from a game-theoretic view, which includes clean images, adversarially perturbed images and corrupted images. It characterizes the misclassification with the distribution, order, and sign of the interactions, and numerical results show three ty... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This work empirically investigates various types of image misclassification from a game-theoretic view, which includes clean images, adversarially perturbed images and corrupted images. It characterizes the misclassification with the distribution, order, and sign of the interactions, and numerical results show ... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.