review stringlengths 5 16.6k | score stringclasses 5
values | area stringclasses 12
values | text stringlengths 31 5.65k |
|---|---|---|---|
The paper introduces a distributed extension of GNNAutoScale (DIGEST) and proposes an asynchronous representation update mechanism (DIGEST-A) to reduce communication overhead for node embedding updates. Theoretical analyses such as forward error bound and convergence analysis are provided. Experiments show promising sp... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper introduces a distributed extension of GNNAutoScale (DIGEST) and proposes an asynchronous representation update mechanism (DIGEST-A) to reduce communication overhead for node embedding updates. Theoretical analyses such as forward error bound and convergence analysis are provided. Experiments show prom... |
The authors propose a self-supervised approach to segment objects from novel views synthesized from a NeRF.
The idea is borrowed from Hamilton et al. (2022) who learned semantic co-segmentation of images in an unsupervised fashion, by distilling the knowledge of DINO. The authors adapted the method to NeRF, to segment... | Recommendation: 8: accept, good paper | Area: Unsupervised and Self-supervised learning | Review:
The authors propose a self-supervised approach to segment objects from novel views synthesized from a NeRF.
The idea is borrowed from Hamilton et al. (2022) who learned semantic co-segmentation of images in an unsupervised fashion, by distilling the knowledge of DINO. The authors adapted the method to NeRF, to... |
The paper proposes a new approach based on Transformer, namely Optformer, to optimize unconstrained black-box optimization problems. The key idea is based on an existing observation that Vision Transformer is similar to the evolutionary algorithm (EAs). Therefore, the paper aims to modify the components of the Transfor... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes a new approach based on Transformer, namely Optformer, to optimize unconstrained black-box optimization problems. The key idea is based on an existing observation that Vision Transformer is similar to the evolutionary algorithm (EAs). Therefore, the paper aims to modify the components of the ... |
The paper presents a learnable aggregation scheme in the context of federated learning. The paper achieves this using meta-learning to generalize the parameters of the aggregator with a proxy dataset. The paper identifies 'period drift' in the current federated learning setup and presents the meta-learning-based aggreg... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The paper presents a learnable aggregation scheme in the context of federated learning. The paper achieves this using meta-learning to generalize the parameters of the aggregator with a proxy dataset. The paper identifies 'period drift' in the current federated learning setup and presents the meta-learning-base... |
This paper presents to perform instance-wise batch label restoration from only the gradient of the final layer, to extract labels from gradients. The core idea is to establish linear equations of the gradients, probabilities and labels to derive the Number of Instances (NoI) per class by the Moore-Penrose pseudoin-vers... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper presents to perform instance-wise batch label restoration from only the gradient of the final layer, to extract labels from gradients. The core idea is to establish linear equations of the gradients, probabilities and labels to derive the Number of Instances (NoI) per class by the Moore-Penrose pseud... |
This paper discusses and proposed one ensemble method of GNN. The key idea is based on the high-order relationship of a target node. The proposed idea is relatively easy to understand and implement, which indicates that future researchers can easily generalize it. The proposed methods also outperform baselines.
The hig... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper discusses and proposed one ensemble method of GNN. The key idea is based on the high-order relationship of a target node. The proposed idea is relatively easy to understand and implement, which indicates that future researchers can easily generalize it. The proposed methods also outperform baselines.... |
The paper proposes an Actor-Critic algorithm in which the actor learns to pursue (imitate) the greedy policy of the critic, using, with continuous actions, a method inspired from the Cross-Entropy method (pursuing the top-K actions from a sample of relatively good actions). With continuous actions, the core contributio... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper proposes an Actor-Critic algorithm in which the actor learns to pursue (imitate) the greedy policy of the critic, using, with continuous actions, a method inspired from the Cross-Entropy method (pursuing the top-K actions from a sample of relatively good actions). With continuous actions, the core con... |
In the paper under review, it is shown that the TD for validation data (validation TD) correlates with the performance of the RL + regularization methods in high UTD (replay ratio) settings.
Based on this finding, a method that selects the regularization method with the smallest validation TD from multiple regularizat... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
In the paper under review, it is shown that the TD for validation data (validation TD) correlates with the performance of the RL + regularization methods in high UTD (replay ratio) settings.
Based on this finding, a method that selects the regularization method with the smallest validation TD from multiple reg... |
This paper first studies the relationship of logit margin loss $\ell_{LM}$ and adversarial robustness and then proposes using one-versus-the-rest (OVR) loss to improve the logit margin loss.
Strength
- The paper is well-written.
- The behavior of the logit margin loss and how OVR loss helps to improve logit margin loss... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper first studies the relationship of logit margin loss $\ell_{LM}$ and adversarial robustness and then proposes using one-versus-the-rest (OVR) loss to improve the logit margin loss.
Strength
- The paper is well-written.
