review stringlengths 5 16.6k | score stringclasses 5
values | area stringclasses 12
values | text stringlengths 31 5.65k |
|---|---|---|---|
Many Factorized Bilinear Pooling (FBiP) uses Hadamard product-based bilinear projection. This paper reveals that the Hadamard product misses a lot of possible projection directions. This paper proposes a general matrix-based bilinear projection based on rank-k matrix base decomposition. This paper uses the proposed bil... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
Many Factorized Bilinear Pooling (FBiP) uses Hadamard product-based bilinear projection. This paper reveals that the Hadamard product misses a lot of possible projection directions. This paper proposes a general matrix-based bilinear projection based on rank-k matrix base decomposition. This paper uses the prop... |
The paper proposes the use of L-length binary codes (TAC) for the use of multi-class classification. Instead of having a one-hot representation, the authors' approach is to produce a binary value from each sliced-up section of the neural network's activations (Figure 1). Models can be trained to output TAC from scratch... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The paper proposes the use of L-length binary codes (TAC) for the use of multi-class classification. Instead of having a one-hot representation, the authors' approach is to produce a binary value from each sliced-up section of the neural network's activations (Figure 1). Models can be trained to output TAC from... |
This paper presents a new method for utilizing hyperbolic geometry that avoids many of the issues with current models.
My understanding of it is as follows. Suppose we have $x_1, \ldots, x_n$ embedded into some hyperbolic space and some kernel map $k(x_i,x_j)$ that **only** depends on the hyperbolic distance between ... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper presents a new method for utilizing hyperbolic geometry that avoids many of the issues with current models.
My understanding of it is as follows. Suppose we have $x_1, \ldots, x_n$ embedded into some hyperbolic space and some kernel map $k(x_i,x_j)$ that **only** depends on the hyperbolic distance ... |
This paper discusses the post-training quantization (PTQ) methodology of a Generative Pre-trained Transformer (GPT) model for a generation task. While formatting the weight to 4 or 3-bit, activation remains as FP. The PTQ methodology is a layer-wise quantization with a small calibration-set, and a part of the weight is... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper discusses the post-training quantization (PTQ) methodology of a Generative Pre-trained Transformer (GPT) model for a generation task. While formatting the weight to 4 or 3-bit, activation remains as FP. The PTQ methodology is a layer-wise quantization with a small calibration-set, and a part of the w... |
This paper proposed a novel neural point-based rendering framework that can achieve view synthesis quality comparable to volume rendering based methods while being 100x faster. The method is based on existing approach but with non-trivial novel techniques that significantly improve the rendering quality, including SH f... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposed a novel neural point-based rendering framework that can achieve view synthesis quality comparable to volume rendering based methods while being 100x faster. The method is based on existing approach but with non-trivial novel techniques that significantly improve the rendering quality, includ... |
This paper proposes a new approach for Federated learning. The idea is to first distill the private data of each client (with respect to the current initialization of their model weights) and send them to the server. On the server side, a global model will learn from these synthetic data and broadcast the learned weigh... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
This paper proposes a new approach for Federated learning. The idea is to first distill the private data of each client (with respect to the current initialization of their model weights) and send them to the server. On the server side, a global model will learn from these synthetic data and broadcast the learn... |
In this paper, the authors aim to solve the measurement problem of graph-level clustering, and proposes an end-to-end method to jointly optimize graph representation and clustering, and the clustering goal can guide the learning of the entire graph, which is more effective than two stages.
Strength: An end-to-end metho... | Recommendation: 3: reject, not good enough | Area: Unsupervised and Self-supervised learning | Review:
In this paper, the authors aim to solve the measurement problem of graph-level clustering, and proposes an end-to-end method to jointly optimize graph representation and clustering, and the clustering goal can guide the learning of the entire graph, which is more effective than two stages.
Strength: An end-to-e... |
The paper presents an extension to the idea of federated learning which works in the situation of heterogeneous data. This is achieved without the need for public data. The approach provides security through the use of differential privacy for the exchanged data.
Strengths:
1. The paper is technically well presented.
2... | Recommendation: 8: accept, good paper | Area: General Machine Learning | Review:
The paper presents an extension to the idea of federated learning which works in the situation of heterogeneous data. This is achieved without the need for public data. The approach provides security through the use of differential privacy for the exchanged data.
Strengths:
1. The paper is technically well pres... |
1. The paper proposed a new user-driven algorithmic framework PROBE (Probabilistically Robust Recourse) that tackles two independently well-studied challenges in algorithmic recourse -- (1) recourse cost, and (2) robustness under noisy human implementation. This is the first study where both recourse cost and robustne... | Recommendation: 6: marginally above the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
1. The paper proposed a new user-driven algorithmic framework PROBE (Probabilistically Robust Recourse) that tackles two independently well-studied challenges in algorithmic recourse -- (1) recourse cost, and (2) robustness under noisy human implementation. This is the first study where both recourse cost and ... |
This paper proposes uses an actor-critic for adapting the simulation mesh, specifically to add or collapse edges. The approach explicitly makes a trade-off between error and computational cost, allowing to tune it at test time.
**Strengths**
The main advantage of the proposed approach is being to trained the re-meshi... | Recommendation: 8: accept, good paper | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper proposes uses an actor-critic for adapting the simulation mesh, specifically to add or collapse edges. The approach explicitly makes a trade-off between error and computational cost, allowing to tune it at test time.
**Strengths**
The main advantage of the proposed approach is being to trained the ... |
This paper proposed a probabilistic graph learning framework named CLEP, which unifies graph generative models with contrastive learning paradigm to extract intra- and inter-graph information. CLEP consists of two major components: 1) the graph generative model, which defines a set of encoders to capture the hidden gra... | Recommendation: 3: reject, not good enough | Area: Unsupervised and Self-supervised learning | Review:
This paper proposed a probabilistic graph learning framework named CLEP, which unifies graph generative models with contrastive learning paradigm to extract intra- and inter-graph information. CLEP consists of two major components: 1) the graph generative model, which defines a set of encoders to capture the hi... |
This paper provides new sample complexity guarantees for convolutional neural networks applied to off-policy Q-learning of MDPs with low-dimensional manifold structure.
The work is novel in that it provides a new characterization of the sample-complexity in terms of the underlying manifold dimension rather than the t... | Recommendation: 8: accept, good paper | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper provides new sample complexity guarantees for convolutional neural networks applied to off-policy Q-learning of MDPs with low-dimensional manifold structure.
