review stringlengths 5 16.6k | score stringclasses 5
values | area stringclasses 12
values | text stringlengths 31 5.65k |
|---|---|---|---|
This paper provides a classical method of approximating VQCs by sampling random frequencies from the VQC as a Fourier series. By showing that VQCs can be efficiently approximated using less than the order of exponential samplings, the work points out the potential problems with quantum advantages of VQCs.
Very good th... | Recommendation: 8: accept, good paper | Area: General Machine Learning | Review:
This paper provides a classical method of approximating VQCs by sampling random frequencies from the VQC as a Fourier series. By showing that VQCs can be efficiently approximated using less than the order of exponential samplings, the work points out the potential problems with quantum advantages of VQCs.
Very... |
The paper proposes a novel way of extracting silver-standard extractive summaries (Oracle) from gold-standard abstractive summaries which they call “OREO”. Earlier methods mostly relied on the Greedy labelling method, which is sub-optimal and deterministic. The authors propose to instead generate the final oracle summa... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper proposes a novel way of extracting silver-standard extractive summaries (Oracle) from gold-standard abstractive summaries which they call “OREO”. Earlier methods mostly relied on the Greedy labelling method, which is sub-optimal and deterministic. The authors propose to instead generate the final orac... |
ML models are known to latch to incidental correlations in the training data, which emphasises the need for controlling what they learn. This paper proposes a learning algorithm that extracts features for classification that can only be revealed from a specified region of pixels (explanation). For example, the explanat... | Recommendation: 5: marginally below the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
ML models are known to latch to incidental correlations in the training data, which emphasises the need for controlling what they learn. This paper proposes a learning algorithm that extracts features for classification that can only be revealed from a specified region of pixels (explanation). For example, the ... |
This paper studies the max margin bias that arise in the gradient flow training procedure of *quasi-homogeneous* neural networks. Quasi-homogeneous networks can be thought of as a generalization of homogeneous networks that are previously studied in the literature (e.g., Lyu & Li (2019) and Ji & Telgarsky (2020)), and ... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper studies the max margin bias that arise in the gradient flow training procedure of *quasi-homogeneous* neural networks. Quasi-homogeneous networks can be thought of as a generalization of homogeneous networks that are previously studied in the literature (e.g., Lyu & Li (2019) and Ji & Telgarsky (2020... |
This paper studies the problem of domain generalization. The authors propose to reduce the gradient conflict between the gradient biased toward the source domains and the unobservable gradient that could minimize risks in unseen domains. Specifically, the authors leverage the large-scale pre-trained model as a loose ap... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper studies the problem of domain generalization. The authors propose to reduce the gradient conflict between the gradient biased toward the source domains and the unobservable gradient that could minimize risks in unseen domains. Specifically, the authors leverage the large-scale pre-trained model as a ... |
The paper proposes a method to infer the state of a dynamical system based on sparse measurements. More specifically, the authors consider a situation in which there is a dynamical system whose state evolves but we cannot observe it directly. Instead we observe only a low dimensional approximation.. Here, the authors p... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper proposes a method to infer the state of a dynamical system based on sparse measurements. More specifically, the authors consider a situation in which there is a dynamical system whose state evolves but we cannot observe it directly. Instead we observe only a low dimensional approximation.. Here, the a... |
This paper proposes a multimodal transformer for molecular generation with multiple property optimization against multiple objectives, using language-based inputs. The authors describe prompt engineering based upon combinations of descriptive text and string representations of molecular properties and molecules as SELF... | Recommendation: 3: reject, not good enough | Area: Generative models | Review:
This paper proposes a multimodal transformer for molecular generation with multiple property optimization against multiple objectives, using language-based inputs. The authors describe prompt engineering based upon combinations of descriptive text and string representations of molecular properties and molecules... |
The paper proposes a graph neural network (GNN) for property prediction for crystals. A key issue in modeling crystals is to handle long range interactions. This is handled by using interatomic potentials as edge features in the GNN and approximating the infinite summation over all pairwise interactions using Ewald sum... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper proposes a graph neural network (GNN) for property prediction for crystals. A key issue in modeling crystals is to handle long range interactions. This is handled by using interatomic potentials as edge features in the GNN and approximating the infinite summation over all pairwise interactions using E... |
The authors consider the problem of making universal adversarial perturbations (UAPs - those that can be applied to any input and trigger classification for a certain class) under a set of transformations. The authors present a new algorithm. It involves applying SGD over examples (in the outer loop) and randomly chos... | Recommendation: 5: marginally below the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The authors consider the problem of making universal adversarial perturbations (UAPs - those that can be applied to any input and trigger classification for a certain class) under a set of transformations. The authors present a new algorithm. It involves applying SGD over examples (in the outer loop) and rando... |
## Updated Score. Changed from 3 to 5
This paper proposes to address the critical question of understanding the inductive bias of machine learning models, i.e., what kind of function/problem models are most suited based on their design choices or other parameters like learning rules, etc.
The authors choose to desi... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
## Updated Score. Changed from 3 to 5
This paper proposes to address the critical question of understanding the inductive bias of machine learning models, i.e., what kind of function/problem models are most suited based on their design choices or other parameters like learning rules, etc.