- The behavior of the logit margin loss and how OVR loss helps to improve logit mar... |
This paper is focused on the few-shot anomaly detection problem. It presents a method that combines contrastive learning for adapting a pre-training model to the target domain, a cross-instance positive pair loss, an option to incorporate negative pair loss, and density-based anomaly detection. They show some promising... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper is focused on the few-shot anomaly detection problem. It presents a method that combines contrastive learning for adapting a pre-training model to the target domain, a cross-instance positive pair loss, an option to incorporate negative pair loss, and density-based anomaly detection. They show some p... |
This paper proposes a new Graph Transformer architecture, Graph Diffuser, to incorporate structural information in graphs, particularly the long-range interactions. Specifically, Graph Diffuser first generates a dense adjacency matrix from the node and edge features, then obtains the virtual edge features by feeding th... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper proposes a new Graph Transformer architecture, Graph Diffuser, to incorporate structural information in graphs, particularly the long-range interactions. Specifically, Graph Diffuser first generates a dense adjacency matrix from the node and edge features, then obtains the virtual edge features by fe... |
This paper investigates the problem of learning a NE in two-players zero-sum games via no-regret algorithms. While most of the state-of-the-art algorithms guarantee convergence in terms of average strategies, existent methods that guarantee last-iterate convergence require strong assumptions like the uniqueness of the ... | Recommendation: 5: marginally below the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper investigates the problem of learning a NE in two-players zero-sum games via no-regret algorithms. While most of the state-of-the-art algorithms guarantee convergence in terms of average strategies, existent methods that guarantee last-iterate convergence require strong assumptions like the uniqueness... |
This paper proposes ThinkSum: a two-stage paradigm for performing inference with large language models (LLMs) with no gradient update. The authors distinguish ThinkSum from chain-of-though like prompting methods by claiming that ThinkSum performs probabilistic inference instead of using LLMs to directly generate answer... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes ThinkSum: a two-stage paradigm for performing inference with large language models (LLMs) with no gradient update. The authors distinguish ThinkSum from chain-of-though like prompting methods by claiming that ThinkSum performs probabilistic inference instead of using LLMs to directly generat... |
This paper proposes a new graph contrastive paradigm named CLEP, which learns the latent community structure via variational inference for contrastive learning. It achieves better graph classification performance under both self-supervised and semi-supervised settings. The thorough ablation study also demonstrates the ... | Recommendation: 8: accept, good paper | Area: Unsupervised and Self-supervised learning | Review:
This paper proposes a new graph contrastive paradigm named CLEP, which learns the latent community structure via variational inference for contrastive learning. It achieves better graph classification performance under both self-supervised and semi-supervised settings. The thorough ablation study also demonstra... |
Defending against all perturbed instances is the target of the most current defense mechanisms, and yet in practice only a subset of instances might be selected by the attacker. This paper aims at a new defense mechanism to minimize the worst case loss across all subsets. To solve this optimization problem, the authors... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
Defending against all perturbed instances is the target of the most current defense mechanisms, and yet in practice only a subset of instances might be selected by the attacker. This paper aims at a new defense mechanism to minimize the worst case loss across all subsets. To solve this optimization problem, the... |
This paper investigates the problem of training neural network policies in many-player games that are capable of generalizing to any number of player, rather than just a specific number N. They accomplish this through two methods. The first is augmentation, in which the number of players N is an input into the policy n... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper investigates the problem of training neural network policies in many-player games that are capable of generalizing to any number of player, rather than just a specific number N. They accomplish this through two methods. The first is augmentation, in which the number of players N is an input into the ... |
The authors investigate the generalization error in optimizing the expected adversarial loss. They utilize the generalization-optimization decomposition and techniques from stability analyses to reveal the minimax optimal optimization algorithm for the adversarial loss. For the lower bound, they show that any algorithm... | Recommendation: 5: marginally below the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The authors investigate the generalization error in optimizing the expected adversarial loss. They utilize the generalization-optimization decomposition and techniques from stability analyses to reveal the minimax optimal optimization algorithm for the adversarial loss. For the lower bound, they show that any a... |
This paper proposes to use a large dataset to train a pre-trained BERT-like deep model for patient representation using unsupervised learning approach. It then fine-tine the pre-trained model using data with labels for Treatment Effect Estimation. The motivation is based on the fact that the data with labels is limited... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper proposes to use a large dataset to train a pre-trained BERT-like deep model for patient representation using unsupervised learning approach. It then fine-tine the pre-trained model using data with labels for Treatment Effect Estimation. The motivation is based on the fact that the data with labels is... |
This paper experimentally shows that errors occur as the batch sizes for training and inference vary in
a deep learning model by caused as the GEMM operation on the GPU.
An interesting finding! It experimentally finds that the different batch sizes during training and inference will affect the model performance due to ... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper experimentally shows that errors occur as the batch sizes for training and inference vary in
a deep learning model by caused as the GEMM operation on the GPU.
An interesting finding! It experimentally finds that the different batch sizes during training and inference will affect the model performance... |
This paper introduces a set of data augmentation techniques for predicting dense error maps in semantic segmentation.
Strengths:
1. The empirical observation that two separate encoding networks (one for the input image and the other for the prediction map) can boost the prediction performance is a plus.
Weaknesses:
... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper introduces a set of data augmentation techniques for predicting dense error maps in semantic segmentation.
Strengths:
1. The empirical observation that two separate encoding networks (one for the input image and the other for the prediction map) can boost the prediction performance is a plus.
Weakn... |
Main concern. Transform GCN explainability into subgraph-graph matching.
Protocol for quantifying the explainability: “As the ground-truth explanations are usually unknown, it is tough to quantitatively evaluate the excellence of explanations. There, we follow Wang et al. (2021b) and employ the predictive accu- racy ... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
Main concern. Transform GCN explainability into subgraph-graph matching.
Protocol for quantifying the explainability: “As the ground-truth explanations are usually unknown, it is tough to quantitatively evaluate the excellence of explanations. There, we follow Wang et al. (2021b) and employ the predictive acc... |
The paper studies linear contextual bandits and linear MDPs with misspecifications. It proposes a new thresholding scheme for data selection, which can be used with OFUL and LSVI seamlessly. For linear contextual bandit, DS-OFUL yields regret upper bounds of same order as that of the well-specified setting. The paper a... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper studies linear contextual bandits and linear MDPs with misspecifications. It proposes a new thresholding scheme for data selection, which can be used with OFUL and LSVI seamlessly. For linear contextual bandit, DS-OFUL yields regret upper bounds of same order as that of the well-specified setting. The... |
The paper argues that SGD with freezing the first embedding layer can perform competitively with the AdamW for modern vision models, by empirically tests the method on 5 finetuning benchmarks.
### Strengths
1. The SGD (freeze-embed) is simple but effective on all the models shown in the experiments.
2. The method perf... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The paper argues that SGD with freezing the first embedding layer can perform competitively with the AdamW for modern vision models, by empirically tests the method on 5 finetuning benchmarks.