The work is novel in that it provides a new characterization of the sample-complexity in terms of the underlying manifold dimension rather th... |
This paper replaces the value target in TD-based methods with a lower bound of the optimal value function. Convergence was shown in the tabular case and the authors introduced several lower bounds that can be applied to practical algorithms. The authors demonstrated the effectiveness of their method on a series of chal... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper replaces the value target in TD-based methods with a lower bound of the optimal value function. Convergence was shown in the tabular case and the authors introduced several lower bounds that can be applied to practical algorithms. The authors demonstrated the effectiveness of their method on a series... |
The paper introduces a method to extend neural attention to longer text with limited computations. With inspiration from the P2P Chord protocol, it rotates the token indices to allow the model to attend to tokens in different scales of distances. It is proved to have a large receptive field compared to traditional tran... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper introduces a method to extend neural attention to longer text with limited computations. With inspiration from the P2P Chord protocol, it rotates the token indices to allow the model to attend to tokens in different scales of distances. It is proved to have a large receptive field compared to traditio... |
This paper proposes a method called “Radial Spike and Slab Bayesian Neural Network (BNN)” to detect ransomware attacks when the data is sparse and imbalanced. This method is a type of BNN with the posterior distribution represented by a mixture of distributions resulting in a Radial Spike and Slab distribution. The usa... | Recommendation: 6: marginally above the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
This paper proposes a method called “Radial Spike and Slab Bayesian Neural Network (BNN)” to detect ransomware attacks when the data is sparse and imbalanced. This method is a type of BNN with the posterior distribution represented by a mixture of distributions resulting in a Radial Spike and Slab distribution.... |
The paper introduces a deep evolutionary convolution network (DECN) for continuous black-box optimization. DECN is composed of two modules: convolution-based reasoning module (CRM) and selection module (SM). The paper describes both modules and shows how to integrate them. It also contains a description of the process ... | Recommendation: 6: marginally above the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
The paper introduces a deep evolutionary convolution network (DECN) for continuous black-box optimization. DECN is composed of two modules: convolution-based reasoning module (CRM) and selection module (SM). The paper describes both modules and shows how to integrate them. It also contains a description of the ... |
The authors attempt to understand catastrophic overfitting in adversarial training with an FGSM adversary. To analyze catastrophic overfitting, they proposed to induce it using controlled manipulations of the dataset technique. Depending on the magnitude of the data perturbations, they observed that it is possible to i... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The authors attempt to understand catastrophic overfitting in adversarial training with an FGSM adversary. To analyze catastrophic overfitting, they proposed to induce it using controlled manipulations of the dataset technique. Depending on the magnitude of the data perturbations, they observed that it is possi... |
The paper proposes a method to generate post-hoc recourses which remain valid under model shift, called DiRRAc. The model parameters are considered as a random vector modeled according to a mixture distribution (whose parameters are fit by taking models trained on trainset samples). Future model shifts are modeled by c... | 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 proposes a method to generate post-hoc recourses which remain valid under model shift, called DiRRAc. The model parameters are considered as a random vector modeled according to a mixture distribution (whose parameters are fit by taking models trained on trainset samples). Future model shifts are mode... |
The paper proposes a way to learn the dynamics between interacting particles by learning the constraints between them. It shows it can predict trajectories on synthetic and real datasets.
Strength:
- The paper is well-written and easy to follow
- Paper shows good empirical results.
Weaknesses:
- Limited novelty, the m... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
The paper proposes a way to learn the dynamics between interacting particles by learning the constraints between them. It shows it can predict trajectories on synthetic and real datasets.
Strength:
- The paper is well-written and easy to follow
- Paper shows good empirical results.
Weaknesses:
- Limited novelt... |
This paper proposes GraphAug, a new automated graph augmentation method. The method selects augmentation operation and ratio, considering the label-invariance of the operation. The experimental results show that the proposed method improves accuracy on various graph classification tasks, compared to the baseline method... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper proposes GraphAug, a new automated graph augmentation method. The method selects augmentation operation and ratio, considering the label-invariance of the operation. The experimental results show that the proposed method improves accuracy on various graph classification tasks, compared to the baselin... |
A new pruning method for deep networks is proposed. The method is motivated as computing weight saliencies so that after pruning, the spectrum of the NTK remains as close as possible to the original NTK. In practice, a series of tricks are used to make computing saliencies more efficient. At the end of the day, salienc... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
A new pruning method for deep networks is proposed. The method is motivated as computing weight saliencies so that after pruning, the spectrum of the NTK remains as close as possible to the original NTK. In practice, a series of tricks are used to make computing saliencies more efficient. At the end of the day,... |
This paper proposes ResGrad, a diffusion-based model that models the residual between the model output and the GT mel. For efficient sampling, ResGrad refines the output of the pre-trained TTS model (FastSpeech 2) by modeling the residual. ResGrad speeds up the inference for DDPM by using the pre-trained TTS model. The... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes ResGrad, a diffusion-based model that models the residual between the model output and the GT mel. For efficient sampling, ResGrad refines the output of the pre-trained TTS model (FastSpeech 2) by modeling the residual. ResGrad speeds up the inference for DDPM by using the pre-trained TTS mo... |
This paper studies the regret minimization problem of finite horizon Latent MDP with context in hindsight, which assumes the true context of the MDP will be revealed at the end of each episode. It proposes a sample efficient algorithm and proves the polynomial regret bound for the algorithm. The algorithm utilizes the ... | Recommendation: 5: marginally below the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper studies the regret minimization problem of finite horizon Latent MDP with context in hindsight, which assumes the true context of the MDP will be revealed at the end of each episode. It proposes a sample efficient algorithm and proves the polynomial regret bound for the algorithm. The algorithm utili... |
This paper considers learning a good representation for offline reinforcement learning algorithms with pixel-based visual observation space. The authors aim to extract the representation, which ignores any control-irrelevant information. The authors choose multi-step inverse models to learn the observation representati... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper considers learning a good representation for offline reinforcement learning algorithms with pixel-based visual observation space. The authors aim to extract the representation, which ignores any control-irrelevant information. The authors choose multi-step inverse models to learn the observation repr... |
Slot attention (Locatello et al., 2020) was originally designed to learn object-centric representations of images. The authors adapt slot attention to learn text representations, in the hope that those slots would similarly extract meaningful units from sequences of characters. The authors made minor changes to the ori... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
Slot attention (Locatello et al., 2020) was originally designed to learn object-centric representations of images. The authors adapt slot attention to learn text representations, in the hope that those slots would similarly extract meaningful units from sequences of characters. The authors made minor changes to... |
