review stringlengths 5 16.6k | score stringclasses 5
values | area stringclasses 12
values | text stringlengths 31 5.65k |
|---|---|---|---|
This paper analyzes the articulation structure that could emerge in emergent language. The paper follows the studies of deep learning-based emergent communication models, which have been studied for half of the decade. In particular, the study adopts the framework of Chaabouni et al. (2020).
The authors tested if they ... | Recommendation: 8: accept, good paper | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper analyzes the articulation structure that could emerge in emergent language. The paper follows the studies of deep learning-based emergent communication models, which have been studied for half of the decade. In particular, the study adopts the framework of Chaabouni et al. (2020).
The authors tested ... |
The papers approaches the problem of contrastive learning of time series representations by learning parameters for time series augmentations. This allows good "views" of the time series to be directly incorporated into SimCLR, leading to reasonable performance gains against the baselines.
Strenghts:
- The paper is eas... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
The papers approaches the problem of contrastive learning of time series representations by learning parameters for time series augmentations. This allows good "views" of the time series to be directly incorporated into SimCLR, leading to reasonable performance gains against the baselines.
Strenghts:
- The pape... |
This paper proposes to use a frozen vision and language pre-trained model (CLIP) for open-vocabulary object detection. The image encoder from CLIP is used as the backbone and the textual encoder from CLIP is used for region proposal classification. A trainable detector head is added to generate region proposals (for lo... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes to use a frozen vision and language pre-trained model (CLIP) for open-vocabulary object detection. The image encoder from CLIP is used as the backbone and the textual encoder from CLIP is used for region proposal classification. A trainable detector head is added to generate region proposals... |
In this paper, the authors propose a method that tend to learn task-independent representations. Specifically, the proposed method uses multiple masks to generate different subspace representations to avoid feature suppression. The proposed method also employs the uncertainty modeling technique against variance from st... | Recommendation: 8: accept, good paper | Area: Unsupervised and Self-supervised learning | Review:
In this paper, the authors propose a method that tend to learn task-independent representations. Specifically, the proposed method uses multiple masks to generate different subspace representations to avoid feature suppression. The proposed method also employs the uncertainty modeling technique against variance... |
The paper proposes a DAT label smoothing approach. The results are backed by the theoretical evidences. However, the main novelty is minimalistic.
Strength:
- The paper is well written.
- The empirical results are backed by theoretical evidences
- Simple solution
Weakness
- The paper used an already established idea w... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes a DAT label smoothing approach. The results are backed by the theoretical evidences. However, the main novelty is minimalistic.
Strength:
- The paper is well written.
- The empirical results are backed by theoretical evidences
- Simple solution
Weakness
- The paper used an already establishe... |
This paper proposes to incorporate model-specific prior knowledge into optimizers by modifying the gradients according to a set of model-specific hyper-parameters. They name the method as Gradient Re-parameterization, and the optimizers are named RepOptimizers. They show that a VGG-style plain network can be trained wi... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes to incorporate model-specific prior knowledge into optimizers by modifying the gradients according to a set of model-specific hyper-parameters. They name the method as Gradient Re-parameterization, and the optimizers are named RepOptimizers. They show that a VGG-style plain network can be tr... |
This work considers the problem of designing quantum algorithms to perform adversarial training of classifiers using randomized smoothing. Using classical method the number of required samples (and therefore queries to the classifier) is roughly $O(1/\epsilon^2)$, where $\epsilon$ is the target error parameter for the... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
This work considers the problem of designing quantum algorithms to perform adversarial training of classifiers using randomized smoothing. Using classical method the number of required samples (and therefore queries to the classifier) is roughly $O(1/\epsilon^2)$, where $\epsilon$ is the target error parameter... |
This paper provides a tool for understanding the performance changes between different updates of ML as service applications. The developed tool is interactive and helps the users discover meaningful coherent different data slices where the performance is changing on, called ChangeLists. The model is built on CLIP and ... | 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 provides a tool for understanding the performance changes between different updates of ML as service applications. The developed tool is interactive and helps the users discover meaningful coherent different data slices where the performance is changing on, called ChangeLists. The model is built on C... |
The authors propose a new framework for insights selection in health data based on reinforcement learning. Insights are actionable interpretations of analysis of data that originates from users' behavior.
The paper proposes to create a large list of candidate insights, which are then scored and filtered based on these ... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The authors propose a new framework for insights selection in health data based on reinforcement learning. Insights are actionable interpretations of analysis of data that originates from users' behavior.
The paper proposes to create a large list of candidate insights, which are then scored and filtered based o... |
In this paper, the authors introduce a novel framework, called Maximum-Entropy Rewarded Reinforcement Learning (MERRL), which can select training data to cover more possible queries that may appear in unknown worlds.
Strengths:
In INTRODUCTION, the authors first proved the relationship between `training set entropy’ an... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
In this paper, the authors introduce a novel framework, called Maximum-Entropy Rewarded Reinforcement Learning (MERRL), which can select training data to cover more possible queries that may appear in unknown worlds.