The authors choose... |
The authors introduce Simplicial Hopfield Networks — an architecture that extends the original Hopfield Network by allowing setwise connections in addition to pairwise connections between all neurons in the network. This reformulation allows higher memory capacity and performance and significantly increases the memory ... | Recommendation: 8: accept, good paper | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
The authors introduce Simplicial Hopfield Networks — an architecture that extends the original Hopfield Network by allowing setwise connections in addition to pairwise connections between all neurons in the network. This reformulation allows higher memory capacity and performance and significantly increases the... |
This paper presents a novel theoretical analysis of the Transformer networks, showing that with a special preprocessing, Transformer can universally approximate any polynomial function. The theoretical analysis is based on a simplified transformer which only has a single attention head and only handle fixed input size.... | Recommendation: 5: marginally below the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper presents a novel theoretical analysis of the Transformer networks, showing that with a special preprocessing, Transformer can universally approximate any polynomial function. The theoretical analysis is based on a simplified transformer which only has a single attention head and only handle fixed inp... |
In this study, authors proposed a self-consistent learning algorithm to generate data samples to improve downstream performance. The self-consistent learning was designed on generator-discriminator framework which is similar as GAN, but different from that of GAN, the generator and discriminator were cooperatively lear... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
In this study, authors proposed a self-consistent learning algorithm to generate data samples to improve downstream performance. The self-consistent learning was designed on generator-discriminator framework which is similar as GAN, but different from that of GAN, the generator and discriminator were cooperativ... |
This paper considers a finite set of attackers and a finite set of defenders and views them in a game-theoretic framework to choose the optimal choices for attacker and defender.
Specifically, it solves a min-max problem to find the optimal weights for attackers and defender.
Strength
- The paper mentions to many ... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper considers a finite set of attackers and a finite set of defenders and views them in a game-theoretic framework to choose the optimal choices for attacker and defender.
Specifically, it solves a min-max problem to find the optimal weights for attackers and defender.
Strength
- The paper mentions ... |
This work is an application paper that deployed the existing federated learning technologies namely hypernetwork to achieve personalized learning in reinforcement learning across multiple agents. The problem statement is based on a simplified hypothetical situation where multiple communities are powered by renewable en... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This work is an application paper that deployed the existing federated learning technologies namely hypernetwork to achieve personalized learning in reinforcement learning across multiple agents. The problem statement is based on a simplified hypothetical situation where multiple communities are powered by rene... |
The paper introduces an improvement of a learning method from samples and rules, named conceptual learning task. In particular the aim could be that of debug a movie recommendation.
- clarity
- experimental evaluation
+ incremental interesting approach
The notation adopted in the paper to describe the rules is difficu... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper introduces an improvement of a learning method from samples and rules, named conceptual learning task. In particular the aim could be that of debug a movie recommendation.
- clarity
- experimental evaluation
+ incremental interesting approach
The notation adopted in the paper to describe the rules is... |
The paper addresses the problem of collaborative MARL with language for specifying instructions and environmental dynamics, focusing on the problem of distributing subgoals among multiple agents. The approach proposed in the paper builds upon the grounding architecture from Zhong et al, 2019 and Hanjie et al. 2021, wit... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper addresses the problem of collaborative MARL with language for specifying instructions and environmental dynamics, focusing on the problem of distributing subgoals among multiple agents. The approach proposed in the paper builds upon the grounding architecture from Zhong et al, 2019 and Hanjie et al. 2... |
In this paper, an image autoencoder is trained with two loss functions: 1. Image reconstruction loss and 2. Poisson loss to optimize representation similarity between artificial and biological neurons.
The authors show that by training in the above way:
1. Image reconstruction is better than the case without Poisson ... | Recommendation: 5: marginally below the acceptance threshold | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
In this paper, an image autoencoder is trained with two loss functions: 1. Image reconstruction loss and 2. Poisson loss to optimize representation similarity between artificial and biological neurons.
The authors show that by training in the above way:
1. Image reconstruction is better than the case without ... |
In this paper, the authors propose an algorithm for single-image depth prediction. The key contribution of the algorithm is the V-layer, which utilizes the depth gradient to make a more accurate depth prediction. It takes the gradients in the two-axis directions and their reliability as input and obtains the depth in a... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
In this paper, the authors propose an algorithm for single-image depth prediction. The key contribution of the algorithm is the V-layer, which utilizes the depth gradient to make a more accurate depth prediction. It takes the gradients in the two-axis directions and their reliability as input and obtains the de... |
This paper proposes a robust RL policy training framework which applies adversarial training using coordinated traffic flow. For building the traffic flow, it uses the so-called “Social Value Orientations (SVOs)” (which is a weight that balances the proportion of each agent’s own driving reward and its surrounding agen... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes a robust RL policy training framework which applies adversarial training using coordinated traffic flow. For building the traffic flow, it uses the so-called “Social Value Orientations (SVOs)” (which is a weight that balances the proportion of each agent’s own driving reward and its surround... |
The paper presents a CNN/transformer based variational autoencoder for generation of molecular 3d structure from molecular graphs.
Strengths:
The paper deals with an important problem and presents a working solution that is fairly simple and effective.
Weaknesses:
The novelty is limited - the main technical contribut... | Recommendation: 5: marginally below the acceptance threshold | Area: Generative models | Review:
The paper presents a CNN/transformer based variational autoencoder for generation of molecular 3d structure from molecular graphs.