This paper proposes a neural network architecture, Bispectral Neural Networks(BNNs), which aims to learn the group-invariant representations of the actions of compact commutative groups. This simple architecture is composed of two layers: a single learnable linear layer, followed by a fixed collection of triple product... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper proposes a neural network architecture, Bispectral Neural Networks(BNNs), which aims to learn the group-invariant representations of the actions of compact commutative groups. This simple architecture is composed of two layers: a single learnable linear layer, followed by a fixed collection of triple... |
This paper presents a transformer based technique for embodied reference understanding. In this problem, the task is to determine the object being referred to (by the human in the image and the text) in an image. The model takes in visual and text features (from a CNN and BERT respectively) and outputs 1) the bounding ... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper presents a transformer based technique for embodied reference understanding. In this problem, the task is to determine the object being referred to (by the human in the image and the text) in an image. The model takes in visual and text features (from a CNN and BERT respectively) and outputs 1) the b... |
The paper discusses an over smoothing issue of representation in Graph Learning and message-passing like computation unit (attention) as the depth of the models increase. The paper argues that attention maps are not a very good indictors of over smoothing, due to the fact that a feature can have low similarity while ha... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper discusses an over smoothing issue of representation in Graph Learning and message-passing like computation unit (attention) as the depth of the models increase. The paper argues that attention maps are not a very good indictors of over smoothing, due to the fact that a feature can have low similarity ... |
The paper shows constructs coreset for (k,z,m)-robust clustering problem, where k is the number of clusters over the distance measure between two points are || - ||_2^z and m is the number of outlier points. The coreset size linearly depends on m, poly(k,eps^(-1)) and exponentially depends on z. The running time of the... | Recommendation: 8: accept, good paper | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The paper shows constructs coreset for (k,z,m)-robust clustering problem, where k is the number of clusters over the distance measure between two points are || - ||_2^z and m is the number of outlier points. The coreset size linearly depends on m, poly(k,eps^(-1)) and exponentially depends on z. The running tim... |
This paper considers compression of CNN weight tensors by using a kind of tensor decomposition. The purpose of this is to reduce the number of parameters and speed up the convolution computation.
It is quite clear that the authors have compressed various white spaces in the paper (e.g., around equations, in captions, ... | Recommendation: 1: strong reject | Area: Deep Learning and representational learning | Review:
This paper considers compression of CNN weight tensors by using a kind of tensor decomposition. The purpose of this is to reduce the number of parameters and speed up the convolution computation.
It is quite clear that the authors have compressed various white spaces in the paper (e.g., around equations, in ca... |
This paper tries to study the over-smoothing issue of GNNs. A detailed discussion on the potential 'pitfalls' of the existing works (focusing on preventing over-smoothing) is given. Then, a deeply-supervised GNN framework, based on the idea of layer-wise supervision, is proposed for problem-solving. Extensive experimen... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper tries to study the over-smoothing issue of GNNs. A detailed discussion on the potential 'pitfalls' of the existing works (focusing on preventing over-smoothing) is given. Then, a deeply-supervised GNN framework, based on the idea of layer-wise supervision, is proposed for problem-solving. Extensive e... |
The authors propose a method where a control policy, to simultaniously follow a nominal trajectory and to keep obstacles at a distance, is learned online as to adapt to the disturbance characteristic that is present for the given problem instance.
Algorithms for doing so are introduced and regret bounds are proven.
The... | Recommendation: 6: marginally above the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
The authors propose a method where a control policy, to simultaniously follow a nominal trajectory and to keep obstacles at a distance, is learned online as to adapt to the disturbance characteristic that is present for the given problem instance.
Algorithms for doing so are introduced and regret bounds are pro... |
The paper introduces a robust approach for dictionary learning on graphs. The method is based on a robust version of the Gromov-Wasserstein discrepancy, which involves a minimax optimization problem that is neither convex/concave. The authors proved several properties of the RGWD. The authors demonstrate the efficacy o... | Recommendation: 8: accept, good paper | Area: Unsupervised and Self-supervised learning | Review:
The paper introduces a robust approach for dictionary learning on graphs. The method is based on a robust version of the Gromov-Wasserstein discrepancy, which involves a minimax optimization problem that is neither convex/concave. The authors proved several properties of the RGWD. The authors demonstrate the ef... |
This paper extends SLAC [1] to make it applicable to constrained POMDPs. The proposed approach combines model-free RL with a latent representation that is learned by a generative model. Constraints are satisfied by learning a constraint critic and optimizing a Lagrangian, which boils down to extending the regular SAC o... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper extends SLAC [1] to make it applicable to constrained POMDPs. The proposed approach combines model-free RL with a latent representation that is learned by a generative model. Constraints are satisfied by learning a constraint critic and optimizing a Lagrangian, which boils down to extending the regul... |
This paper proposes a new method for multi-object trajectory prediction. Specifically, the proposed method consists of two different encoders and one motion decoder. The conventional motion encoder inputs past trajectories and outputs the features and hidden motion state of each agent. And the motion decoder receives ... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a new method for multi-object trajectory prediction. Specifically, the proposed method consists of two different encoders and one motion decoder. The conventional motion encoder inputs past trajectories and outputs the features and hidden motion state of each agent. And the motion decoder r... |
This work proposes SIMOL, a soft improvement based multi-objective optimization algorithm, for efficient meta learning. The proposed method first formulates meta learning as a multi-objective optimization problem, where each task is an objective to optimize. Then, to handle the huge number of tasks (e.g., can be up to ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This work proposes SIMOL, a soft improvement based multi-objective optimization algorithm, for efficient meta learning. The proposed method first formulates meta learning as a multi-objective optimization problem, where each task is an objective to optimize. Then, to handle the huge number of tasks (e.g., can b... |
This manuscript tackles the challenging problem of graph learning in 3D. The authors proposed the first use of quantum computing to learn quantum 3D latent representation of 3D structures from the Hilbert space composed of the Bloch sphere of each qubit. The proposed method preserves equivariance and invariance of 3D g... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This manuscript tackles the challenging problem of graph learning in 3D. The authors proposed the first use of quantum computing to learn quantum 3D latent representation of 3D structures from the Hilbert space composed of the Bloch sphere of each qubit. The proposed method preserves equivariance and invariance... |
The paper tries to tackle the problem of learning a generative model for sampling identically distributed but dependent data points from a distribution. Like previous normalizing flows, they try to learn a bijective map of data to an assumed latent space but unlike previous works which assume a diagonal Gaussian, they ... | Recommendation: 6: marginally above the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The paper tries to tackle the problem of learning a generative model for sampling identically distributed but dependent data points from a distribution. Like previous normalizing flows, they try to learn a bijective map of data to an assumed latent space but unlike previous works which assume a diagonal Gaussia... |
This paper proposes a new MoE model Architecture to improve the parameter efficiency of MoE by learning a soft combination of a global set of expert layers.