### Strengths
1. The SGD (freeze-embed) is simple but effective on all the models shown in the experiments.
2. The met... |
This paper focuses on investigating the performance of existing
non-contrastive methods for link prediction in both transductive and inductive
settings. In their experiments, they find that BGRL generally performs well in transductive settings, but poorly in the more realistic inductive settings, which motivates them... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper focuses on investigating the performance of existing
non-contrastive methods for link prediction in both transductive and inductive
settings. In their experiments, they find that BGRL generally performs well in transductive settings, but poorly in the more realistic inductive settings, which motiva... |
Multicalibration requires calibration of a predictor even conditioned on membership in S where S comes form a family of protected subsets. Formally, we require $E[y] = E[f(x)]$ conditioned on x belonging to S and $f(x) \approx p$. If we relax this to allow $E[(y - f(x))] = \alpha$, conditioned on $f(x) \approx p$, we h... | Recommendation: 3: reject, not good enough | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
Multicalibration requires calibration of a predictor even conditioned on membership in S where S comes form a family of protected subsets. Formally, we require $E[y] = E[f(x)]$ conditioned on x belonging to S and $f(x) \approx p$. If we relax this to allow $E[(y - f(x))] = \alpha$, conditioned on $f(x) \approx ... |
For many protein design applications, the user specifies a motif, and then needs to build a scaffold of additional amino acids that will act to stabilize the motif. This paper proposes an interesting technique for generating scaffolds conditional on motifs. First, a 3D diffusion model of protein structures is fit. Then... | Recommendation: 8: accept, good paper | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
For many protein design applications, the user specifies a motif, and then needs to build a scaffold of additional amino acids that will act to stabilize the motif. This paper proposes an interesting technique for generating scaffolds conditional on motifs. First, a 3D diffusion model of protein structures is f... |
A framework called ModReduce was proposed to distillate knowledge from a teacher model by combining response-based, feature-based, and relation-based methods. The proposed distillation process is also divided into two phases: an offline distillation followed by an online distillation. The experimental results show that... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
A framework called ModReduce was proposed to distillate knowledge from a teacher model by combining response-based, feature-based, and relation-based methods. The proposed distillation process is also divided into two phases: an offline distillation followed by an online distillation. The experimental results s... |
This paper proposes kNN Prompting, which addresses the problem of in-context learning (ICL) that the context length is limited, so that the performance cannot scale with the number of available training examples. Similar to ICL, kNN Prompting also formulates the downstream tasks as prompt completion tasks. Differently,... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper proposes kNN Prompting, which addresses the problem of in-context learning (ICL) that the context length is limited, so that the performance cannot scale with the number of available training examples. Similar to ICL, kNN Prompting also formulates the downstream tasks as prompt completion tasks. Diff... |
In this paper, the authors developed an auto-encoder-based anomaly detection model called DyAD (Dynamic system Anomaly Detection) for time series anomaly detection. They primarily dealt with the high-dimensionality issue caused via viewing time series anomaly detection as hypothesis testing on dynamical systems. They d... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
In this paper, the authors developed an auto-encoder-based anomaly detection model called DyAD (Dynamic system Anomaly Detection) for time series anomaly detection. They primarily dealt with the high-dimensionality issue caused via viewing time series anomaly detection as hypothesis testing on dynamical systems... |
To improve upon the current best performance of models that predict molecular properties in typical drug discovery scenarios (where the number of molecules with known properties are very small), the authors propose MHNfs, an embedding-based few-shot learner using Modern Hopfield Networks to provide learned associative ... | Recommendation: 8: accept, good paper | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
To improve upon the current best performance of models that predict molecular properties in typical drug discovery scenarios (where the number of molecules with known properties are very small), the authors propose MHNfs, an embedding-based few-shot learner using Modern Hopfield Networks to provide learned asso... |
This paper proposes to use a pre-trained large language model to specify rewards for reinforcement learning. The model takes a description of the task and reward, and optionally a small set of demonstrations and outputs a binary reward for a new queried game state. Based on this reward, an RL agent is trained to comple... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes to use a pre-trained large language model to specify rewards for reinforcement learning. The model takes a description of the task and reward, and optionally a small set of demonstrations and outputs a binary reward for a new queried game state. Based on this reward, an RL agent is trained t... |
The paper describes simple pre-training strategy for multi-mer contact prediction. The strategy is shown to provide empirical benefit across several benchmark tasks. However the paper does not compare to some of the most widely used methods for predicting contacts (AlphaFold and RosettaFold).
Strengths:
* The chief b... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper describes simple pre-training strategy for multi-mer contact prediction. The strategy is shown to provide empirical benefit across several benchmark tasks. However the paper does not compare to some of the most widely used methods for predicting contacts (AlphaFold and RosettaFold).
Strengths:
* The... |
This paper proposed SuaVE, a new identifiable model for causal representations learning. SuaVE is an extension of i-VAE model in the sense that: 1, an additional latent causal variables layer is generated conditioned on the latent noise variables; 2, the latent causal variables are modelled via linear Gaussian models ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposed SuaVE, a new identifiable model for causal representations learning. SuaVE is an extension of i-VAE model in the sense that: 1, an additional latent causal variables layer is generated conditioned on the latent noise variables; 2, the latent causal variables are modelled via linear Gaussian... |
This paper studies the forward pass of a convolutional decoder in the frequency domain. It finds that the CNN layer forward propagates each frequency component in the
spectrum map independently to other frequency components.
It also finds that the CNN operations make a convolutional decoder network more likely to weake... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper studies the forward pass of a convolutional decoder in the frequency domain. It finds that the CNN layer forward propagates each frequency component in the
spectrum map independently to other frequency components.
It also finds that the CNN operations make a convolutional decoder network more likely ... |
This paper proposes an autoencoder architecture for structured representation learning. In particular, the proposed model incorporates latent variables at different layers of the convolutional decoder through Structural Transform (StrTfm) layers (inspired by FiLM (Perez et al., 2018) and Ada-In (Karras et al., 2019) la... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes an autoencoder architecture for structured representation learning. In particular, the proposed model incorporates latent variables at different layers of the convolutional decoder through Structural Transform (StrTfm) layers (inspired by FiLM (Perez et al., 2018) and Ada-In (Karras et al., ... |
First, I am an emergent reviewer and not an expert on OT. Therefore, I'm not sure whether the literature is reviewed properly and whether the evaluation setup is fine.
This work presents a formulation of the unbalanced optimal transport problem. To solve the unbalance problem, it relies on the formalism of semi-coupli... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
First, I am an emergent reviewer and not an expert on OT. Therefore, I'm not sure whether the literature is reviewed properly and whether the evaluation setup is fine.
This work presents a formulation of the unbalanced optimal transport problem. To solve the unbalance problem, it relies on the formalism of sem... |
The paper proposes a new model that is able to construct the instance-dependent prompt through reinforcement learning. It allows the agent to perform different editing techniques to update instructions, few-shot exemplars, and verbalizers at test time to construct query-dependent prompts efficiently. Results on differe... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper proposes a new model that is able to construct the instance-dependent prompt through reinforcement learning. It allows the agent to perform different editing techniques to update instructions, few-shot exemplars, and verbalizers at test time to construct query-dependent prompts efficiently. Results on... |
This paper comprehensively study how to mitigate or get rid of catastrophic overfitting from perspectives: data initialisation, network architecture and optimisation, and thus provides a bag of tricks for how to improve and stabilise fast adversarial training.