Strengths:
The paper deals with an important problem and presents a working solution that is fairly simple and effective.
Weaknesses:
The novelty is limited - the main technical c... |
This paper analyzes the training dynamics of gradient descent based algorithms using a particular orthonormal basis for the function space spanned by the (function) iterates during training. Convergence to the function in the model space that minimizes the expected l2 loss is shown to occur at a novel convergence rate ... | Recommendation: 6: marginally above the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper analyzes the training dynamics of gradient descent based algorithms using a particular orthonormal basis for the function space spanned by the (function) iterates during training. Convergence to the function in the model space that minimizes the expected l2 loss is shown to occur at a novel convergen... |
The paper proposes RulE, a model for link prediction on knowledge graphs that jointly learns representations for entities, relations and rules in the embedding space. More specifically, RulE applies on top of the RotatE KGE model, and additionally includes a component for learning compositional rule embeddings: Given a... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The paper proposes RulE, a model for link prediction on knowledge graphs that jointly learns representations for entities, relations and rules in the embedding space. More specifically, RulE applies on top of the RotatE KGE model, and additionally includes a component for learning compositional rule embeddings:... |
This paper studies the portfolio optimization problem with factor learning, where the authors use deep reinforcement learning models to learn the factors, and combined that with the classic continuous time finance factor model. The deep RL component is based on the deep deterministic policy learning (DDPG) model from t... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper studies the portfolio optimization problem with factor learning, where the authors use deep reinforcement learning models to learn the factors, and combined that with the classic continuous time finance factor model. The deep RL component is based on the deep deterministic policy learning (DDPG) mode... |
The paper considers the offline-to-online RL setting and proposes policy expansion (PEX) by constructing the stochastic online policy: a mixture of the offline policy (trained using the offline data) and the online policy (trained with the online rollouts). In the experiments, the paper shows that PEX has some performa... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper considers the offline-to-online RL setting and proposes policy expansion (PEX) by constructing the stochastic online policy: a mixture of the offline policy (trained using the offline data) and the online policy (trained with the online rollouts). In the experiments, the paper shows that PEX has some ... |
This paper proposes a sentence encoder, SPE, to improve the quality of adversarial examples in terms of minimizing semantics change. Specifically, the authors train multiple text classifiers and average the output embeddings from these classifiers to get the final representation of an input sentence. The paper also pro... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a sentence encoder, SPE, to improve the quality of adversarial examples in terms of minimizing semantics change. Specifically, the authors train multiple text classifiers and average the output embeddings from these classifiers to get the final representation of an input sentence. The paper ... |
Basically, this paper aims to study the problem of evaluating the counterfactual statements based on neural models. To achieve this goal, the authors firstly detail the previous work on the causal neural models, and then propose to identify counterfactual quantities based on the first two causal layers. The authors hav... | Recommendation: 8: accept, good paper | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
Basically, this paper aims to study the problem of evaluating the counterfactual statements based on neural models. To achieve this goal, the authors firstly detail the previous work on the causal neural models, and then propose to identify counterfactual quantities based on the first two causal layers. The aut... |
This paper provides an algorithm for learning the rationalizable action profiles for multi-agent games with bandit feedback, based on which another algorithm for learning the approximate rationalizable CE and CCE. Both algorithms are computationally efficient and have been proven to improve significantly compared to pr... | Recommendation: 8: accept, good paper | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper provides an algorithm for learning the rationalizable action profiles for multi-agent games with bandit feedback, based on which another algorithm for learning the approximate rationalizable CE and CCE. Both algorithms are computationally efficient and have been proven to improve significantly compar... |
The paper proposes to use heuristic method to discover association rules between variables in an optimization problem.
+ An interesting idea compared to the state of the art works
- Poor results
The introduction of the problem is unclear. For instance in the introduction in not clear the the meaning of samples in an o... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
The paper proposes to use heuristic method to discover association rules between variables in an optimization problem.
+ An interesting idea compared to the state of the art works
- Poor results
The introduction of the problem is unclear. For instance in the introduction in not clear the the meaning of samples... |
This paper proposes a multi-task learning method to reduce conflict gradients during multi-task optimisation. The proposed method, name Recon is very simple in design and intuitive – it first computes the cosine similarity score for all combinations of task pairs in each shared layer, and sets this shared layer to be t... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
This paper proposes a multi-task learning method to reduce conflict gradients during multi-task optimisation. The proposed method, name Recon is very simple in design and intuitive – it first computes the cosine similarity score for all combinations of task pairs in each shared layer, and sets this shared layer... |
The paper considers the problem of hyper-parameter optimization for reinforcement learning algorithms. A Bayesian optimization based algorithm is proposed to tackle the problem where the key idea is to model the reward curve of a candidate hyper-parameter configuration with a generalized logistic function. Gaussian Pro... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper considers the problem of hyper-parameter optimization for reinforcement learning algorithms. A Bayesian optimization based algorithm is proposed to tackle the problem where the key idea is to model the reward curve of a candidate hyper-parameter configuration with a generalized logistic function. Gaus... |
The paper proposes prototypical calibration for zero-shot and few-shot classification tasks when evaluating large language models. The approach first uses Gaussian mixture distribution to estimate the prototypical clusters for all categories of the classification task and then assigns each cluster to the corresponding ... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper proposes prototypical calibration for zero-shot and few-shot classification tasks when evaluating large language models. The approach first uses Gaussian mixture distribution to estimate the prototypical clusters for all categories of the classification task and then assigns each cluster to the corres... |
Following the well-established Sliced-Wasserstein (SW) used as a Wasserstein-based metric to compare measures supported on the Euclidean space $\mathbb{R}^d$, this paper introduces a natural Spherical-Sliced-Wasserstein (SSW) distance to compare measures supported on an hypersphere.