Weaknesses:
(1). This paper carried out analysis first and listed three challenges from analysis. However, I did not know which MoE model does this paper stud... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper proposes a new MoE model Architecture to improve the parameter efficiency of MoE by learning a soft combination of a global set of expert layers.
Weaknesses:
(1). This paper carried out analysis first and listed three challenges from analysis. However, I did not know which MoE model does this pa... |
This paper is about unsupervised learning of image classification. It is focused on patch-level contrastive by an attention mechanism. Experiments on ImageNet-1k, MSCOCO, ADE-20k verify the effectiveness.
--Strength
The paper introduces ADCLR, a novel patch-level contrastive learning method, improves the quality of d... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper is about unsupervised learning of image classification. It is focused on patch-level contrastive by an attention mechanism. Experiments on ImageNet-1k, MSCOCO, ADE-20k verify the effectiveness.
--Strength
The paper introduces ADCLR, a novel patch-level contrastive learning method, improves the qual... |
The paper proposes a method for point cloud triangulation by creating local per-point triangulations and then post-processing these local triangulations into a global point cloud triangulation. The first step is performed using a graph neural network and it is the main contribution of the paper. Specifically, the paper... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper proposes a method for point cloud triangulation by creating local per-point triangulations and then post-processing these local triangulations into a global point cloud triangulation. The first step is performed using a graph neural network and it is the main contribution of the paper. Specifically, t... |
This paper presents a new decentralized learning algorithm that bases on the proposed cross gradient aggregation (CGA) algorithm. Specifically, the authors leverage the existing self-gradient and cross-gradient concepts to develop the neighborhood gradient clustering (NGC) algorithm. This proposed method replaces the l... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper presents a new decentralized learning algorithm that bases on the proposed cross gradient aggregation (CGA) algorithm. Specifically, the authors leverage the existing self-gradient and cross-gradient concepts to develop the neighborhood gradient clustering (NGC) algorithm. This proposed method replac... |
The paper assumes that each patient belongs to different domains given by some unobserved covariates. Mutual reconstruction is used to learn the domain covariates, which are then removed by orthogonal projection. The proposed methods improve prediction performance on healthcare monitoring and EHR datasets.
Strength:
... | Recommendation: 3: reject, not good enough | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper assumes that each patient belongs to different domains given by some unobserved covariates. Mutual reconstruction is used to learn the domain covariates, which are then removed by orthogonal projection. The proposed methods improve prediction performance on healthcare monitoring and EHR datasets.
St... |
The paper proposes an unbiased stochastic quantization method called QUIC-FL for rotation-based mean estimation, in order to reduce the communication by transmitting full-precision vectors. The method uses a 'shared randomness' approach which reduces the computation cost at the server to recover the compressed signals.... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
The paper proposes an unbiased stochastic quantization method called QUIC-FL for rotation-based mean estimation, in order to reduce the communication by transmitting full-precision vectors. The method uses a 'shared randomness' approach which reduces the computation cost at the server to recover the compressed ... |
This paper presents that we can significantly improve the sample-efficient of prior deep RL approaches by increasing the number of updates per environment steps, and shows that resetting all parameters or part of parameters is critical. The paper shows improved performance on a variety of benchmark tasks, and provides ... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper presents that we can significantly improve the sample-efficient of prior deep RL approaches by increasing the number of updates per environment steps, and shows that resetting all parameters or part of parameters is critical. The paper shows improved performance on a variety of benchmark tasks, and p... |
This paper designed new regularization terms of coupled cross entropy minimization criterion in crowdsourcing tasks. It relaxed the assumption of conditional independence among the annotators.
Strength: This paper relaxed the assumption conditional independence among the annotators in crowdsourcing tasks with coupled c... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper designed new regularization terms of coupled cross entropy minimization criterion in crowdsourcing tasks. It relaxed the assumption of conditional independence among the annotators.
Strength: This paper relaxed the assumption conditional independence among the annotators in crowdsourcing tasks with c... |
This work proposes a new domain adaptation method for both single-source and multi-source conditions. The main idea is to alleviate the error accumulation of target pseudo labels and reduce the domain gap at the prediction level. Technically, they use a generative model (normalizing flows) to model the class-wise distr... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This work proposes a new domain adaptation method for both single-source and multi-source conditions. The main idea is to alleviate the error accumulation of target pseudo labels and reduce the domain gap at the prediction level. Technically, they use a generative model (normalizing flows) to model the class-wi... |
This paper empirically measures correlations between various generalization metrics and model quality (BLEU score) on a machine translation dataset across a range of hyperparameters. Compare to prior work, the main differences are:
- Focusing on NLP, rather than computer vision
- Focusing on metrics that predict test e... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper empirically measures correlations between various generalization metrics and model quality (BLEU score) on a machine translation dataset across a range of hyperparameters. Compare to prior work, the main differences are:
- Focusing on NLP, rather than computer vision
- Focusing on metrics that predic... |
The paper proposes a test time augmentation using a cascade loss prediction method which only requires a single forward pass of the transformation predictor to select multiple transformations. In contrast to the repeated usage of one loss predictor in cyclic-TTA, the proposed method uses RNN to capture the semantic inf... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper proposes a test time augmentation using a cascade loss prediction method which only requires a single forward pass of the transformation predictor to select multiple transformations. In contrast to the repeated usage of one loss predictor in cyclic-TTA, the proposed method uses RNN to capture the sema... |
The paper proposes to group states in continuous space into clusters, according to a learned reachability metric. Such an operation allows turning the independent transition trajectories into a graph, allowing running value iteration for robust credit assignment. With shortest-path planning and an action translator, th... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper proposes to group states in continuous space into clusters, according to a learned reachability metric. Such an operation allows turning the independent transition trajectories into a graph, allowing running value iteration for robust credit assignment. With shortest-path planning and an action transl... |
This paper analyze the soft sampling method for training DNNs, which selects a subset uniformly at random with replacement from the full data set in each epoch. Analysis of convergence rate and coverage are conducted.