Strengths:
In INTRODUCTION, the authors first proved the relationship between `training set ent... |
Motivated by the data inefficiency problem of off-policy algorithms, like PPO, this paper proposed a parallel version of the Q-learning algorithm. Specifically, this approach uses three parallel processes, namely actor, V-learner and P-learner, to perform a DDPG-like learning process. Empirical evaluations on six Isaac... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
Motivated by the data inefficiency problem of off-policy algorithms, like PPO, this paper proposed a parallel version of the Q-learning algorithm. Specifically, this approach uses three parallel processes, namely actor, V-learner and P-learner, to perform a DDPG-like learning process. Empirical evaluations on s... |
This paper presents an approach to learn target (goal) and support (obstacle) gradient fields for navigation planning from a set of example configurations. As opposed to many other learning approaches for navigation planning, it does not rely on demonstration trajectories but only on sets of successful target configura... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper presents an approach to learn target (goal) and support (obstacle) gradient fields for navigation planning from a set of example configurations. As opposed to many other learning approaches for navigation planning, it does not rely on demonstration trajectories but only on sets of successful target c... |
This paper proposes a meta-learning approach for few shot node classification on a graph. In particular, a ProtoNet metric-based approach is adopted. However, unlike traditional ProtoNet where the class prototypes are computed from class samples in an unweighted manner, this paper assigns different importance scores to... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a meta-learning approach for few shot node classification on a graph. In particular, a ProtoNet metric-based approach is adopted. However, unlike traditional ProtoNet where the class prototypes are computed from class samples in an unweighted manner, this paper assigns different importance s... |
The authors propose a novel loss function which is tailored specifically for learning A*-heuristics. $L$* captures (1) heuristic monotonicity and (2) optimal A*-search efficiency in terms of #node expansions. The authors motivates theoretically why $L$* overcome known limitations to L_p loss in terms of A* heuristics, ... | Recommendation: 8: accept, good paper | Area: General Machine Learning | Review:
The authors propose a novel loss function which is tailored specifically for learning A*-heuristics. $L$* captures (1) heuristic monotonicity and (2) optimal A*-search efficiency in terms of #node expansions. The authors motivates theoretically why $L$* overcome known limitations to L_p loss in terms of A* heur... |
This paper studies the problem of graph neural network unlearning. Unlike the partitioning-based unlearning strategies, the proposed solution works on two defined loss functions that reflect the target of edge deletion kind of unlearning request. One loss function measures the Deleted Edge Consistency: the predictivene... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper studies the problem of graph neural network unlearning. Unlike the partitioning-based unlearning strategies, the proposed solution works on two defined loss functions that reflect the target of edge deletion kind of unlearning request. One loss function measures the Deleted Edge Consistency: the pred... |
The work is about verbalizing entities and the relationships between them. According to the authors, existing
works deal with entity and relation verbalization separately. This work aims to solve entity and relation verbalization
in a unified form by adaptive generative language models to the task of entity and relat... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The work is about verbalizing entities and the relationships between them. According to the authors, existing
works deal with entity and relation verbalization separately. This work aims to solve entity and relation verbalization
in a unified form by adaptive generative language models to the task of entity a... |
This paper address the problem of multi-task dense scene understanding, including semantic segmentation, human parsing etc. A transformer based method is proposed, and task-specific prompt tokens are designed to enable the transformer architecture to utilize its full capacity on all tasks. Authors modified the multi-he... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper address the problem of multi-task dense scene understanding, including semantic segmentation, human parsing etc. A transformer based method is proposed, and task-specific prompt tokens are designed to enable the transformer architecture to utilize its full capacity on all tasks. Authors modified the ... |
This paper studied multimodal VAE. Compared to existing methods that rely on the conditional independence assumption, this paper relaxes such assumption by introducing Set Multimodal VAE (SMVAE). The author conducted experiments on three multimodal datasets to verify the effectiveness of SMVAE.
Strength:
+ This paper... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper studied multimodal VAE. Compared to existing methods that rely on the conditional independence assumption, this paper relaxes such assumption by introducing Set Multimodal VAE (SMVAE). The author conducted experiments on three multimodal datasets to verify the effectiveness of SMVAE.
Strength:
+ Th... |
This paper proposes a representation learning method for image-based state observations for RL learning algorithms. This is proposed as an auxiliary task built around self-supervision between successive observations, with a distinct focus on disentangling the learned representations to differentiate between relevant an... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes a representation learning method for image-based state observations for RL learning algorithms. This is proposed as an auxiliary task built around self-supervision between successive observations, with a distinct focus on disentangling the learned representations to differentiate between rel... |
This paper proposes a new meta-training approach for language models. For an instance of task instruction, input and label, their method, called Flipped Learning, increases the likelihood of the task instruction given the input and correct label, while decreasing the likelihood of the task instruction given the input a... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a new meta-training approach for language models. For an instance of task instruction, input and label, their method, called Flipped Learning, increases the likelihood of the task instruction given the input and correct label, while decreasing the likelihood of the task instruction given the... |
This paper introduces a new MIM framework that dynamically focuses on patch reconstructions based on the degree of difficulty (i.e., the nearby visible patches) during pre-training. Besides, this paper proposes to self-distill intermediate features from the momentum encoder. Experiments show that it outperforms previou... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper introduces a new MIM framework that dynamically focuses on patch reconstructions based on the degree of difficulty (i.e., the nearby visible patches) during pre-training. Besides, this paper proposes to self-distill intermediate features from the momentum encoder. Experiments show that it outperforms... |
This work extends the 1-block model in Taylor et al. (2022) to the d-block model. Compared with the LIF model, the d-block model achieves accelerated computing on GPU by using fewer sequential operations.
Strength:
The proposed model indeed enables accelerated computing on GPU, and achieves sota results on some benchm... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
This work extends the 1-block model in Taylor et al. (2022) to the d-block model. Compared with the LIF model, the d-block model achieves accelerated computing on GPU by using fewer sequential operations.
Strength:
The proposed model indeed enables accelerated computing on GPU, and achieves sota results on som... |
This paper proposes the LU mechanism for the phenomenon of Grokking observed by Power et. al. 2020 - which states that the train and test curves follow an L and U shaped curve with weight norms respectively. They verify this observation in a student teacher setup, and show that it can arise in non-algorithmic datasets ... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper proposes the LU mechanism for the phenomenon of Grokking observed by Power et. al. 2020 - which states that the train and test curves follow an L and U shaped curve with weight norms respectively. They verify this observation in a student teacher setup, and show that it can arise in non-algorithmic d... |
The authors study federated learning starting from pre-trained models in both the image and text setting. They study a wide range of FL algorithms. Some key observations are that (a) the gap of iid to non-iid closes (b) ranking of FL algorithms is different under pre-training (c) effect of heterogneity are not severed.... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The authors study federated learning starting from pre-trained models in both the image and text setting. They study a wide range of FL algorithms. Some key observations are that (a) the gap of iid to non-iid closes (b) ranking of FL algorithms is different under pre-training (c) effect of heterogneity are not ... |
This paper studies transfer learning for tabular data. The authors take MetaMIMIC for experiments, where there are 12 medical targets. They train on 11 targets and test on the leaveout target. They find that transfer learning provides definitive advantages over gradient boosted decision trees. They further compare self... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper studies transfer learning for tabular data. The authors take MetaMIMIC for experiments, where there are 12 medical targets. They train on 11 targets and test on the leaveout target. They find that transfer learning provides definitive advantages over gradient boosted decision trees. They further comp... |
The paper proposes an amortized inference approach combined with a temporal attention mechanism for parallel training of dynamical models given long irregularly sampled trajectories. Experimental results on three synthetic datasets demonstrate improved mean squared error compared to baselines.