The construction technique follows... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
Following the well-established Sliced-Wasserstein (SW) used as a Wasserstein-based metric to compare measures supported on the Euclidean space $\mathbb{R}^d$, this paper introduces a natural Spherical-Sliced-Wasserstein (SSW) distance to compare measures supported on an hypersphere.
The construction technique... |
This paper argues that effective exploration strategies can improve generalization in contextual MDPs (CMDPs). The paper proposes a method called exploration via distributional ensemble (EDE), which uses a deep ensemble to (1) disentangle epistemic uncertainty from aleatoric uncertainty and (2) encourage exploration by... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper argues that effective exploration strategies can improve generalization in contextual MDPs (CMDPs). The paper proposes a method called exploration via distributional ensemble (EDE), which uses a deep ensemble to (1) disentangle epistemic uncertainty from aleatoric uncertainty and (2) encourage explor... |
This paper provides convergence bounds for score-based models under the assumption that score estimate is $L^2$ accurate. Provided that, the paper derives some remarkable bounds for the score-based sampling methods under pretty weak assumptions.
Strength: The results are pretty impressive given the state of the field.
... | Recommendation: 8: accept, good paper | Area: Generative models | Review:
This paper provides convergence bounds for score-based models under the assumption that score estimate is $L^2$ accurate. Provided that, the paper derives some remarkable bounds for the score-based sampling methods under pretty weak assumptions.
Strength: The results are pretty impressive given the state of the... |
The paper proposes a novel cost function for representation learning/stochastic embedding of data. Specifically, they adapt the information bottleneck functional, which aims at maximizing $I(Y;Z)-\beta I(X;Z)$ by removing the term corresponding to the entropy of $Z$. Essentially, the authors result at maximizing $I(Y;Z... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The paper proposes a novel cost function for representation learning/stochastic embedding of data. Specifically, they adapt the information bottleneck functional, which aims at maximizing $I(Y;Z)-\beta I(X;Z)$ by removing the term corresponding to the entropy of $Z$. Essentially, the authors result at maximizin... |
The authors present a retrieval-based LM that can generate phrases. Their base model is a standard transformer, and the retrieval component uses BERT. Their model is trained using in-batch positives / negatives. Retrieval is actually done at document-level, then phrase extracted using a segmentation algorithm. For eval... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The authors present a retrieval-based LM that can generate phrases. Their base model is a standard transformer, and the retrieval component uses BERT. Their model is trained using in-batch positives / negatives. Retrieval is actually done at document-level, then phrase extracted using a segmentation algorithm. ... |
The authors tackle a practical problem of federated learning where asynchronous communication is frequent due to the computational heterogeneity. The authors propose a method considering taleness of local updates when performing model aggregation. They demonstrate the effectiveness of their method.
**Strength**
- They... | Recommendation: 3: reject, not good enough | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The authors tackle a practical problem of federated learning where asynchronous communication is frequent due to the computational heterogeneity. The authors propose a method considering taleness of local updates when performing model aggregation. They demonstrate the effectiveness of their method.
**Strength*... |
This paper presents a hybrid voxel-grid- and MLP-based neural surface reconstruction method called Voxurf, that is both fast and accurate. It uses the volumetric representation of a voxel grid to expedite training, while proposing a two-stage training process, a novel network that captures the relationship between colo... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper presents a hybrid voxel-grid- and MLP-based neural surface reconstruction method called Voxurf, that is both fast and accurate. It uses the volumetric representation of a voxel grid to expedite training, while proposing a two-stage training process, a novel network that captures the relationship betw... |
It is well known that the performance of machine learning models is highly dependent on the distribution of the data on which it is evaluated: model performance deteriorates when tested on data generated from a distribution shifted with respect to the training data generating process. Identifying and mitigating the eff... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
It is well known that the performance of machine learning models is highly dependent on the distribution of the data on which it is evaluated: model performance deteriorates when tested on data generated from a distribution shifted with respect to the training data generating process. Identifying and mitigating... |
It is well known that the function represented by a ReLU network is piecewise linear, with each linear region corresponding to an activation pattern. There have been papers showing that despite the potential number of possible activation patterns being exponential in the number of neurons, the actual number for a typic... | Recommendation: 3: reject, not good enough | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
It is well known that the function represented by a ReLU network is piecewise linear, with each linear region corresponding to an activation pattern. There have been papers showing that despite the potential number of possible activation patterns being exponential in the number of neurons, the actual number for... |
The method studies the generalization capabilities of unsupervised RL empirically, and develops a hybrid planner with strong asymptotic performance and high sample efficiency in standard benchmarks.