Strength:
1. The theoretical analysis of convergence and coverage of the sampling method could be a us... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper analyze the soft sampling method for training DNNs, which selects a subset uniformly at random with replacement from the full data set in each epoch. Analysis of convergence rate and coverage are conducted.
Strength:
1. The theoretical analysis of convergence and coverage of the sampling method could... |
Recent work shows that large batch size can give better privacy-utility trade-off in differentia private ML training. However, training with large batch sizes is computationally expensive. The paper observes (empirically) that the privacy budget of differential private ML training depends mostly on the total amount of ... | Recommendation: 6: marginally above the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
Recent work shows that large batch size can give better privacy-utility trade-off in differentia private ML training. However, training with large batch sizes is computationally expensive. The paper observes (empirically) that the privacy budget of differential private ML training depends mostly on the total am... |
This paper studies the benefits of model-based generalization by comparing the performance of DQN on the rollouts of a learned model, and DQN on a replay buffer. Theoretically, this paper proves that learning a transition model can be more efficient (in terms of pruning invalid solutions) than learning a Q-function usi... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper studies the benefits of model-based generalization by comparing the performance of DQN on the rollouts of a learned model, and DQN on a replay buffer. Theoretically, this paper proves that learning a transition model can be more efficient (in terms of pruning invalid solutions) than learning a Q-func... |
The authors consider the task of classifying data with long-tailed distributions of class samples under resource constraints. The resource constraints are addressed by using binary neural networks. The authors propose to use floating point networks pretrained on standard benchmarks and to retrain the classifier layer o... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The authors consider the task of classifying data with long-tailed distributions of class samples under resource constraints. The resource constraints are addressed by using binary neural networks. The authors propose to use floating point networks pretrained on standard benchmarks and to retrain the classifier... |
This paper studies the convergence of the temporal-difference (TD) algorithm with overparametrized neural networks. Compared with previous works, it weakens the projection step radius to a constant level by using a different analyzing techniques. The paper also provides simulation results to support the theories.
Stren... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper studies the convergence of the temporal-difference (TD) algorithm with overparametrized neural networks. Compared with previous works, it weakens the projection step radius to a constant level by using a different analyzing techniques. The paper also provides simulation results to support the theorie... |
This paper aims to scale up model capacity and improve the generalization performance across tasks with offline Q-learning methods. In contrast, prior works mainly centered around small-scale, single-task problems where broad generalization and learning general-purpose representations are not expected. To this end, the... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper aims to scale up model capacity and improve the generalization performance across tasks with offline Q-learning methods. In contrast, prior works mainly centered around small-scale, single-task problems where broad generalization and learning general-purpose representations are not expected. To this ... |
The authors propose a new knowledge representation method named StarGraph, which serves as an encoder and aims to generate high-quality entity embeddings for downstream score functions. The proposed StarGraph mainly consists of two stages: subgraph sampling and encoding, and the authors manage to optimize the efficienc... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The authors propose a new knowledge representation method named StarGraph, which serves as an encoder and aims to generate high-quality entity embeddings for downstream score functions. The proposed StarGraph mainly consists of two stages: subgraph sampling and encoding, and the authors manage to optimize the e... |
This paper proposes momentum tracking in decentralized learning. The authors establish a main theorem that shows the proposed momentum tracking method is invariant to the data heterogeneity bound. The authors conduct several experiments comparing momentum tracking with other baseline methods including gradient tracking... | Recommendation: 5: marginally below the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper proposes momentum tracking in decentralized learning. The authors establish a main theorem that shows the proposed momentum tracking method is invariant to the data heterogeneity bound. The authors conduct several experiments comparing momentum tracking with other baseline methods including gradient ... |
This paper presents an architecture which divides entities into two groups, one of locally relevant entities passed directly to the agent's utility network and another of less relevant entities which are passed into a global coach which compresses global information. This architecture produces improved performance on m... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper presents an architecture which divides entities into two groups, one of locally relevant entities passed directly to the agent's utility network and another of less relevant entities which are passed into a global coach which compresses global information. This architecture produces improved performa... |
This paper proposes a vector representation of a 2-parameter persistent homology. This is a very important problem. Application of multiparameter persistence can have significance impact in developing topology-based learning methods.
The proposed method is based on the generalized rank invariance by Kim and Memoli an... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a vector representation of a 2-parameter persistent homology. This is a very important problem. Application of multiparameter persistence can have significance impact in developing topology-based learning methods.