**Strengths**
- The paper... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
The paper proposes an amortized inference approach combined with a temporal attention mechanism for parallel training of dynamical models given long irregularly sampled trajectories. Experimental results on three synthetic datasets demonstrate improved mean squared error compared to baselines.
**Strengths**
- T... |
This paper applied tensor completion to search for the optimal hyperparameters for machine learning methods. Precisely, the combinations of parameter candidates are mapped on the tensor values with multiple modes. By the efficient sampling trick (Zhang, 2019) followed by tensor completion, the optimal parameters are se... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
This paper applied tensor completion to search for the optimal hyperparameters for machine learning methods. Precisely, the combinations of parameter candidates are mapped on the tensor values with multiple modes. By the efficient sampling trick (Zhang, 2019) followed by tensor completion, the optimal parameter... |
This paper proves the memorization capacity of Transformer. The main technical theory is: given $N$ input-output pairs of sequence, this paper constructs a transformer that memorizes them with permutation equivariance, under slightly stronger assumptions. The mathematical tools used in this paper adopts the approach fr... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper proves the memorization capacity of Transformer. The main technical theory is: given $N$ input-output pairs of sequence, this paper constructs a transformer that memorizes them with permutation equivariance, under slightly stronger assumptions. The mathematical tools used in this paper adopts the app... |
This paper introduces a novel labeling scheme for extractive text summarization. Extractive summarization datasets are most often derived from abstractive datasets using a greedy labeling scheme, providing a single extractive reference per document. Greedy multi-hot labels cause sparsity/under-fitting issues so the aut... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper introduces a novel labeling scheme for extractive text summarization. Extractive summarization datasets are most often derived from abstractive datasets using a greedy labeling scheme, providing a single extractive reference per document. Greedy multi-hot labels cause sparsity/under-fitting issues so... |
This paper studies the dynamics of a quadratic model trained with quadratic loss.
The effective dynamics of NTK is derived.
Authors argue that edge of stability behavior in neural networks is correlated with the behavior in quadratic regression models.
Strength:
The topic should be interesting to optimization communi... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper studies the dynamics of a quadratic model trained with quadratic loss.
The effective dynamics of NTK is derived.
Authors argue that edge of stability behavior in neural networks is correlated with the behavior in quadratic regression models.
Strength:
The topic should be interesting to optimization... |
This paper focuses on the bias in continual learning. According to the bias, this paper designs a novel method, named learning without prejudices (LwP) to discourage malignant forgetting and encourage benign forgetting. The main contributions can be summarized as:
(a) Presenting a novel framework, termed "continual un... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper focuses on the bias in continual learning. According to the bias, this paper designs a novel method, named learning without prejudices (LwP) to discourage malignant forgetting and encourage benign forgetting. The main contributions can be summarized as:
(a) Presenting a novel framework, termed "cont... |
This paper presents a new metric to measure the fairness of language models on a toxicity labeled dataset. Authors exploit the fact that a fair model should yield higher perplexity scores for toxic sentences. The proposed metric is bounded between [0, 1] and utilizes Mann-Whitney U statistical test to quantify the tend... | 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 presents a new metric to measure the fairness of language models on a toxicity labeled dataset. Authors exploit the fact that a fair model should yield higher perplexity scores for toxic sentences. The proposed metric is bounded between [0, 1] and utilizes Mann-Whitney U statistical test to quantify ... |
The paper proposes multi-agent hybrid-POMDPS which uses a centralized training scheme as well as models a communication process between the agents. Experiments are conducted on standard benchmarks to show the superiority of the proposed method.
strengths:
-the paper proposes an approach called hybrid-POMDPs
-it adds t... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper proposes multi-agent hybrid-POMDPS which uses a centralized training scheme as well as models a communication process between the agents. Experiments are conducted on standard benchmarks to show the superiority of the proposed method.
strengths:
-the paper proposes an approach called hybrid-POMDPs
-i... |
This submission proposed a distributed computing scheme for matrix multiplication that can handle two problems: adversarial attacks and straggler effects. The experimental results demonstrate the robustness and fastness of the proposed approach by comparing it with two baselines.
Strength:
1. The motivation is clear.... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
This submission proposed a distributed computing scheme for matrix multiplication that can handle two problems: adversarial attacks and straggler effects. The experimental results demonstrate the robustness and fastness of the proposed approach by comparing it with two baselines.