Specifically, the authors find that randomly masking out so... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper comprehensively study how to mitigate or get rid of catastrophic overfitting from perspectives: data initialisation, network architecture and optimisation, and thus provides a bag of tricks for how to improve and stabilise fast adversarial training.
Specifically, the authors find that randomly maskin... |
The paper shows how piecewise-linear networks limit the tightness of "piecewise-linear limited" (PLL) certification procedures (such as Lipschitz-based certification). In particular, it demonstrates why piecewise-linear networks that require a small capacity to produce a robust boundary can require a much larger capaci... | 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 shows how piecewise-linear networks limit the tightness of "piecewise-linear limited" (PLL) certification procedures (such as Lipschitz-based certification). In particular, it demonstrates why piecewise-linear networks that require a small capacity to produce a robust boundary can require a much large... |
This paper studies the InfoNCE loss widely used in contrastive learning from the view of Lovasz theta function. In particular, regardless the positive-pair term of InfoNCE, minimizing InfoNCE corresponds to solving the Lovasz theta function. Inspired by this, the authors relax the constraint of Lovasz theta function be... | Recommendation: 3: reject, not good enough | Area: Unsupervised and Self-supervised learning | Review:
This paper studies the InfoNCE loss widely used in contrastive learning from the view of Lovasz theta function. In particular, regardless the positive-pair term of InfoNCE, minimizing InfoNCE corresponds to solving the Lovasz theta function. Inspired by this, the authors relax the constraint of Lovasz theta fun... |
This paper focuses on byzantine robustness of FL. Typically, the server uses robust aggregation rules to ensure that byzantine clients do not hinder learning. However, the performance of most aggregation rules is degraded when data is non-IID across different clients. The authors reveal the root causes of 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:
This paper focuses on byzantine robustness of FL. Typically, the server uses robust aggregation rules to ensure that byzantine clients do not hinder learning. However, the performance of most aggregation rules is degraded when data is non-IID across different clients. The authors reveal the root causes of the p... |
This paper presents a new NeRF model, PAC-NeRF, that incorporate physical particle dynamic into the estimation to simultaneously estimate particle parameters and geometries. The particle dynamics are assumed to follow the conservation laws of continuum mechanics, and a hybrid Eulerian-Lagrangian representation is propo... | Recommendation: 8: accept, good paper | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper presents a new NeRF model, PAC-NeRF, that incorporate physical particle dynamic into the estimation to simultaneously estimate particle parameters and geometries. The particle dynamics are assumed to follow the conservation laws of continuum mechanics, and a hybrid Eulerian-Lagrangian representation ... |
This paper studies the influence of warm-start in actor-critic algorithms. The authors provide a finite-time analysis of the impact of the approximation error and the sub-optimality of the initial policy.
### Strength:
- This paper is overall well-written and easy to follow
- Treating actor-critic algorithms as a per... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper studies the influence of warm-start in actor-critic algorithms. The authors provide a finite-time analysis of the impact of the approximation error and the sub-optimality of the initial policy.
### Strength:
- This paper is overall well-written and easy to follow
- Treating actor-critic algorithms ... |
To learn a personalized and lightweight recommendation model for each client, this paper proposes to learn user-specific lightweight models that are deployed on the device side. Furthermore, it introduces a dual personalization mechanism to learn personalized item embedding tables and the score function for each user. ... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
To learn a personalized and lightweight recommendation model for each client, this paper proposes to learn user-specific lightweight models that are deployed on the device side. Furthermore, it introduces a dual personalization mechanism to learn personalized item embedding tables and the score function for eac... |
Authors propose two new metrics to evaluate feature attribution explanations: completeness and soundness which are based on algorithm theory. Authors show limitations of several existing order-based and model-retraining based metrics. The metrics are evaluated on several explainability methods such as GradCAM, DeepSHAP... | Recommendation: 5: marginally below the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
Authors propose two new metrics to evaluate feature attribution explanations: completeness and soundness which are based on algorithm theory. Authors show limitations of several existing order-based and model-retraining based metrics. The metrics are evaluated on several explainability methods such as GradCAM, ... |
This work builds upon the notion of a complementary learning system that consists of multiple memories: episodic memory and the semantic memory in this case. Although this notion is utilized in many works, the main contribution of this work is in designing the episodic memory to contain samples from the current batch t... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This work builds upon the notion of a complementary learning system that consists of multiple memories: episodic memory and the semantic memory in this case. Although this notion is utilized in many works, the main contribution of this work is in designing the episodic memory to contain samples from the current... |
The paper studies the problem of imitation learning, building on the recent IQ-learn framework. Instead of an adversarial reward-policy loss like GAIL, IQ-learn instead parameterizes the Q-function so that the policy can be directly extracted. While IQ-learn works fine, the paper notes that some of the practical tricks... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper studies the problem of imitation learning, building on the recent IQ-learn framework. Instead of an adversarial reward-policy loss like GAIL, IQ-learn instead parameterizes the Q-function so that the policy can be directly extracted. While IQ-learn works fine, the paper notes that some of the practica... |
The authors propose a reinforcement learning approach to road vectorization. In contrast with previous works, the task is modeled as generating a graph as a variable-length edge sequence.
The pipeline consists of multiple stages: semantic segmentation from RGB images (this results in binary maps), a transformer-based... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The authors propose a reinforcement learning approach to road vectorization. In contrast with previous works, the task is modeled as generating a graph as a variable-length edge sequence.
The pipeline consists of multiple stages: semantic segmentation from RGB images (this results in binary maps), a transform... |
This paper provides a theoretical understanding to the advantage of using warm-start in Actor-Critic (AC) type RL algorithms. The authors achieve this by establishing finite-time bounds of warm-start AC algorithms, which explicitly capture the impact of using warm-start. In terms of technical approach, the authors cast... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper provides a theoretical understanding to the advantage of using warm-start in Actor-Critic (AC) type RL algorithms. The authors achieve this by establishing finite-time bounds of warm-start AC algorithms, which explicitly capture the impact of using warm-start. In terms of technical approach, the auth... |
The authors propose to improve the task of protein-protein interaction prediction by providing a GNN approach with additional features on the proteins based on their structures and sequences.