The proposed method is based on the generalized rank invariance by Kim and M... |
This paper focuses on style transfer. Different from previous works, this work shows the equivalent form of UST methods in the frequency domain. Moreover, the proposed method revisits existing UST methods in the frequency domain, showing the effects of Fourier amplitude and phase. With these experiments, this work prop... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper focuses on style transfer. Different from previous works, this work shows the equivalent form of UST methods in the frequency domain. Moreover, the proposed method revisits existing UST methods in the frequency domain, showing the effects of Fourier amplitude and phase. With these experiments, this w... |
This paper proposes new techniques for using public data in differentially private machine learning and makes a significant improvement over the state-of-the-art solution. More specifically, they propose to use synthesized data provided by a generative model trained by the given public data, and a new gradient clipping... | Recommendation: 6: marginally above the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper proposes new techniques for using public data in differentially private machine learning and makes a significant improvement over the state-of-the-art solution. More specifically, they propose to use synthesized data provided by a generative model trained by the given public data, and a new gradient ... |
The paper focuses on providing a learning strategy tailored to tackle the problem of Raven’s Progressive Matrices (RPMs). The main idea consists of (i) distinguishing between attribute-level (determining the shape, color, size and presence of an object) and task-level (associating a symbol to each combination of object... | Recommendation: 3: reject, not good enough | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
The paper focuses on providing a learning strategy tailored to tackle the problem of Raven’s Progressive Matrices (RPMs). The main idea consists of (i) distinguishing between attribute-level (determining the shape, color, size and presence of an object) and task-level (associating a symbol to each combination o... |
The paper proposes the use of large-scale autoregressive generative language models (LLMs) for generating tabular data. Several recent works have similarly attempted the task of generating tabular data using popular computer vision methods such as Variational Autoencoders and Generative Adversarial Networks, but requir... | Recommendation: 8: accept, good paper | Area: Generative models | Review:
The paper proposes the use of large-scale autoregressive generative language models (LLMs) for generating tabular data. Several recent works have similarly attempted the task of generating tabular data using popular computer vision methods such as Variational Autoencoders and Generative Adversarial Networks, bu... |
The paper proposes a domain generalization method with a new discriminative energy-base model. The proposed idea is novel and effective because they adapt the target samples to source distributions instead of generalizing the model to unseen target samples in most previous methods. Theoretically, the authors provide de... | Recommendation: 8: accept, good paper | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The paper proposes a domain generalization method with a new discriminative energy-base model. The proposed idea is novel and effective because they adapt the target samples to source distributions instead of generalizing the model to unseen target samples in most previous methods. Theoretically, the authors pr... |
This paper proposes a new method to only use text to achieve local image editing based on recent text-to-image models. More specifically, they propose to use cross-attention to connect each word and image regions and get rid of user-provided masks. Many interesting applications have been proposed and the results are ve... | Recommendation: 8: accept, good paper | Area: Generative models | Review:
This paper proposes a new method to only use text to achieve local image editing based on recent text-to-image models. More specifically, they propose to use cross-attention to connect each word and image regions and get rid of user-provided masks. Many interesting applications have been proposed and the result... |
The paper studies representations obtained due to using auxiliary tasks and larger networks. The main claim is that tasks derived from randomly initialized networks are sufficient to train strong representations.
The major strength of the work is the fact that it addresses an important problem of augmenting the RL tra... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper studies representations obtained due to using auxiliary tasks and larger networks. The main claim is that tasks derived from randomly initialized networks are sufficient to train strong representations.
The major strength of the work is the fact that it addresses an important problem of augmenting th... |
The paper presents an extension of Fourier neural operators (FNO) to have operators evolving over time. The paper proposed an architecture to cooperate the time embedding into FNO. The experiments are conducted on synthetic datasets given PDEs and on real-world time series data. The model outperforms alternative method... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper presents an extension of Fourier neural operators (FNO) to have operators evolving over time. The paper proposed an architecture to cooperate the time embedding into FNO. The experiments are conducted on synthetic datasets given PDEs and on real-world time series data. The model outperforms alternativ... |
In this paper, the authors consider the problem of computing the 1-Wasserstein distance (Earth Mover’s distance) between two d-dimensional discrete distributions $\nu$ and $\mu$. In this problem the mass distributed over a set of supply nodes according to distribution $\nu$, needs to be transported and distributed to t... | Recommendation: 8: accept, good paper | Area: Optimization (eg, convex and non-convex optimization) | Review:
In this paper, the authors consider the problem of computing the 1-Wasserstein distance (Earth Mover’s distance) between two d-dimensional discrete distributions $\nu$ and $\mu$. In this problem the mass distributed over a set of supply nodes according to distribution $\nu$, needs to be transported and distribu... |
This paper studies the problem of how to improve language models’ ability to perform reasoning. In particular, the authors focus on the task of mathematical reasoning: on both GSM8K and a synthetic task called unit conversion, the authors showed that by training models to generate abstract representations (referred to ... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper studies the problem of how to improve language models’ ability to perform reasoning. In particular, the authors focus on the task of mathematical reasoning: on both GSM8K and a synthetic task called unit conversion, the authors showed that by training models to generate abstract representations (refe... |
The paper proposes a deep extension of variational implicit processes (VIP, Ma et al. (2019)), in which an implicit process, defined as a (nonlinear) transformation of a collection of random variables, is approximated using a Gaussian process. Together with a Gaussian likelihood, this gives a Gaussian process posterior... | Recommendation: 6: marginally above the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The paper proposes a deep extension of variational implicit processes (VIP, Ma et al. (2019)), in which an implicit process, defined as a (nonlinear) transformation of a collection of random variables, is approximated using a Gaussian process. Together with a Gaussian likelihood, this gives a Gaussian process p... |
The authors identify an implicit characteristic of convolutional layer operations and exploit that to show benefits in operational speed/efficiency for CNN transposed CNN layers and pooling layers in deep nets. The idea itself is neat and clear. And the paper is presented well. The authors evaluate the proposed on non-... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The authors identify an implicit characteristic of convolutional layer operations and exploit that to show benefits in operational speed/efficiency for CNN transposed CNN layers and pooling layers in deep nets. The idea itself is neat and clear. And the paper is presented well. The authors evaluate the proposed... |
The paper tackles the problem of learning representations that are robust to biases in the data. The authors first clarify why existing contrastive losses fail to deal with biased data, and further derive a novel formulation of supervised contrastive loss to provide more accurate control of minimal distance between pos... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper tackles the problem of learning representations that are robust to biases in the data. The authors first clarify why existing contrastive losses fail to deal with biased data, and further derive a novel formulation of supervised contrastive loss to provide more accurate control of minimal distance bet... |
Summary:
- The paper proposes a new task called Visual Transformation Telling (VTT), which is defined as predicting the intermediate steps ('steps' are referred to as ‘transformations’ in the paper) in forms of natural language given the start and end state of each step.
- Two existing instructional video understandin... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
Summary:
- The paper proposes a new task called Visual Transformation Telling (VTT), which is defined as predicting the intermediate steps ('steps' are referred to as ‘transformations’ in the paper) in forms of natural language given the start and end state of each step.
- Two existing instructional video unde... |
* This paper proposes an approach to few-shot classification developing on related works such as MAML and Body Only Inner Loop (BOIL).
* In few-shot learning settings which involve domain shift between the training and evaluation settings,
models necessarily require some adaptation in their parameters in order to be a... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
* This paper proposes an approach to few-shot classification developing on related works such as MAML and Body Only Inner Loop (BOIL).