Strength:
1. The motivation i... |
The aim of this paper is to learn probability distributions in high dimensions with dominance constraints. To this end, the authors are exploiting the Choquet order between probability distributions, and introduce the notion of Variational Dominance Criterion, a divergence measure that captures the relative spread of a... | Recommendation: 8: accept, good paper | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The aim of this paper is to learn probability distributions in high dimensions with dominance constraints. To this end, the authors are exploiting the Choquet order between probability distributions, and introduce the notion of Variational Dominance Criterion, a divergence measure that captures the relative spr... |
The paper studies the delayed impact of fairness in machine learning and provides a new model that models this effect as opposed to the commonly studied static notions of fairness. The paper introduces an algorithm called ELF (Enforcing Long-term Fairness) to provide high confidence delayed impact fairness guarantees (... | 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 studies the delayed impact of fairness in machine learning and provides a new model that models this effect as opposed to the commonly studied static notions of fairness. The paper introduces an algorithm called ELF (Enforcing Long-term Fairness) to provide high confidence delayed impact fairness guar... |
This paper proposed a method of communication-efficient federated learning framework for a vertically distributed graph. Specifically, a lazy aggregation rule is proposed to reduce the communication rounds. A new strategy called stale updates skips aggregation in specific iterations to reduce the cost during vertical t... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper proposed a method of communication-efficient federated learning framework for a vertically distributed graph. Specifically, a lazy aggregation rule is proposed to reduce the communication rounds. A new strategy called stale updates skips aggregation in specific iterations to reduce the cost during ve... |
This paper studies the robustness of neural ODEs, by studying the condition when neural ODEs are contractive, i.e., the trajectories of the ODE converge to the same value exponentially fast. Based on existing lemmas on the condition of contraction, the authors propose a weight regularization based approach during train... | Recommendation: 3: reject, not good enough | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper studies the robustness of neural ODEs, by studying the condition when neural ODEs are contractive, i.e., the trajectories of the ODE converge to the same value exponentially fast. Based on existing lemmas on the condition of contraction, the authors propose a weight regularization based approach duri... |
This paper introduces GLM-130B, a bilingual English and Chinese LLM. The model is trained for 400B tokens (half on English and half on Chinese). The model follows a bidirectional attention architecture with the self-supervised blank infilling objective. They also include a small percentage of MIP tokens. Int4 quantiza... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper introduces GLM-130B, a bilingual English and Chinese LLM. The model is trained for 400B tokens (half on English and half on Chinese). The model follows a bidirectional attention architecture with the self-supervised blank infilling objective. They also include a small percentage of MIP tokens. Int4 ... |
This paper studies the domain adversarial training problem and propose to use label smoothing for domain discrimination. Specifically, motivated by the observation that different domains from VLCS dataset show small difference, thus a soft domain label should be applied to domain adversarial training instead of hard la... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper studies the domain adversarial training problem and propose to use label smoothing for domain discrimination. Specifically, motivated by the observation that different domains from VLCS dataset show small difference, thus a soft domain label should be applied to domain adversarial training instead of... |
This paper presents a new analysis framework based on differential inclusion (DI) theory for analyzing value-based reinforcement learning (RL) methods with discontinuous behavior policy changes (such as Q-learning with an epsilon-greedy behavior.) Using the new framework, the paper provides an explanation of the asympt... | Recommendation: 8: accept, good paper | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper presents a new analysis framework based on differential inclusion (DI) theory for analyzing value-based reinforcement learning (RL) methods with discontinuous behavior policy changes (such as Q-learning with an epsilon-greedy behavior.) Using the new framework, the paper provides an explanation of th... |
The paper proposes a computational drug repurposing framework that not only predicts the "treatment probabilities" between drugs and diseases but also predicts the path-based, "testable" mechanisms of action as their biomedical explanations, all based on a "massive" biomedical knowledge graph. Specifically, the paper u... | Recommendation: 3: reject, not good enough | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper proposes a computational drug repurposing framework that not only predicts the "treatment probabilities" between drugs and diseases but also predicts the path-based, "testable" mechanisms of action as their biomedical explanations, all based on a "massive" biomedical knowledge graph. Specifically, the... |
The paper proposes an approach for initializing sparse Mixture of Expert (MoE) models from dense checkpoints. The approach called upcycling reuses an already trained dense checkpoint by copying all the parameters from the original checkpoint (except for the MoE router parameters). The experts in this new model are repl... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The paper proposes an approach for initializing sparse Mixture of Expert (MoE) models from dense checkpoints. The approach called upcycling reuses an already trained dense checkpoint by copying all the parameters from the original checkpoint (except for the MoE router parameters). The experts in this new model ... |
This paper presents a module to perform layer attention for CNNs and Vision Transformers in linear-complexity. The core idea is to leverage the representations of previous layers in a recurrent form so that the quadratic-complexity self-attention is avoided. Experiments on image classification, object detection, and in... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper presents a module to perform layer attention for CNNs and Vision Transformers in linear-complexity. The core idea is to leverage the representations of previous layers in a recurrent form so that the quadratic-complexity self-attention is avoided. Experiments on image classification, object detection... |
This paper proposes a method that extends Q-learning to train offline on counterexamples of unsafe states, generated through a safety specification combined with set covering algorithm.
## Strengths
**************Method:**************
- The set cover formulation is very interesting and seems like a nice way to abstr... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes a method that extends Q-learning to train offline on counterexamples of unsafe states, generated through a safety specification combined with set covering algorithm.
## Strengths
**************Method:**************
- The set cover formulation is very interesting and seems like a nice way ... |
The paper proposes a new way of integrating logical constraints in the training of deep neural networks. In particular, the authors try to solve the problem that the current models often learn to satisfy the constraints by learning the *obvious solution* (e.g., given $A \to B$ the models learn to set $A = \bot$). The l... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes a new way of integrating logical constraints in the training of deep neural networks. In particular, the authors try to solve the problem that the current models often learn to satisfy the constraints by learning the *obvious solution* (e.g., given $A \to B$ the models learn to set $A = \bot$... |
The paper empirically evaluates how well the NTK and empirical NTK approach captures the performance of neural networks. For that, it analyses the scaling behavior of the kernels in comparison to the original network wrt. dataset size and network width. The paper finds that neither NTK nor empirical NTK can capture the... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The paper empirically evaluates how well the NTK and empirical NTK approach captures the performance of neural networks. For that, it analyses the scaling behavior of the kernels in comparison to the original network wrt. dataset size and network width. The paper finds that neither NTK nor empirical NTK can cap... |
This paper first proves that the self-attention mechanism in neural networks is a special form of support vector regression. Based on this conclusion, the authors proposed a principle to develop attention and designed the Attention-BN and the Attention-SH for improving accuracy and efficiency. Experimental results demo... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper first proves that the self-attention mechanism in neural networks is a special form of support vector regression. Based on this conclusion, the authors proposed a principle to develop attention and designed the Attention-BN and the Attention-SH for improving accuracy and efficiency. Experimental resu... |
This paper investigate the existence of trojan model for backdoor attack --- a model that is in theory similar to the Bayes optimal yet performs poorly under some universal backdoor trigger. The paper formalizes its notion, proves its conditional existence and show an attack algorithm that generates universal backdoor ... | 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 investigate the existence of trojan model for backdoor attack --- a model that is in theory similar to the Bayes optimal yet performs poorly under some universal backdoor trigger. The paper formalizes its notion, proves its conditional existence and show an attack algorithm that generates universal b... |
This paper considered a multi-task reinforcement learning problem where all tasks share the same state space, action spaces, and the transition dynamics. A skill machine is proposed to solve complex task involving temporal and concurrent composition which can be learned from reward machine. The authors prove that an ag... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper considered a multi-task reinforcement learning problem where all tasks share the same state space, action spaces, and the transition dynamics. A skill machine is proposed to solve complex task involving temporal and concurrent composition which can be learned from reward machine. The authors prove th... |
This paper introduces a novel method based on parameter averaging to estimate accurate and robust feature importance in tabular data setting. The proposed method first initializes and trains multiple instances of a shallow network (referred as local masks) with different random seeds for a downstream task and then obta... | 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 introduces a novel method based on parameter averaging to estimate accurate and robust feature importance in tabular data setting. The proposed method first initializes and trains multiple instances of a shallow network (referred as local masks) with different random seeds for a downstream task and t... |
The paper proposes two “orthogonal” methods to improve existing deep-learning based image and video compression models. First, the authors swap out existing scalar quantization modules with a uniform vector quantization map - they show that this VQ map doesn’t need to be learned but can be uniform (e.g. hexagonal or oc... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes two “orthogonal” methods to improve existing deep-learning based image and video compression models. First, the authors swap out existing scalar quantization modules with a uniform vector quantization map - they show that this VQ map doesn’t need to be learned but can be uniform (e.g. hexagon... |
This paper proposes the idea of random layerwise token dropping method. Extensive experimental restuls empirically demostrate that randomly dropping tokens during pre-training preserves the task accuracy but saves the computational cost for pre-training large language models.