Strength
* This work offers a large-scale benchmarks for different URL techniques and compare the usefulness of differen... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The method studies the generalization capabilities of unsupervised RL empirically, and develops a hybrid planner with strong asymptotic performance and high sample efficiency in standard benchmarks.
Strength
* This work offers a large-scale benchmarks for different URL techniques and compare the usefulness of ... |
This paper looks at a critical but rarely studied problem: the certification of robustness to Universal Perturbations (UPs). In particular, UPs (e.g., universal adversarial noise, backdoor, or neural trojan attacks) have become alarming lines of threats that make it hard to use Machine Learning as a service safely and ... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper looks at a critical but rarely studied problem: the certification of robustness to Universal Perturbations (UPs). In particular, UPs (e.g., universal adversarial noise, backdoor, or neural trojan attacks) have become alarming lines of threats that make it hard to use Machine Learning as a service saf... |
The paper proposes several techniques to benefit the pre-training of MAE. However, it seems like an ensemble of different tricks. The contributions should be further highlighted.
Weaknesses:
(1) The writing of the paper is a little messy. The authors should highlight their contributions clearly in the Introduction. ... | Recommendation: 1: strong reject | Area: Deep Learning and representational learning | Review:
The paper proposes several techniques to benefit the pre-training of MAE. However, it seems like an ensemble of different tricks. The contributions should be further highlighted.
Weaknesses:
(1) The writing of the paper is a little messy. The authors should highlight their contributions clearly in the Introd... |
The authors propose Comprehensive Attention Benchmark (CAB) under a fine-grained attention taxonomy with four distinguishable attention patterns, namely, noncausal self, causal self, noncausal cross, and causal cross attentions. The benchmark collects seven real-world tasks from diverse fields of computer vision, natu... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The authors propose Comprehensive Attention Benchmark (CAB) under a fine-grained attention taxonomy with four distinguishable attention patterns, namely, noncausal self, causal self, noncausal cross, and causal cross attentions. The benchmark collects seven real-world tasks from diverse fields of computer visi... |
This paper introduces a method for semi-supervised learning in the offline RL setting where the unlabelled part of the dataset consists of action-free state trajectories and the labelled part consists of the full trajectories. They use an inverse dynamics model to learn actions that give rise to state transitions and u... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper introduces a method for semi-supervised learning in the offline RL setting where the unlabelled part of the dataset consists of action-free state trajectories and the labelled part consists of the full trajectories. They use an inverse dynamics model to learn actions that give rise to state transitio... |
This paper develops a new optimization method for domain generalization. The idea of the proposed PGrad approach is to use a robust gradient direction for parameter update. The robust direction is estimated in 2 steps. First, the rollout of the weights trajectory is collected by sequentially applying training updates i... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper develops a new optimization method for domain generalization. The idea of the proposed PGrad approach is to use a robust gradient direction for parameter update. The robust direction is estimated in 2 steps. First, the rollout of the weights trajectory is collected by sequentially applying training u... |
This paper explores the task of dexterous manipulation with human demonstrations. The paper introduces an interesting set of simulated benchmarking tasks, along with a baseline approach to solve these. The benchmark builds on prior work (Plasticinelab), using a simulated shadow robot hand. The proposed baseline relies ... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper explores the task of dexterous manipulation with human demonstrations. The paper introduces an interesting set of simulated benchmarking tasks, along with a baseline approach to solve these. The benchmark builds on prior work (Plasticinelab), using a simulated shadow robot hand. The proposed baseline... |
This work proposes to train continuous normalizing flows by specifying a continuum of distributions using an interpolation function. The neural network model is trained to estimate the expected velocity of the interpolation function though MSE minimization. This training approach does not require numerical simulation o... | Recommendation: 3: reject, not good enough | Area: Generative models | Review:
This work proposes to train continuous normalizing flows by specifying a continuum of distributions using an interpolation function. The neural network model is trained to estimate the expected velocity of the interpolation function though MSE minimization. This training approach does not require numerical simu... |
The paper addresses a crucial problem in computer vision: Video Highlights Detection (VHD) which aims to target the appealing domains for the given video(s). More general approaches are based on world assumptions which lead to poor scalability and to address this issue the author proposes a new method called Global Pro... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper addresses a crucial problem in computer vision: Video Highlights Detection (VHD) which aims to target the appealing domains for the given video(s). More general approaches are based on world assumptions which lead to poor scalability and to address this issue the author proposes a new method called Gl... |
The work extends the contrastive learning based nonlinear mixture identification framework in (von Kugelgen et al. 2021) and considers a case where the two modalities have two different generative functions. The major contribution is the models’ identifiability analysis. The setting and the analysis can be considered a... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
The work extends the contrastive learning based nonlinear mixture identification framework in (von Kugelgen et al. 2021) and considers a case where the two modalities have two different generative functions. The major contribution is the models’ identifiability analysis. The setting and the analysis can be cons... |
["Long" summary]
The paper deals with sample efficiency in deep reinforcement learning (DRL). In particular, there was a known problem in DRL with high update-to-data (UTD) ratio, addressed by several recent studies with different regularizers. The authors aim to understand why DRL usually fail with high UTD and how w... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
["Long" summary]
The paper deals with sample efficiency in deep reinforcement learning (DRL). In particular, there was a known problem in DRL with high update-to-data (UTD) ratio, addressed by several recent studies with different regularizers. The authors aim to understand why DRL usually fail with high UTD a... |
This paper studies the effect of adversarial training on the action-value estimates of value-based deep RL agents. The paper considers both a toy learning problem where the optimal function approximator with and without adversarial regularization can be computed analytically, as well as networks trained on the arcade l... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
This paper studies the effect of adversarial training on the action-value estimates of value-based deep RL agents. The paper considers both a toy learning problem where the optimal function approximator with and without adversarial regularization can be computed analytically, as well as networks trained on the ... |
This paper proposes a control algorithm to explore transition paths between reactant and product basins in molecular dynamics simulations. The work is strictly based on the theory of Knappen and Ruiz (2016), and implements it for molecular systems.