* In few-shot learning settings which involve domain shift between the training and evaluation settings,
models necessarily require some adaptation in their parameters in order... |
Summary:
The authors propose a variational prompt generator conditioned on an input instance. This is learned by adding a residual (which is input conditioned) to a fixed set of learnt prompts. During inference, multiple residuals can be sampled to generate different prompts. The authors show that the proposed approac... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
Summary:
The authors propose a variational prompt generator conditioned on an input instance. This is learned by adding a residual (which is input conditioned) to a fixed set of learnt prompts. During inference, multiple residuals can be sampled to generate different prompts. The authors show that the proposed... |
In crowd-computing tasks there are often two causes for wrong answers: 1) low reliability of worker (i.e. possibly a spammer or someone making a random guess), 2) confusion due to task (or question) difficulty (i.e. the question is simply harder than others or the potential answers could be confusing). The David-Skene ... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
In crowd-computing tasks there are often two causes for wrong answers: 1) low reliability of worker (i.e. possibly a spammer or someone making a random guess), 2) confusion due to task (or question) difficulty (i.e. the question is simply harder than others or the potential answers could be confusing). The Davi... |
There is a recent emergence of interests in adapting diffusion models to handle discrete data and this work adds to that literature. To understand the main difference from prior work, it is important to note that there are two (almost) equivalent frameworks for learning diffusion models: the variational lower bound fra... | Recommendation: 5: marginally below the acceptance threshold | Area: Generative models | Review:
There is a recent emergence of interests in adapting diffusion models to handle discrete data and this work adds to that literature. To understand the main difference from prior work, it is important to note that there are two (almost) equivalent frameworks for learning diffusion models: the variational lower b... |
In this paper, the authors point out two limitations in using external knowledge with pretrained language models (PLMs): (1) it is time-costly to index and retrieve on large-scale knowledge bases, and (2) retrieved knowledge could be noisy and misleading. With these limitations as the motivation, they propose a scoring... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
In this paper, the authors point out two limitations in using external knowledge with pretrained language models (PLMs): (1) it is time-costly to index and retrieve on large-scale knowledge bases, and (2) retrieved knowledge could be noisy and misleading. With these limitations as the motivation, they propose a... |
The authors propose a new method to train spiking neural networks (SNNs) directly. The most efficient current method to do so is surrogate gradient (SG) learning where for the spike threshold the Heaviside step function is used in the forward pass, but a surrogate gradient (usually the derivative of a sigmoid function)... | Recommendation: 8: accept, good paper | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
The authors propose a new method to train spiking neural networks (SNNs) directly. The most efficient current method to do so is surrogate gradient (SG) learning where for the spike threshold the Heaviside step function is used in the forward pass, but a surrogate gradient (usually the derivative of a sigmoid f... |
The focus of this paper is on communication-efficient and Byzantine-robust distributed optimization. To this end, the authors propose an algorithm Byz-VR-MARINA that combines ideas from the Byzantine-robust distributed optimization work of Karimireddy et al., 2021 and communication-efficient distributed optimization wo... | Recommendation: 6: marginally above the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
The focus of this paper is on communication-efficient and Byzantine-robust distributed optimization. To this end, the authors propose an algorithm Byz-VR-MARINA that combines ideas from the Byzantine-robust distributed optimization work of Karimireddy et al., 2021 and communication-efficient distributed optimiz... |
This paper proposes Symbolic Physics Learner (SPL), a method for symbolic regression and discovery of dynamics.
The authors formalize mathematical operations using expression trees and context-free grammars,
and employ Monte Carlo Tree Search (MCTS) to explore the space and select
optimal solutions.
Furthermore, they p... | Recommendation: 8: accept, good paper | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper proposes Symbolic Physics Learner (SPL), a method for symbolic regression and discovery of dynamics.
The authors formalize mathematical operations using expression trees and context-free grammars,
and employ Monte Carlo Tree Search (MCTS) to explore the space and select
optimal solutions.
Furthermore... |
This paper provides a novel aspect of using quantum ML for geometric data modeling. This is the first work in this research line, especially for quantum chemistry.
Strengths:
- This paper introduces using the quantum computing for geometric data modeling. This is a promising direction.
- The paper is self-contained, co... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper provides a novel aspect of using quantum ML for geometric data modeling. This is the first work in this research line, especially for quantum chemistry.
Strengths:
- This paper introduces using the quantum computing for geometric data modeling. This is a promising direction.
- The paper is self-conta... |
In this paper, the authors propose a locally dense, globally sparse alternative to the feedforward neural network architecture. The proposed architecture consists of interconnected blocks where the neurons within each block are connected via feedforward, attention, and product-between-neurons mechanisms. The authors co... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
In this paper, the authors propose a locally dense, globally sparse alternative to the feedforward neural network architecture. The proposed architecture consists of interconnected blocks where the neurons within each block are connected via feedforward, attention, and product-between-neurons mechanisms. The au... |
The paper proposes an algorithm to learn low-dimensional representations in the Federated Learning setting. The paper borrows idea from a recent work which formulated a loss that could maximize the code rate difference between the entire dataset and summation of individual classes. The work takes a further step by iden... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes an algorithm to learn low-dimensional representations in the Federated Learning setting. The paper borrows idea from a recent work which formulated a loss that could maximize the code rate difference between the entire dataset and summation of individual classes. The work takes a further step... |
This paper investigates whether using multiple frames for training is necessary for video-language downstream tasks. The authors propose a single-frame framework for video-language understanding. Results indicate that with large-scale pre-training and a proper frame ensemble strategy at inference time, a single-frame t... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper investigates whether using multiple frames for training is necessary for video-language downstream tasks. The authors propose a single-frame framework for video-language understanding. Results indicate that with large-scale pre-training and a proper frame ensemble strategy at inference time, a single... |
This paper proposes a method for neural network model fusion based on the concepts of Wasserstein barycenter, and Gromov-Wasserstein barycenter. The method uses OT couplings from each previous layer to construct couplings between the subsequent layer that minimize a cost based off the previous layers' couplings. Signif... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a method for neural network model fusion based on the concepts of Wasserstein barycenter, and Gromov-Wasserstein barycenter. The method uses OT couplings from each previous layer to construct couplings between the subsequent layer that minimize a cost based off the previous layers' couplings... |
The paper presents an auto-regressive style model for joint causal graph learning and missing data imputation for unstructured temporal data. The model is trained in a two-step procedure alternating between training the imputation model and the graph learning (EM-style algorithm). The model parametrises the graph using... | Recommendation: 6: marginally above the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The paper presents an auto-regressive style model for joint causal graph learning and missing data imputation for unstructured temporal data. The model is trained in a two-step procedure alternating between training the imputation model and the graph learning (EM-style algorithm). The model parametrises the gra... |
This paper proposes a novel knowledge distillation methods which decide the progressive distillation process by the cost of linear regression on validation set, which achieves good performance on multiple student-teacher settings.