### Strengths
- Simple but effective ide... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes the idea of random layerwise token dropping method. Extensive experimental restuls empirically demostrate that randomly dropping tokens during pre-training preserves the task accuracy but saves the computational cost for pre-training large language models.
### Strengths
- Simple but effec... |
The authors present GAMR, a brain-inspired architecture for visual reasoning. This architecture is transformer-based, but additionally utilizes an LSTM controller module and a memory module. The architecture empirically works well on visual reasoning task. This is exciting from an engineering perspective, but also of i... | Recommendation: 8: accept, good paper | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
The authors present GAMR, a brain-inspired architecture for visual reasoning. This architecture is transformer-based, but additionally utilizes an LSTM controller module and a memory module. The architecture empirically works well on visual reasoning task. This is exciting from an engineering perspective, but a... |
This paper presents a framework called Entity-Factorized Markov Decision Process (EFMDP), where the task is factorized across different entities, and the state and goal configurations of the entities are permutation-invariant. Based on this framework, the paper studies various permutation-invariant architecture designs... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper presents a framework called Entity-Factorized Markov Decision Process (EFMDP), where the task is factorized across different entities, and the state and goal configurations of the entities are permutation-invariant. Based on this framework, the paper studies various permutation-invariant architecture... |
The paper proposes a robust explanation method for NLP tasks using Causal Proxy Model (CPM). Given a black-box model to be explained, the CPM tries to simulate both the factual and counterfactual performance of that model. With CPM, one can have 1) an explanation of the black-box model, 2) comparable factual performanc... | 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 robust explanation method for NLP tasks using Causal Proxy Model (CPM). Given a black-box model to be explained, the CPM tries to simulate both the factual and counterfactual performance of that model. With CPM, one can have 1) an explanation of the black-box model, 2) comparable factual pe... |
Motivated by the insight that DP budgets are conservative (i.e. a given budget overestimates the privacy/membership inference information that can be extracted realistically), this work designs a canary-based method to produce a more realistic estimate of the privacy budget in FL settings. The threat model is
- safe bu... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
Motivated by the insight that DP budgets are conservative (i.e. a given budget overestimates the privacy/membership inference information that can be extracted realistically), this work designs a canary-based method to produce a more realistic estimate of the privacy budget in FL settings. The threat model is
-... |
The paper deals with the credit scoring problem. The paper formulates the problem as a logistic regression tree problem, where the data points are split into segments and each segment has a logistic regression model to model the probability of default. The paper then proposes an algorithm that can jointly learn segment... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper deals with the credit scoring problem. The paper formulates the problem as a logistic regression tree problem, where the data points are split into segments and each segment has a logistic regression model to model the probability of default. The paper then proposes an algorithm that can jointly learn... |
This paper analyzes the global style transfer method from the perspective of Fourier analysis. Specifically, the formula of style transfer is converted into Fourier transform under the umbrella of global statistics. Thus the relationships between Fourier phase (Fourier amplitude) and content loss (Gram matrix) are achi... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper analyzes the global style transfer method from the perspective of Fourier analysis. Specifically, the formula of style transfer is converted into Fourier transform under the umbrella of global statistics. Thus the relationships between Fourier phase (Fourier amplitude) and content loss (Gram matrix) ... |
The paper proposed a novel and nice procedure for automating interactive theorem proving, which combines state-of-the-art approaches to ML for (interactive) theorem proving. The key idea is to have a proof written in natural language (either provided by a human user or generated by a language model) translated into a f... | Recommendation: 8: accept, good paper | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper proposed a novel and nice procedure for automating interactive theorem proving, which combines state-of-the-art approaches to ML for (interactive) theorem proving. The key idea is to have a proof written in natural language (either provided by a human user or generated by a language model) translated ... |
The paper introduce a novel dataset difficulty metric based on how long humans have to view an image in order to classify it correctly. The authors release the difficulty metrics for ImageNet and ObjectNet datasets as well as distribution of image difficulties in those datasets. The paper also introduce a new metric pr... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper introduce a novel dataset difficulty metric based on how long humans have to view an image in order to classify it correctly. The authors release the difficulty metrics for ImageNet and ObjectNet datasets as well as distribution of image difficulties in those datasets. The paper also introduce a new m... |
This submission proposed a new plug-in module which decomposes the 3D vision attention in video modeling. The proposed module, T2D, decomposes a self-attention among 3D spaces (spatial: XY and temporal: T) into three 2D space attention: XY, XT and YT. The proposed idea is very similar to what has been shown in R(2+1)D ... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This submission proposed a new plug-in module which decomposes the 3D vision attention in video modeling. The proposed module, T2D, decomposes a self-attention among 3D spaces (spatial: XY and temporal: T) into three 2D space attention: XY, XT and YT. The proposed idea is very similar to what has been shown in ... |
The problem of domain extension with language is addressed in this paper. The proposed method (LADS) uses a CLIP model's domain-level knowledge to learn a latent feature augmentation of the training set. It does not require any unseen domain samples and instead relies on written descriptions of the training and unseen ... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The problem of domain extension with language is addressed in this paper. The proposed method (LADS) uses a CLIP model's domain-level knowledge to learn a latent feature augmentation of the training set. It does not require any unseen domain samples and instead relies on written descriptions of the training and... |
The authors introduce an energy-constrained diffusion model which encodes a batch of instances from a dataset into evolutionary states that progressively incorporate other instances’ information by their interactions. In their method, the diffusion process is constrained by descent criteria w.r.t. a principled energy f... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The authors introduce an energy-constrained diffusion model which encodes a batch of instances from a dataset into evolutionary states that progressively incorporate other instances’ information by their interactions. In their method, the diffusion process is constrained by descent criteria w.r.t. a principled ... |
The authors propose a novel method to handle the noisy pseudo-label issue in UDA settings. They first train normalizing flow models to estimate the distribution of the target domain bases on generated pseudo-labels. Then they use the generative models to construct the D-CFA on features in the source domain. Besides, th... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The authors propose a novel method to handle the noisy pseudo-label issue in UDA settings. They first train normalizing flow models to estimate the distribution of the target domain bases on generated pseudo-labels. Then they use the generative models to construct the D-CFA on features in the source domain. Bes... |