Good
the approach is easy to combine/add to previous GNN models and seems to provide a consistent improvement of 5%
Bad
The n... | Recommendation: 3: reject, not good enough | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The authors propose to improve the task of protein-protein interaction prediction by providing a GNN approach with additional features on the proteins based on their structures and sequences.
Good
the approach is easy to combine/add to previous GNN models and seems to provide a consistent improvement of 5%
B... |
The paper builds on a recent modelling of the hierarchical clustering of [Zugner et al. 2021] to learn discrete structures by gradient descent.
The structure is modelled using two row-stochastic parameter matrices: $A$ parametrizes the probability distributions of each sample of belonging to a cluster ($n$ distribution... | Recommendation: 6: marginally above the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The paper builds on a recent modelling of the hierarchical clustering of [Zugner et al. 2021] to learn discrete structures by gradient descent.
The structure is modelled using two row-stochastic parameter matrices: $A$ parametrizes the probability distributions of each sample of belonging to a cluster ($n$ dist... |
The paper studies the problem of learning DAGs from observational data. In contrast to previous work under the continuous framework, the authors aim to learn a DAG that is transportable to other distributions over the same observables that respect a set of conditional independences. In a few words, the approach consist... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
The paper studies the problem of learning DAGs from observational data. In contrast to previous work under the continuous framework, the authors aim to learn a DAG that is transportable to other distributions over the same observables that respect a set of conditional independences. In a few words, the approach... |
This paper proposes a way to regularize GFlowNet policies using an optimal transport cost that can encourage either similarity or dissimilarity of policies at states that are close in the state space. This cost is converted into several proxy regularization terms that can be added to the trajectory balance objective in... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a way to regularize GFlowNet policies using an optimal transport cost that can encourage either similarity or dissimilarity of policies at states that are close in the state space. This cost is converted into several proxy regularization terms that can be added to the trajectory balance obje... |
The paper develops an algorithm, ESCHER, to reduce the variance of CFR with deep learning by avoiding importance sampling.
The intuition looks pretty reasonable to me. However, I still have some questions regarding ESCHER before I can recommend accepting.
1. In Section 3 it's claimed sampling using a fixed distribution... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The paper develops an algorithm, ESCHER, to reduce the variance of CFR with deep learning by avoiding importance sampling.
The intuition looks pretty reasonable to me. However, I still have some questions regarding ESCHER before I can recommend accepting.
1. In Section 3 it's claimed sampling using a fixed dist... |
This paper presents a comparison of different deep learning methods for tourist flow 1-hour forecast in 30 touristic locations of Salzburg city. The data is an hourly entry number of visitors at each of the 30 sites during 3 years, with addition of some information on the day (holidays or not, etc.), and then a test us... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper presents a comparison of different deep learning methods for tourist flow 1-hour forecast in 30 touristic locations of Salzburg city. The data is an hourly entry number of visitors at each of the 30 sites during 3 years, with addition of some information on the day (holidays or not, etc.), and then a... |
This paper proposes a technique to improve the efficiency of multi-head attention.
The idea is to apply the multi-resolution decomposition with orthogonal bases inspired by wavelets to attention matrics or the V matrices.
The Haar Wavelet (box-car like) is considered in this paper, resulting in up/down sampling combi... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper proposes a technique to improve the efficiency of multi-head attention.
The idea is to apply the multi-resolution decomposition with orthogonal bases inspired by wavelets to attention matrics or the V matrices.
The Haar Wavelet (box-car like) is considered in this paper, resulting in up/down sampli... |
In this paper, they propose a new formulation of scaled dot-product attention for token by token inference on continual stream. That reduces the time complexity from O(n^2 d) to O(n d) while keeping the outputs and weights identical to the original Transformer outputs and weights.
Strengths:
- The proposed approach is ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
In this paper, they propose a new formulation of scaled dot-product attention for token by token inference on continual stream. That reduces the time complexity from O(n^2 d) to O(n d) while keeping the outputs and weights identical to the original Transformer outputs and weights.
Strengths:
- The proposed appr... |
This paper studies the problem of differentially private federated learning. More specifically, in order to improve the utility of the differentially private model in federated learning, the authors propose to combine secure multiparty computation (MPC) with differential privacy (DP) when solving the regularized logist... | 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 studies the problem of differentially private federated learning. More specifically, in order to improve the utility of the differentially private model in federated learning, the authors propose to combine secure multiparty computation (MPC) with differential privacy (DP) when solving the regularize... |
The authors propose a new mechanism based on the energy-models for the task of out of distribution detection on graphs. They show that the graph neural networks trained with supervised loss objective can be intrinsically effective to detect OOD data on which the model should avoid prediction. The energy based method pr... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The authors propose a new mechanism based on the energy-models for the task of out of distribution detection on graphs. They show that the graph neural networks trained with supervised loss objective can be intrinsically effective to detect OOD data on which the model should avoid prediction. The energy based m... |
The paper considers a timely topic: connection between fairness and privacy-preservation of an ML model. The angle is interesting: fairness is considered from the point of view that the model performance gains for user's data should correlate with the privacy leakage. It is well known that e.g. training with differenti... | Recommendation: 3: reject, not good enough | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper considers a timely topic: connection between fairness and privacy-preservation of an ML model. The angle is interesting: fairness is considered from the point of view that the model performance gains for user's data should correlate with the privacy leakage. It is well known that e.g. training with di... |
Authors compare multiple neural contextual bandit algorithms on multiple data sets along with non-neural baselines.
Strength:
* This type of empirical investigation is important and undersupplied in the literature.
Weakness:
* Empirical evidence is conditional, so to draw broad conclusions, a massive number of ... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
Authors compare multiple neural contextual bandit algorithms on multiple data sets along with non-neural baselines.
Strength:
* This type of empirical investigation is important and undersupplied in the literature.