The loss fucntion that is optimized is a combination of two:
Distanc... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper proposes a control algorithm to explore transition paths between reactant and product basins in molecular dynamics simulations. The work is strictly based on the theory of Knappen and Ruiz (2016), and implements it for molecular systems.
The loss fucntion that is optimized is a combination of two:
... |
The paper proposes using sequence-to-sequence translation from NLP as the core of a Multiple Sequence Alignment. They train their model on phylogenetic simulators and compare against standard alignment methods.
Strengths:
- The paper likely improves the SOTA for some applications of multiple sequence alignments.
- The... | 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 using sequence-to-sequence translation from NLP as the core of a Multiple Sequence Alignment. They train their model on phylogenetic simulators and compare against standard alignment methods.
Strengths:
- The paper likely improves the SOTA for some applications of multiple sequence alignment... |
This paper combines graph neural networks with point cloud information to improve the performance of the popular graph network simulators. Standard graph network simulators accumulate prediction errors, causing predicted dynamic states to drift over time. The key idea in this paper is using point cloud information when... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper combines graph neural networks with point cloud information to improve the performance of the popular graph network simulators. Standard graph network simulators accumulate prediction errors, causing predicted dynamic states to drift over time. The key idea in this paper is using point cloud informat... |
The paper explores the problem of mechanism design which studies how to design reward functions and environmental rules defining mathematical games. The applications of mechanism design spans across many domains from optimizing social welfare with economic policies to designing governmental policies. The conventional p... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper explores the problem of mechanism design which studies how to design reward functions and environmental rules defining mathematical games. The applications of mechanism design spans across many domains from optimizing social welfare with economic policies to designing governmental policies. The conven... |
This work proposes to parametrize meta-RL policies with a symbolic policy in the hopes of improving generalization, interpretability, and efficiency.
This work has two main weaknesses. First, the clarity of the writing and presentation could be significantly improved, and is currently challenging to understand. I elabo... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This work proposes to parametrize meta-RL policies with a symbolic policy in the hopes of improving generalization, interpretability, and efficiency.
This work has two main weaknesses. First, the clarity of the writing and presentation could be significantly improved, and is currently challenging to understand.... |
This paper studies the problem of building a coreset for fitting rational functions to time series data. In particular, suppose $y_1, ..., y_n \in\mathbb{R}$ is a time series. Then, in rational function fitting, we want to find a rational function $r$ of degree $k$ such that $r(i) \approx y_i$. We formally measure erro... | Recommendation: 3: reject, not good enough | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper studies the problem of building a coreset for fitting rational functions to time series data. In particular, suppose $y_1, ..., y_n \in\mathbb{R}$ is a time series. Then, in rational function fitting, we want to find a rational function $r$ of degree $k$ such that $r(i) \approx y_i$. We formally meas... |
This work aims to address the robust overfitting issue in adversarial training by using data augmentations. Prior works have shown that adversarial training does not benefit from augmentations such as AutoAugment. The authors firstly study the role of hardness and diversity of augmentations in robustness and accuracy, ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This work aims to address the robust overfitting issue in adversarial training by using data augmentations. Prior works have shown that adversarial training does not benefit from augmentations such as AutoAugment. The authors firstly study the role of hardness and diversity of augmentations in robustness and ac... |
This paper introduces the approach of incorporating an uncertainty estimator into the PPO model. The authors define the uncertainty of RL models by the standard deviation of logits from neural models and propose capturing these uncertainties by utilizing a previously proposed model named Masksembles. To evaluate the mo... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper introduces the approach of incorporating an uncertainty estimator into the PPO model. The authors define the uncertainty of RL models by the standard deviation of logits from neural models and propose capturing these uncertainties by utilizing a previously proposed model named Masksembles. To evaluat... |
The paper proposes ASIF, which can retrieve relevant captions given an image **without training**, utilizing 1) two pre-trained unimodal encoders and 2) a relatively small multimodal datasets. The key intuitive is that “captions of similar images should be themselves similar”; thus the retrieval is done through the rel... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes ASIF, which can retrieve relevant captions given an image **without training**, utilizing 1) two pre-trained unimodal encoders and 2) a relatively small multimodal datasets. The key intuitive is that “captions of similar images should be themselves similar”; thus the retrieval is done through... |
This paper focuses on robust representation learning and attempts to propose a method that is robust to input perturbations. To achieve this goal, the authors propose Noise Injection Node Regularization (NINR) which severs as a “regularizer”. Experimental results demonstrate in some cases, this proposed method outperfo... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper focuses on robust representation learning and attempts to propose a method that is robust to input perturbations. To achieve this goal, the authors propose Noise Injection Node Regularization (NINR) which severs as a “regularizer”. Experimental results demonstrate in some cases, this proposed method ... |
This paper proposes to meta-learn a network capable of embedding the mixed-relevance target dataset and dynamically adapting inference for target examples using cross-attention. It further analysis the method performance in the domain-supervised adaption.