This paper proposed a strategy to augment the image feature into different domains. The process is guided by CLIP to not just transfer to another domain but also keep the semantic content. In the experiment ,we can see that the proposed method, LADS, indeed performs well in both in-distribution and out-of-distribution.... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposed a strategy to augment the image feature into different domains. The process is guided by CLIP to not just transfer to another domain but also keep the semantic content. In the experiment ,we can see that the proposed method, LADS, indeed performs well in both in-distribution and out-of-distr... |
This paper focuses on long tail behavior in deep probabilistic forecasting. Specifically, the authors first observe from the real-world datasets that the tail classes of the data do not necessarily correspond to the tail classes of the error. Then, in order to solve the long tail behavior in prediction error for deep p... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper focuses on long tail behavior in deep probabilistic forecasting. Specifically, the authors first observe from the real-world datasets that the tail classes of the data do not necessarily correspond to the tail classes of the error. Then, in order to solve the long tail behavior in prediction error fo... |
This paper studies the generalization of bilevel optimization problems which are widely used in machine learning, e.g., meta-learning, hyper-parameter optimization, and
reinforcement learning. Specifically, the authors conduct a thorough analysis of the generalization of first-order (gradient-based) methods for the bi... | Recommendation: 5: marginally below the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper studies the generalization of bilevel optimization problems which are widely used in machine learning, e.g., meta-learning, hyper-parameter optimization, and
reinforcement learning. Specifically, the authors conduct a thorough analysis of the generalization of first-order (gradient-based) methods fo... |
This paper discusses ascent regularization for training continuous normalizing flows (CNFs). This is motivated from Wasserstein gradient flows and results in an interesting regularization that encourages the learned model to be similar to the target distribution around a large interval of time values. I find this to be... | Recommendation: 6: marginally above the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
This paper discusses ascent regularization for training continuous normalizing flows (CNFs). This is motivated from Wasserstein gradient flows and results in an interesting regularization that encourages the learned model to be similar to the target distribution around a large interval of time values. I find th... |
This paper proposes a virtual OoD sample synthesizer suitable for federated learning, called FOSTER. It uses the class information collected by the client to train the generator on the server and broadcast. Under the premise of ensuring privacy, FOSTER uses the knowledge of other non-iid federated partners to generate ... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper proposes a virtual OoD sample synthesizer suitable for federated learning, called FOSTER. It uses the class information collected by the client to train the generator on the server and broadcast. Under the premise of ensuring privacy, FOSTER uses the knowledge of other non-iid federated partners to g... |
This paper address UDA with the mean teacher model where the teacher model generate pseudo labels to supervise the student model, and the teacher model is updated as the moving average of the student model. Mixup is adopted to help enhance generalization performance. Experiments on two datasets verify the effectiveness... | Recommendation: 3: reject, not good enough | Area: Unsupervised and Self-supervised learning | Review:
This paper address UDA with the mean teacher model where the teacher model generate pseudo labels to supervise the student model, and the teacher model is updated as the moving average of the student model. Mixup is adopted to help enhance generalization performance. Experiments on two datasets verify the effec... |
This paper propose a method to find the differences of how differently trained models make predictions. More precisely, it gives human-interpretable features (or patterns) on how different models recognize different classes. The proposed method heavily relies on how influences of training data on test data are measured... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper propose a method to find the differences of how differently trained models make predictions. More precisely, it gives human-interpretable features (or patterns) on how different models recognize different classes. The proposed method heavily relies on how influences of training data on test data are ... |
This paper presents F-VLM, which tackles open-vocabulary detection based on frozen pre-trained vision-language models (VLMs). The authors observe that frozen VLMs retain locality-sensitive features and are strong region classifiers. By only training a detection head upon frozen VLMs and utilizing regional VLM features ... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper presents F-VLM, which tackles open-vocabulary detection based on frozen pre-trained vision-language models (VLMs). The authors observe that frozen VLMs retain locality-sensitive features and are strong region classifiers. By only training a detection head upon frozen VLMs and utilizing regional VLM f... |
The paper conducts extensive experiments to study the scaling behavior of different neural network models on language problems. The experiment results show:
1. Different models architectures scale differently when the sizes of the models are changed,
2. At different compute regions, the model that perform best is diff... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The paper conducts extensive experiments to study the scaling behavior of different neural network models on language problems. The experiment results show:
1. Different models architectures scale differently when the sizes of the models are changed,
2. At different compute regions, the model that perform best... |
This paper proposes a model compression approach: Random Operation Access Specific Tile (ROAST) hashing. The authors consider three operations for block-based hashing to reduce memory usage, and they also introduce a global memory-sharing method to improve model accuracy. Experiments on BERT and ResNet show that the pr... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
This paper proposes a model compression approach: Random Operation Access Specific Tile (ROAST) hashing. The authors consider three operations for block-based hashing to reduce memory usage, and they also introduce a global memory-sharing method to improve model accuracy. Experiments on BERT and ResNet show tha... |
The paper proposes to apply Data Programming to semi-supervised Continual Learning (SSCL). The core idea is to first generate weak labeling functions (WLF) with existing tools like Snuba, where the labeling functions are then used to pseudo-label the unlabeled data points that are eventually all fed into training the d... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
The paper proposes to apply Data Programming to semi-supervised Continual Learning (SSCL). The core idea is to first generate weak labeling functions (WLF) with existing tools like Snuba, where the labeling functions are then used to pseudo-label the unlabeled data points that are eventually all fed into traini... |
The paper works on adversarial training, a well-known defense for adversarial samples. It aims to address the problem of robustness over-fitting, which refers to the phenomenon that training with a fixed budget degenerates model performance. The authors propose a method called Strength-Adaptive Adversarial Training, wh... | 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 works on adversarial training, a well-known defense for adversarial samples. It aims to address the problem of robustness over-fitting, which refers to the phenomenon that training with a fixed budget degenerates model performance. The authors propose a method called Strength-Adaptive Adversarial Trai... |
Online ML systems usually have explicit positive samples (E.g., ads conversion, click, buy, like, etc.) to train on. However, negative samples are often implicit and have a much higher volume concerning positives. A common technique in constructing a training data set is to sample from the negatives to having more reas... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
Online ML systems usually have explicit positive samples (E.g., ads conversion, click, buy, like, etc.) to train on. However, negative samples are often implicit and have a much higher volume concerning positives. A common technique in constructing a training data set is to sample from the negatives to having m... |
The paper presents a physical-based neural rendering network. The network codes lights and BRDF using 2 subnet-light sampling field network and material network. Then they are feed to construct light transport to generate radiance.