Weakness:
* Empirical evidence is conditional, so to draw broad conclusions, a massive nu... |
This paper focuses on the problem of debiasing representations learned by encoder-decoder models when demographic information is missing. The authors propose Unsupervised Locality-based Proxy Label assignment (ULPL), which assigns a binary “proxy” label to each example based on whether the model incurs a larger or smal... | 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 focuses on the problem of debiasing representations learned by encoder-decoder models when demographic information is missing. The authors propose Unsupervised Locality-based Proxy Label assignment (ULPL), which assigns a binary “proxy” label to each example based on whether the model incurs a larger... |
A multi-agent RL algorithm is proposed to solve the multi-SKU inventory optimization in a single-echelon problem. The goal is to obtain the order quantity for each SKU while the overall capacity constraint is held and the total profit is maximized. The profit is calculated by getting the sell amount, purchase cost, hol... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
A multi-agent RL algorithm is proposed to solve the multi-SKU inventory optimization in a single-echelon problem. The goal is to obtain the order quantity for each SKU while the overall capacity constraint is held and the total profit is maximized. The profit is calculated by getting the sell amount, purchase c... |
This paper describes a method for subset sampling based on conditional Poisson sampling, called neural conditional Poisson subset sampling. This approach builds on the general conditional Poisson sampling approach, which samples each element in the subset independently, and conditions this procedure to return subsets ... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper describes a method for subset sampling based on conditional Poisson sampling, called neural conditional Poisson subset sampling. This approach builds on the general conditional Poisson sampling approach, which samples each element in the subset independently, and conditions this procedure to return ... |
This paper introduces a continual learning technique for the active learning. In particular, a set of the continual learning techniques are adopted and integrated into each model training procedure at each cycle of active learning. The proposed CAL regards the newly-queried sample at each round a new task as in the con... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper introduces a continual learning technique for the active learning. In particular, a set of the continual learning techniques are adopted and integrated into each model training procedure at each cycle of active learning. The proposed CAL regards the newly-queried sample at each round a new task as in... |
This paper studied the Constrained MDP (CMDP) problem, under the infinite-horizon average-reward setup of linear MDP. The model-free algorithm proposed in this paper has better regret bound in terms of total steps $T$ compared with the previous algorithm for tabular CMDP, under weaker assumptions. Under the same assump... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper studied the Constrained MDP (CMDP) problem, under the infinite-horizon average-reward setup of linear MDP. The model-free algorithm proposed in this paper has better regret bound in terms of total steps $T$ compared with the previous algorithm for tabular CMDP, under weaker assumptions. Under the sam... |
The paper proposes a framework based on probabilistic logical reasoning that allows to train object detection models with arbitrary types of weak supervision. The proposed architecture consists of two components: symbolic reasoning and deep learning architecture.
Such a structure is easy to implement weakly supervised... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper proposes a framework based on probabilistic logical reasoning that allows to train object detection models with arbitrary types of weak supervision. The proposed architecture consists of two components: symbolic reasoning and deep learning architecture.
Such a structure is easy to implement weakly su... |
This paper proves that transformers with single-head attention and a specific embedding scheme can express any polynomial by constructing the weights required for forming a polynomial.
Pros:
- The construction is clear and interesting.
Cons:
- Most of the claims in the paper are just simple consequences of the exact c... | Recommendation: 5: marginally below the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper proves that transformers with single-head attention and a specific embedding scheme can express any polynomial by constructing the weights required for forming a polynomial.
Pros:
- The construction is clear and interesting.
Cons:
- Most of the claims in the paper are just simple consequences of the... |
This paper addresses the issue of image copy detection and retrieval from large databases by incorporating a robust feature extractor and a scalable, efficient search algorithm. The feature extractor is a neural network that has been trained to extract descriptors that are robust to various transformations. The search ... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper addresses the issue of image copy detection and retrieval from large databases by incorporating a robust feature extractor and a scalable, efficient search algorithm. The feature extractor is a neural network that has been trained to extract descriptors that are robust to various transformations. The... |
The paper shows that any no-regret dynamic converges to a Nash equilibrium in strongly convex games, assuming the convexity is sufficiently strong relative to the Lipschitz constant of the gradient function. Somewhat better bounds are shown for specific no-regret dynamics.
The results are interesting but not extremely ... | Recommendation: 5: marginally below the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The paper shows that any no-regret dynamic converges to a Nash equilibrium in strongly convex games, assuming the convexity is sufficiently strong relative to the Lipschitz constant of the gradient function. Somewhat better bounds are shown for specific no-regret dynamics.
The results are interesting but not ex... |
The paper describes a solution for improving performance in multi-agent reinforcement learning. The term "consciousness" in this work essentially refers to attention based mechanisms and to the fact that the authors adopt and approach that resides in the idea of dividing the general problem into sub-parts in order to t... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper describes a solution for improving performance in multi-agent reinforcement learning. The term "consciousness" in this work essentially refers to attention based mechanisms and to the fact that the authors adopt and approach that resides in the idea of dividing the general problem into sub-parts in or... |
This paper presents a transformer memory architecture for processing events from an event camera. A block of events are converted into positional embeddings, passed through self attention, and then fused with a memory tensor which is updated with events over time. The memory tensor can then be decoded at any time to pr... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper presents a transformer memory architecture for processing events from an event camera. A block of events are converted into positional embeddings, passed through self attention, and then fused with a memory tensor which is updated with events over time. The memory tensor can then be decoded at any ti... |
This paper extends the contrastive learning on graph by replacing the contrastive between negative samples with that between class clusters with theoretical analysis. It possesses the attractive characteristic of reducing complexity. Experimental evaluations demonstrate its effectiveness and efficiency.
Strength
- It m... | Recommendation: 3: reject, not good enough | Area: Unsupervised and Self-supervised learning | Review:
This paper extends the contrastive learning on graph by replacing the contrastive between negative samples with that between class clusters with theoretical analysis. It possesses the attractive characteristic of reducing complexity. Experimental evaluations demonstrate its effectiveness and efficiency.
Strengt... |
This paper proposes the first clean-image backdoor attack, which only poisons the training labels without touching the training samples. Experimental results demonstrate that the proposed method can achieve an attack success rate of up to 98.2% on the images containing the trigger pattern.
Strength:
1. Generally speaki... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes the first clean-image backdoor attack, which only poisons the training labels without touching the training samples. Experimental results demonstrate that the proposed method can achieve an attack success rate of up to 98.2% on the images containing the trigger pattern.
Strength:
1. Generall... |
This paper studies the effect of cascaded convolutional decoder network on the spectrum domain, and presents several findings summarized in the experiment section.
Strengths:
1. The authors performed a detailed study on the effect of different components in a decoder network on the spectrum domain by discrete Fourier ... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper studies the effect of cascaded convolutional decoder network on the spectrum domain, and presents several findings summarized in the experiment section.