Strength:
the writing of this papers seems easy to follow and ... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper proposes to meta-learn a network capable of embedding the mixed-relevance target dataset and dynamically adapting inference for target examples using cross-attention. It further analysis the method performance in the domain-supervised adaption.
Strength:
the writing of this papers seems easy to fol... |
Motivated by the fact that architecture parameters in one-shot differentiable NAS are not indicative of the true operator importance, the authors integrate the Lite-Transformer into the one-shot model, and train it jointly with the model, to predict the ranking of operators at each edge / layer.
The operations are pa... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
Motivated by the fact that architecture parameters in one-shot differentiable NAS are not indicative of the true operator importance, the authors integrate the Lite-Transformer into the one-shot model, and train it jointly with the model, to predict the ranking of operators at each edge / layer.
The operation... |
This paper explores the problem of learning agent policies from action limited datasets and proposes ALPT, which pretrains an inverse dynamics model (IDM) on multiple environments to provide accurate action labels for a decision transformer (DT) agent on an action limited target environment dataset. The experiments and... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper explores the problem of learning agent policies from action limited datasets and proposes ALPT, which pretrains an inverse dynamics model (IDM) on multiple environments to provide accurate action labels for a decision transformer (DT) agent on an action limited target environment dataset. The experim... |
Local SGD (LSGD) is a communication-efficient variant of SGD for large-scale training, where multiple GPUs perform SGD independently and average the model parameters periodically. It has been shown previously that LSGD generalizes better when with a small enough learning rate and sufficient training time. To analyze th... | Recommendation: 8: accept, good paper | Area: Optimization (eg, convex and non-convex optimization) | Review:
Local SGD (LSGD) is a communication-efficient variant of SGD for large-scale training, where multiple GPUs perform SGD independently and average the model parameters periodically. It has been shown previously that LSGD generalizes better when with a small enough learning rate and sufficient training time. To an... |
The method addresses the problem of asymmetric image retrieval, in which different models are deployed for query and gallery - usally for computation efficiency reasons - and that consists on aligning their embedding spaces. The authors argue that current approaches, which try to enforce the consitency of features betw... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The method addresses the problem of asymmetric image retrieval, in which different models are deployed for query and gallery - usally for computation efficiency reasons - and that consists on aligning their embedding spaces. The authors argue that current approaches, which try to enforce the consitency of featu... |
This paper investigated a proper bit-precision for a block floating-point format for deep neural network training and revealed that training with a small number of mantissa bits can be compensated if the training iterations with a large mantissa bit are followed. The authors also investigate the interplay between the b... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
This paper investigated a proper bit-precision for a block floating-point format for deep neural network training and revealed that training with a small number of mantissa bits can be compensated if the training iterations with a large mantissa bit are followed. The authors also investigate the interplay betwe... |
This work proposed a novel learning problem for binary classifications, namely Confidence-Difference (ConfDiff) classification, in which the learner is only given unlabeled data pairs $x,x'$ equipped with confidence difference specifying the difference in the probabilities of being positive $P(y'=1|x') - P(y=1|x)$. The... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
This work proposed a novel learning problem for binary classifications, namely Confidence-Difference (ConfDiff) classification, in which the learner is only given unlabeled data pairs $x,x'$ equipped with confidence difference specifying the difference in the probabilities of being positive $P(y'=1|x') - P(y=1|... |
The paper introduces a new method to rank feature importance in a neural network. The authors propose to train multiple shallow networks and take the average of the weights for the final ranking. The proposed method is evaluated in several datasets to prove its accuracy and stability.
Strength
- The paper motivates th... | 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 introduces a new method to rank feature importance in a neural network. The authors propose to train multiple shallow networks and take the average of the weights for the final ranking. The proposed method is evaluated in several datasets to prove its accuracy and stability.
Strength
- The paper moti... |
The paper aims to understand the recently discovered "Edge of Stability" phenomena in a second-order regression model. The authors consider second-order NTK models and show for simple settings, under proper conditions, one can observe progressive sharpening close to initialization. Finally, the authors simulate the si... | Recommendation: 5: marginally below the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The paper aims to understand the recently discovered "Edge of Stability" phenomena in a second-order regression model. The authors consider second-order NTK models and show for simple settings, under proper conditions, one can observe progressive sharpening close to initialization. Finally, the authors simulat... |
The authors create a novel approach for symbolic regression using two different mechanisms. The first is a math language model, and the second is an adaptable strategy for alternating fitness functions during evolution. They experiment with their approach in synthetic and real-world data and achieve state-of-the-art re... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
The authors create a novel approach for symbolic regression using two different mechanisms. The first is a math language model, and the second is an adaptable strategy for alternating fitness functions during evolution. They experiment with their approach in synthetic and real-world data and achieve state-of-th... |
This paper proposed an asynchronous method for distributed bilevel optimization. The proposed method can tackle both nonconvex upper-level and lower-level objective functions. Convergence analysis has been provided.