S:
The paper proposes to model light into direct and indirect lighting which takes local... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper presents a physical-based neural rendering network. The network codes lights and BRDF using 2 subnet-light sampling field network and material network. Then they are feed to construct light transport to generate radiance.
S:
The paper proposes to model light into direct and indirect lighting which tak... |
This paper presents a novel active learning-based approach that accurately addresses an long-existing pain point in deep learning compilation, namely tuning speed. Its contribution could be summarized as:
* Presented a careful investigation into the root cause of the unnecessarily long tuning speed, which is lack of di... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper presents a novel active learning-based approach that accurately addresses an long-existing pain point in deep learning compilation, namely tuning speed. Its contribution could be summarized as:
* Presented a careful investigation into the root cause of the unnecessarily long tuning speed, which is la... |
This paper investigates the pretraining practice on tabular data. It shows that using the object target labels during the pretraining stage is beneficial for the downstream performance. Several target-aware pretraining objectives are proposed. Experiments on various datasets validate the claim.
The paper compares vari... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper investigates the pretraining practice on tabular data. It shows that using the object target labels during the pretraining stage is beneficial for the downstream performance. Several target-aware pretraining objectives are proposed. Experiments on various datasets validate the claim.
The paper compa... |
In this work the authors present m6araw, a method for detecting m6A modifications directly from raw ONT long reads. As the authors nicely explain, current methods require a two step process: One for translating the raw signal from ONT into RNA sequence, another one for detecting the m6A modifications in those sequence ... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
In this work the authors present m6araw, a method for detecting m6A modifications directly from raw ONT long reads. As the authors nicely explain, current methods require a two step process: One for translating the raw signal from ONT into RNA sequence, another one for detecting the m6A modifications in those s... |
This submission proposes a quasistatic approach to derive the optimization algorithms's behavior on the manifold of minima. It has tried to understand the role of some parameters such as learning rate and batch size. Some unrealistic assumptions have been made, the results are not supported by rigorous analysis and are... | Recommendation: 3: reject, not good enough | Area: Optimization (eg, convex and non-convex optimization) | Review:
This submission proposes a quasistatic approach to derive the optimization algorithms's behavior on the manifold of minima. It has tried to understand the role of some parameters such as learning rate and batch size. Some unrealistic assumptions have been made, the results are not supported by rigorous analysis... |
This paper studies option learning in RL, and introduces a new framework (HiT-MDP). The authors prove that HiT-MDP is homomorphic equivalent to the standard SMDP formulation, and derive an on-policy policy gradient method with the new HiT-MDP formulation. On Mujuco environments, the proposed method is shown to be effec... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper studies option learning in RL, and introduces a new framework (HiT-MDP). The authors prove that HiT-MDP is homomorphic equivalent to the standard SMDP formulation, and derive an on-policy policy gradient method with the new HiT-MDP formulation. On Mujuco environments, the proposed method is shown to ... |
This paper presents GoBigger, a platform for studying cooperative and competitive behavior at large scale. The game is inspired by Agar. Even if the design of an environment like that described in the paper might be valuable for the community, it is difficult to identify clear technical contributions in this work. The ... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper presents GoBigger, a platform for studying cooperative and competitive behavior at large scale. The game is inspired by Agar. Even if the design of an environment like that described in the paper might be valuable for the community, it is difficult to identify clear technical contributions in this wo... |
This paper presents a new method for learning topological representations of 2D and 3D images. In contrast to existing work, the paper proposes a method that leverages the intrinsic structure of the "representation space", making it possible to derive (a) topological representations in a probabilistic fashion, while (b... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper presents a new method for learning topological representations of 2D and 3D images. In contrast to existing work, the paper proposes a method that leverages the intrinsic structure of the "representation space", making it possible to derive (a) topological representations in a probabilistic fashion, ... |
This paper presents an analysis of adaptive gradient methods with an emphasis on providing insights into the optimal setting of hyperparameters. The presented convergence results suggest settings that are in line with practical observations. From these results, the paper derives a "critical batch size", which minimizes... | Recommendation: 5: marginally below the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper presents an analysis of adaptive gradient methods with an emphasis on providing insights into the optimal setting of hyperparameters. The presented convergence results suggest settings that are in line with practical observations. From these results, the paper derives a "critical batch size", which m... |
This paper proposes a new quantization method to deal with high dimensional categorical features in recommendation systems.
Strengths:
The background and related work parts are very well written. It is easy to understand the different approaches, even a bit easier than the proposed approach itself.
The method is ... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a new quantization method to deal with high dimensional categorical features in recommendation systems.
Strengths:
The background and related work parts are very well written. It is easy to understand the different approaches, even a bit easier than the proposed approach itself.