Strengths:
1. The authors performed a detailed study on the effect of different components in a decoder network on the spectrum domain by discrete ... |
In this papers, the authors propose a simple transformer-based model to predict minimum inhibitory concentration (MICs) . The trasnformer is built on the top of XGBoost-extracted features. The model is trained to simultaneously predict MIC of 14 antibiotics for Salmonela.
======== Post-rebuttal update ========
The au... | Recommendation: 3: reject, not good enough | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
In this papers, the authors propose a simple transformer-based model to predict minimum inhibitory concentration (MICs) . The trasnformer is built on the top of XGBoost-extracted features. The model is trained to simultaneously predict MIC of 14 antibiotics for Salmonela.
======== Post-rebuttal update ========... |
This paper expands prompt engineering techniques to multi-modality scenarios. Specifically, the paper sought to find how best to prompt a set of unimodal and multi-modal models to solve multi-modal tasks that are difficult to solve independently or would require large amount of training data to tackle. The paper also c... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper expands prompt engineering techniques to multi-modality scenarios. Specifically, the paper sought to find how best to prompt a set of unimodal and multi-modal models to solve multi-modal tasks that are difficult to solve independently or would require large amount of training data to tackle. The pape... |
The paper proposes the use of convolutional neural networks to act as a recombination operator in a population-based black-box optimization algorithm (aka evolutionary algorithm). That operator is combining the solutions making the population, ranked according to their fitness, with weights provided by the learned conv... | Recommendation: 3: reject, not good enough | Area: Optimization (eg, convex and non-convex optimization) | Review:
The paper proposes the use of convolutional neural networks to act as a recombination operator in a population-based black-box optimization algorithm (aka evolutionary algorithm). That operator is combining the solutions making the population, ranked according to their fitness, with weights provided by the lear... |
This paper studies the phenomenon of oversmoothing in GNNs and transformers; and presents the ContraNorm layer. The experiments on real-world dataset show the efficacy of this method.
Strength:
It’s a nice method with good motivations and theoretical insights.
The paper is generally well written.
The experiments show ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper studies the phenomenon of oversmoothing in GNNs and transformers; and presents the ContraNorm layer. The experiments on real-world dataset show the efficacy of this method.
Strength:
It’s a nice method with good motivations and theoretical insights.
The paper is generally well written.
The experimen... |
This paper presented the mini-batch extension of STP method in the finite-sum problem. The paper analyzed the complexity of minibatch STP in both non-convex and convex cases and examine the performance of MiSTP under various settings. The experiment showed that MiSTP converged faster than STP in terms of number of epoc... | Recommendation: 5: marginally below the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper presented the mini-batch extension of STP method in the finite-sum problem. The paper analyzed the complexity of minibatch STP in both non-convex and convex cases and examine the performance of MiSTP under various settings. The experiment showed that MiSTP converged faster than STP in terms of number... |
This paper proposes two variance-reduced (VR) algorithms for a specific nonconvex strongly concave min-max problem obtained after reformulating the policy evaluation problem as a min-max game with nonlinear function approximation for the value function of the policy. They provide the first O(1/K) convergence and the po... | Recommendation: 6: marginally above the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper proposes two variance-reduced (VR) algorithms for a specific nonconvex strongly concave min-max problem obtained after reformulating the policy evaluation problem as a min-max game with nonlinear function approximation for the value function of the policy. They provide the first O(1/K) convergence an... |
This paper studies the distributionally robust optimization problem by leveraging topological knowledge. Specifically, two types of topologies are considered, namely physical-based topology and data-driven topology. The physical-based topology is based on neighborhood information, and the data-driven topology is based ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper studies the distributionally robust optimization problem by leveraging topological knowledge. Specifically, two types of topologies are considered, namely physical-based topology and data-driven topology. The physical-based topology is based on neighborhood information, and the data-driven topology i... |
This paper proposed a more robust training method for neurotrophic devices
Strength:
* The paper is well struct and easy to follow. Results are clearly listed in tables and figures.
Weaknesses:
* I think the claim of "hardware-oriented" is a bit of stretch as the core of the target problem in improving robustness on ... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
This paper proposed a more robust training method for neurotrophic devices
Strength:
* The paper is well struct and easy to follow. Results are clearly listed in tables and figures.
Weaknesses:
* I think the claim of "hardware-oriented" is a bit of stretch as the core of the target problem in improving robust... |
The paper proposes a new learning setup called cascade in which an agent observes a dynamical system and then changes its initial state’s conditions to achieve a desired goal. The paper also proposes an interesting approach for learning a probabilistic scoring function over an “event tree” data structure to efficiently... | Recommendation: 8: accept, good paper | Area: General Machine Learning | Review:
The paper proposes a new learning setup called cascade in which an agent observes a dynamical system and then changes its initial state’s conditions to achieve a desired goal. The paper also proposes an interesting approach for learning a probabilistic scoring function over an “event tree” data structure to eff... |
This paper proves that attacking a polynomial-time classifier is NP-complete, and training a polynomial-time model that is robust on a
single input is Σ_2^P-complete. The authors proposed a method that evaluates on the fly if a model is robust on a specific point by running
an adversarial attack on the input. Based on ... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper proves that attacking a polynomial-time classifier is NP-complete, and training a polynomial-time model that is robust on a
single input is Σ_2^P-complete. The authors proposed a method that evaluates on the fly if a model is robust on a specific point by running
an adversarial attack on the input. B... |
This paper presents theoretical proofs of perfectly secure (zero KL divergence) steganography, particularly that it depends on being induced by a coupling (necessary and sufficient condition) and that for efficiency it will also depend on a minimum entropy coupling (that benefit from very recent advances in the approxi... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper presents theoretical proofs of perfectly secure (zero KL divergence) steganography, particularly that it depends on being induced by a coupling (necessary and sufficient condition) and that for efficiency it will also depend on a minimum entropy coupling (that benefit from very recent advances in the... |
Calibration is important for models used for predictions in online advertising systems. It can be challenging because there is a form of adverse selection, where any over-predictions are more likely to result in these ads being shown. There is also a challenge because the training data is typically the ads that were sh... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
Calibration is important for models used for predictions in online advertising systems. It can be challenging because there is a form of adverse selection, where any over-predictions are more likely to result in these ads being shown. There is also a challenge because the training data is typically the ads that... |
This paper uses distribution matching and randomization to reduce Trojan specificity. The author proposes to use 1-Wasserstein distance for the design of the loss function. The authors conduct experiments on MNIST, CIFAR-10 and CIFAR-100, show that the method proposed by the authors can improve the difficulty of Trojan... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper uses distribution matching and randomization to reduce Trojan specificity. The author proposes to use 1-Wasserstein distance for the design of the loss function. The authors conduct experiments on MNIST, CIFAR-10 and CIFAR-100, show that the method proposed by the authors can improve the difficulty o... |
The paper "Property inference attacks against t-SNE plots" suggests a way to infer some properties of the data set from an unlabeled t-SNE plot of that data set. For example, if the data contain images of males and females, then the attack would aim to infer the fraction of males from the shape of the unlabeled t-SNE p... | Recommendation: 5: marginally below the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper "Property inference attacks against t-SNE plots" suggests a way to infer some properties of the data set from an unlabeled t-SNE plot of that data set. For example, if the data contain images of males and females, then the attack would aim to infer the fraction of males from the shape of the unlabeled... |
The paper proposes an end-to-end encoder-decoder approach for ligand-protein binding (blind docking) where the ligand can conform while the protein structure is assumed fixed. In the encoder part, the individual atomic representation of ligand and residue-based representation of ligand is learnt through graph networks ... | 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 an end-to-end encoder-decoder approach for ligand-protein binding (blind docking) where the ligand can conform while the protein structure is assumed fixed. In the encoder part, the individual atomic representation of ligand and residue-based representation of ligand is learnt through graph n... |
This paper attempts to solve OOD problem in object detection.