Strength:
1. This paper proposes to choose the policy of student-teacher in knowledge di... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper proposes a novel knowledge distillation methods which decide the progressive distillation process by the cost of linear regression on validation set, which achieves good performance on multiple student-teacher settings.
Strength:
1. This paper proposes to choose the policy of student-teacher in know... |
This paper deals with training neural network to be robust to geometric transformations of the inputs. Here, robustness is taken to mean *certified* robustness, which can be formally proven, even in the worst case scenario.
The contribution are two-fold:
- Improvements in terms of speed of evaluating bounds on the out... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper deals with training neural network to be robust to geometric transformations of the inputs. Here, robustness is taken to mean *certified* robustness, which can be formally proven, even in the worst case scenario.
The contribution are two-fold:
- Improvements in terms of speed of evaluating bounds on... |
Linear programs (LPs) are a special case of constraint satisfaction problems (CSPs).
This work claims that GNN's may approximate the properties of LPs: if it is feasible, unbounded and its optimal solution if bounded.
namely: feasibility, optimal objective value, and optimal solution.
The work shows (in theorems 4.1 an... | Recommendation: 5: marginally below the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
Linear programs (LPs) are a special case of constraint satisfaction problems (CSPs).
This work claims that GNN's may approximate the properties of LPs: if it is feasible, unbounded and its optimal solution if bounded.
namely: feasibility, optimal objective value, and optimal solution.
The work shows (in theorem... |
This paper studies the single domain generalization problem for 3D point cloud classification, which aims to generalize the model trained on a single source domain to an unknown target domain. The method splits the training domain into two subdomains, and models the alignment between the two sub-domains, expecting the ... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper studies the single domain generalization problem for 3D point cloud classification, which aims to generalize the model trained on a single source domain to an unknown target domain. The method splits the training domain into two subdomains, and models the alignment between the two sub-domains, expect... |
This paper presents Pix2Struct, a pretrained image-to-text model. The model is pretrained to parse masked screenshots of web pages and the authors show that it can be fine-tuned on multiple tasks containing visually-situated language (VQA, Image captioning, Infographic VQA). The fine-tuned models achieve state-of-the-a... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper presents Pix2Struct, a pretrained image-to-text model. The model is pretrained to parse masked screenshots of web pages and the authors show that it can be fine-tuned on multiple tasks containing visually-situated language (VQA, Image captioning, Infographic VQA). The fine-tuned models achieve state-... |
A meta-learning framework, EDGE, for new drug recommendation is proposed using relationships between drugs and diseases.
Strengths
- Paper is very well written
- Uses two real-world datasets MMIC-IV and CLAIMS
- The Edge framework is compared against several well known techniques from literature such as GameNet, SafeDr... | Recommendation: 3: reject, not good enough | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
A meta-learning framework, EDGE, for new drug recommendation is proposed using relationships between drugs and diseases.
Strengths
- Paper is very well written
- Uses two real-world datasets MMIC-IV and CLAIMS
- The Edge framework is compared against several well known techniques from literature such as GameNet... |
The paper tackles the question of when offline policy selection (OPS, selecting a well-performing policy in offline RL) can be performed efficiently. For this they determine bounds based on bounds by offline policy evaluation (OPE): OPS is lower-bounded by the number of samples/episodes required for OPE. The paper expl... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper tackles the question of when offline policy selection (OPS, selecting a well-performing policy in offline RL) can be performed efficiently. For this they determine bounds based on bounds by offline policy evaluation (OPE): OPS is lower-bounded by the number of samples/episodes required for OPE. The pa... |
This paper presents a method, Magnetic Mirror Decent, which can get comparable performance to state of the art in both single-agent and multi-agent RL settings. To the best of my knowledge it is the first algorithm known to have such strong empirical performance in both settings, and to have strong theoretical converg... | Recommendation: 8: accept, good paper | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper presents a method, Magnetic Mirror Decent, which can get comparable performance to state of the art in both single-agent and multi-agent RL settings. To the best of my knowledge it is the first algorithm known to have such strong empirical performance in both settings, and to have strong theoretical... |
The paper studies the problem of editing factual knowledge in transformer-based language models. Authors note that existing model editing techniques fail when introducing a large batch of edits. They present a fairly model-specific algorithm called MEMIT, directly modifies model weights of FFN (MLP) transformer layers ... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper studies the problem of editing factual knowledge in transformer-based language models. Authors note that existing model editing techniques fail when introducing a large batch of edits. They present a fairly model-specific algorithm called MEMIT, directly modifies model weights of FFN (MLP) transformer... |
The paper studies the problem of spurious correlations. Authors propose to improve robustness against spurious correlations through a simple approach (Deep Feature Reweighting): specifically re-train the last layer of a neural network using a small set of reweighting
data where the spurious correlation does not hold. F... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper studies the problem of spurious correlations. Authors propose to improve robustness against spurious correlations through a simple approach (Deep Feature Reweighting): specifically re-train the last layer of a neural network using a small set of reweighting
data where the spurious correlation does not... |
The paper proposes a new objective for model-based RL that lower bounds the true objective and can be used to optimize the encoder, model, and the policy at once. The proposed model shows higher sample efficiency compared to the baselines.
Strengths
- Recent works (especially on offline RL) have shown that we can find ... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper proposes a new objective for model-based RL that lower bounds the true objective and can be used to optimize the encoder, model, and the policy at once. The proposed model shows higher sample efficiency compared to the baselines.
Strengths
- Recent works (especially on offline RL) have shown that we c... |
This paper aims to address unsupervised domain generalization via a proposed cycle-consistent masked autoencoder, which includes a CycleGAN, MaskAutoEncoder, and contrastive learning. The authors claim that the poposed method can significantly improve state-of-the-art performances, which are demonstrated by the experim... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper aims to address unsupervised domain generalization via a proposed cycle-consistent masked autoencoder, which includes a CycleGAN, MaskAutoEncoder, and contrastive learning. The authors claim that the poposed method can significantly improve state-of-the-art performances, which are demonstrated by the... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.