The me... |
This paper studies the problem of one-bit completion: the task is to recover a vector $y \in \\{0, 1\\}^n$ from the observation of a subset of its positive entries. The model assume that there exists a matrix $R \in \\{0, 1\\}^{m \times n}$ containing additional information about the model, e.g. $m$ previous partial ob... | Recommendation: 8: accept, good paper | Area: General Machine Learning | Review:
This paper studies the problem of one-bit completion: the task is to recover a vector $y \in \\{0, 1\\}^n$ from the observation of a subset of its positive entries. The model assume that there exists a matrix $R \in \\{0, 1\\}^{m \times n}$ containing additional information about the model, e.g. $m$ previous pa... |
This paper proposes a new method to manage surprise signals in reinforcement learning. Surprises are stored in a memory for each episode and the memory is wiped after each episode. An autoencoder is used to perform readouts. The memory module can be plugged in existing surprise generators. Benchmark results show that t... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes a new method to manage surprise signals in reinforcement learning. Surprises are stored in a memory for each episode and the memory is wiped after each episode. An autoencoder is used to perform readouts. The memory module can be plugged in existing surprise generators. Benchmark results sho... |
The paper is considering the problem of learning the optimal transport maps with quadratic cost. In order to do so, they propose to apply a constrained optimization algorithm directly on the Monge formulation. In particular, the marginal constraint T#\mu = \nu is formulated as a penalty that involves a distance between... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The paper is considering the problem of learning the optimal transport maps with quadratic cost. In order to do so, they propose to apply a constrained optimization algorithm directly on the Monge formulation. In particular, the marginal constraint T#\mu = \nu is formulated as a penalty that involves a distance... |
This work proposes a formulation of the network motif mining problem as a machine learning task. In this task, the authors consider evaluating the top-K approximate motifs (according to some distance function) through an end-to-end learning process. The paper then proposes MotiFiesta, an architecture that leverages Edg... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
This work proposes a formulation of the network motif mining problem as a machine learning task. In this task, the authors consider evaluating the top-K approximate motifs (according to some distance function) through an end-to-end learning process. The paper then proposes MotiFiesta, an architecture that lever... |
This paper proposes a differentiable temporal logical rule method, TILP for temporal knowledge graph. The proposal can achieve similar performance as state-of-the-art baselines, while can find helpful logic patterns under some restricted scenario which no previous work can find.
## Strength
- This method proposes a di... | Recommendation: 5: marginally below the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
This paper proposes a differentiable temporal logical rule method, TILP for temporal knowledge graph. The proposal can achieve similar performance as state-of-the-art baselines, while can find helpful logic patterns under some restricted scenario which no previous work can find.
## Strength
- This method propo... |
This paper proposes an approach that performs top-k classification with label ranking. The proposed approach has three stages. The first stage trains the base classifier. The second stage relabels training data using one ground truth label and k-1 most likely labels. The third stage trains the top-k classification mode... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
This paper proposes an approach that performs top-k classification with label ranking. The proposed approach has three stages. The first stage trains the base classifier. The second stage relabels training data using one ground truth label and k-1 most likely labels. The third stage trains the top-k classificat... |
This paper proposed the first provably efficient offline RL algorithm for POMDPs with a confounded dataset, which achieves $n^{-1/2}$ under partial coverage assumption. Specifically, the author solves three key difficulties: 1) To tacke the confounder issues, the paper assumes the existence of the proxy variables from ... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposed the first provably efficient offline RL algorithm for POMDPs with a confounded dataset, which achieves $n^{-1/2}$ under partial coverage assumption. Specifically, the author solves three key difficulties: 1) To tacke the confounder issues, the paper assumes the existence of the proxy variabl... |
This paper proposes a novel privacy attack against FL systems, where the proposed attack extracts text sequences that contain targeted, privacy critical phrases. The targeted attack mechanism is achieved by maliciously modifying the model parameters on the server. In particular, the attack consists of three steps: tagg... | Recommendation: 6: marginally above the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper proposes a novel privacy attack against FL systems, where the proposed attack extracts text sequences that contain targeted, privacy critical phrases. The targeted attack mechanism is achieved by maliciously modifying the model parameters on the server. In particular, the attack consists of three ste... |
In this work, a deep learning approach for solving KS-DFT was proposed, which converts the objective function for KS-DFT into an unconstrained equivalent by reparameterizing the orthogonal constraints as a feed-forward computation. By using stochastic gradient descent, the integral was amortized over the optimization s... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
In this work, a deep learning approach for solving KS-DFT was proposed, which converts the objective function for KS-DFT into an unconstrained equivalent by reparameterizing the orthogonal constraints as a feed-forward computation. By using stochastic gradient descent, the integral was amortized over the optimi... |
Citing two limitations of Concept bottleneck model (CBM), (1) the time consuming and labor intensive need to collect labeled data for each of the predefined concepts and (2) their accuracy s often significantly lower than that of a standard neural network, the authors propose Label-free CBM; a framework to transform a... | Recommendation: 6: marginally above the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
Citing two limitations of Concept bottleneck model (CBM), (1) the time consuming and labor intensive need to collect labeled data for each of the predefined concepts and (2) their accuracy s often significantly lower than that of a standard neural network, the authors propose Label-free CBM; a framework to tra... |
This paper presents a new compartment based epidemic modelling for COVID which considers pre-symptomatic and symptomatic phases. The paper improves a prior model by Khan et al, 2021, and show empirically the new model better fits to a data from UAE.
It is a mathematical modelling paper, there is no representation or l... | Recommendation: 3: reject, not good enough | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper presents a new compartment based epidemic modelling for COVID which considers pre-symptomatic and symptomatic phases. The paper improves a prior model by Khan et al, 2021, and show empirically the new model better fits to a data from UAE.
It is a mathematical modelling paper, there is no representat... |
This paper proposes to improve the performance of energy-based models by incorporating latent variables to the model, assuming the positive samples of which come from an encoder optimized by contrastive representation learning. The latent variable EBM is parametrized in a way such that $p(x)$ and $p(z|x)$ can be easily... | Recommendation: 8: accept, good paper | Area: Generative models | Review:
This paper proposes to improve the performance of energy-based models by incorporating latent variables to the model, assuming the positive samples of which come from an encoder optimized by contrastive representation learning. The latent variable EBM is parametrized in a way such that $p(x)$ and $p(z|x)$ can b... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.