It tries to reduce the impact of lacking unknown data for supervision and leverage in-distribution data to improve the model’s discrimination ability with both Information Bottleneck and Reverse Information Bottleneck.
It first uses a standard IB network to ... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper attempts to solve OOD problem in object detection.
It tries to reduce the impact of lacking unknown data for supervision and leverage in-distribution data to improve the model’s discrimination ability with both Information Bottleneck and Reverse Information Bottleneck.
It first uses a standard IB net... |
This paper identified three key ingredients to speed up the learning process for MBRL models. Specifically, the authors, building on TD-MPC, proposed a framework with policy pretraining, seeding, and finetuning with interactive learning, and showed great sample-efficiency improvement on a set of challenging visual cont... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper identified three key ingredients to speed up the learning process for MBRL models. Specifically, the authors, building on TD-MPC, proposed a framework with policy pretraining, seeding, and finetuning with interactive learning, and showed great sample-efficiency improvement on a set of challenging vis... |
The authors analyse via mathematical modelling and 600+ participating students the effects of facial beauty. They find both rewards (including based on perceived trustworthiness) and penalties from +/- beauty.
This is an interesting study from a social perspective. The findings bear merit and all is well described. The... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The authors analyse via mathematical modelling and 600+ participating students the effects of facial beauty. They find both rewards (including based on perceived trustworthiness) and penalties from +/- beauty.
This is an interesting study from a social perspective. The findings bear merit and all is well descri... |
The paper studies efficient method for computing convex conjugate arising in wasserstein OT problem. Although exactly and approximately computing so is believed to be hard in prior work, this work proposes a new method based on amortized approximation scheme could be used to computing the exact conjugate. This amortiza... | Recommendation: 6: marginally above the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
The paper studies efficient method for computing convex conjugate arising in wasserstein OT problem. Although exactly and approximately computing so is believed to be hard in prior work, this work proposes a new method based on amortized approximation scheme could be used to computing the exact conjugate. This ... |
This paper is about explanations in reinforcement learning (RL). The idea is to focus on trajectories in offline settings (in a dataset) and find a cluster of trajectories that best explains a particular decision/behavior. The approach finds embeddings of trajectories, then clusters them, and verifies which of the clus... | Recommendation: 5: marginally below the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper is about explanations in reinforcement learning (RL). The idea is to focus on trajectories in offline settings (in a dataset) and find a cluster of trajectories that best explains a particular decision/behavior. The approach finds embeddings of trajectories, then clusters them, and verifies which of ... |
The paper studies the limitations to deepen graph transformers and argue that substructure learning gets increasingly harder in the canonical formulation of the graph transformers. To address this limitation, the paper proposes a variant of graph transformers that explicitly models substructure attention and encoding. ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper studies the limitations to deepen graph transformers and argue that substructure learning gets increasingly harder in the canonical formulation of the graph transformers. To address this limitation, the paper proposes a variant of graph transformers that explicitly models substructure attention and en... |
The paper studies adversarial attacks on the state observation (sensor inputs) channel of a RL agent. Different from previous work (in particular [Zhang et al., 2020]), the authors consider the stealthiness of the adversarial attacks. The contribution of the paper is that it defines the concept of detectability for adv... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper studies adversarial attacks on the state observation (sensor inputs) channel of a RL agent. Different from previous work (in particular [Zhang et al., 2020]), the authors consider the stealthiness of the adversarial attacks. The contribution of the paper is that it defines the concept of detectability... |
This paper studies the distributionally robust optimization problem. Due to existing methods depending on expensive group annotations, the proposed AGRO method proposes to discover different groups via an adversarial opponent. Specifically, there are two models are designed in the AGRO, namely the task model and the gr... | Recommendation: 5: marginally below the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper studies the distributionally robust optimization problem. Due to existing methods depending on expensive group annotations, the proposed AGRO method proposes to discover different groups via an adversarial opponent. Specifically, there are two models are designed in the AGRO, namely the task model an... |
This paper proposes Metadata Archaeology, a unifying and general framework for uncovering latent metadata categories.
This paper introduces and validate the approach of Metadata Archaeology via Probe Dynamics (MAP-D): leveraging the training dynamics of curated data subsets called probe suites to infer other examples’ ... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
This paper proposes Metadata Archaeology, a unifying and general framework for uncovering latent metadata categories.
This paper introduces and validate the approach of Metadata Archaeology via Probe Dynamics (MAP-D): leveraging the training dynamics of curated data subsets called probe suites to infer other ex... |
This paper introduces training algorithms that are robust to bounded dynamics in the data, with a focus on label updates.
----------------------
Update: I thank the authors for their informative feedback. I have updated my overall recommendation but remain unconfident about it.
The work seems theoretically solid. ... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper introduces training algorithms that are robust to bounded dynamics in the data, with a focus on label updates.
----------------------
Update: I thank the authors for their informative feedback. I have updated my overall recommendation but remain unconfident about it.
The work seems theoretically... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.