Strengths:
1. The proposed method is clearly explained and theoretically analyzed.
2. Experiments have ... | Recommendation: 5: marginally below the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper proposed an asynchronous method for distributed bilevel optimization. The proposed method can tackle both nonconvex upper-level and lower-level objective functions. Convergence analysis has been provided.
Strengths:
1. The proposed method is clearly explained and theoretically analyzed.
2. Experimen... |
The paper presents a new problem, named Sequential Model Editing, that requires a model editor to make a sequence of error corrections while keeping all the previous edits, preserving model performance, as well as generalizing to equivalent inputs. Instead of modifying the model parameters, a set of new parameters (ter... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper presents a new problem, named Sequential Model Editing, that requires a model editor to make a sequence of error corrections while keeping all the previous edits, preserving model performance, as well as generalizing to equivalent inputs. Instead of modifying the model parameters, a set of new paramet... |
Three types of biologically-plausible learning algorithms --- predictive coding, contrastive hebbian learning, and equilibrium propagation --- are reformulated as energy based models. This framework is used to provide a unifying description for when these algorithms closely approximate gradient descent (backprop), and ... | Recommendation: 6: marginally above the acceptance threshold | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
Three types of biologically-plausible learning algorithms --- predictive coding, contrastive hebbian learning, and equilibrium propagation --- are reformulated as energy based models. This framework is used to provide a unifying description for when these algorithms closely approximate gradient descent (backpro... |
This work proposed a novel zero-shot framework for linear image restoration, and the key module is denoising diffusion null-space model. It only requires a pretrained diffusion model, without further training and optimization. Experimental results show that the proposed method outperforms sota zero-shot restoration met... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This work proposed a novel zero-shot framework for linear image restoration, and the key module is denoising diffusion null-space model. It only requires a pretrained diffusion model, without further training and optimization. Experimental results show that the proposed method outperforms sota zero-shot restora... |
The paper discusses information theory-based clustering and self-supervised learning.
The main contribution lies in the use of "reversed" KL divergence and cross-entropy.
The paper is a bit confusing, so I present here my understanding of the theory.
Clustering can be done by maximizing the Mutual Information over t... | Recommendation: 3: reject, not good enough | Area: Unsupervised and Self-supervised learning | Review:
The paper discusses information theory-based clustering and self-supervised learning.
The main contribution lies in the use of "reversed" KL divergence and cross-entropy.
The paper is a bit confusing, so I present here my understanding of the theory.
Clustering can be done by maximizing the Mutual Informatio... |
The paper proposes a new sparse attention mechanism that uses decision trees to cluster the queries and keys such that the queries only attend to the keys that fall into the same leaf. To mitigate the collapsing problem (all keys and queries are allocated to one leaf node) in the proposed TF-Attention (Tree Fine-graine... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes a new sparse attention mechanism that uses decision trees to cluster the queries and keys such that the queries only attend to the keys that fall into the same leaf. To mitigate the collapsing problem (all keys and queries are allocated to one leaf node) in the proposed TF-Attention (Tree Fin... |
The authors introduce a method for local sampling in diffusion models. This is achieved by leveraging the stochasticity of the diffusion models. An input image is propagated through the forward process for t steps then through the reverse process for t steps. In the forward process each step produces an output that is ... | Recommendation: 8: accept, good paper | Area: Generative models | Review:
The authors introduce a method for local sampling in diffusion models. This is achieved by leveraging the stochasticity of the diffusion models. An input image is propagated through the forward process for t steps then through the reverse process for t steps. In the forward process each step produces an output ... |
The paper proposes a self-supervised learning method for tabular data. Specifically, it adopts a pipeline similar to DINO – it matches the categorical distribution produced from an EMA teacher network and a student network but the classifier is built with a queue of past embeddings instead of a learned linear layer. Th... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
The paper proposes a self-supervised learning method for tabular data. Specifically, it adopts a pipeline similar to DINO – it matches the categorical distribution produced from an EMA teacher network and a student network but the classifier is built with a queue of past embeddings instead of a learned linear l... |
### Problem
The paper tackles the problem of prompt tuning for large scale vision-language models. This is a very recent direction.
### Proposed method
The method is built on top of (i) zero-shot CLIP with hand-crafted prompts and (ii) CoOP which proposed to fine-tune context vectors as tunable prompts. The method pro... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
### Problem
The paper tackles the problem of prompt tuning for large scale vision-language models. This is a very recent direction.
### Proposed method
The method is built on top of (i) zero-shot CLIP with hand-crafted prompts and (ii) CoOP which proposed to fine-tune context vectors as tunable prompts. The me... |
This paper considers shallow neural networks trained with gradient methods (stochastic gradient descent or gradient flow) under logistic or exponential loss. The authors prove that a low test error can be achieved in a regime that still exhibits low rotations in the weights, but is capable of exiting the NTK regime (in... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper considers shallow neural networks trained with gradient methods (stochastic gradient descent or gradient flow) under logistic or exponential loss. The authors prove that a low test error can be achieved in a regime that still exhibits low rotations in the weights, but is capable of exiting the NTK re... |
This paper proposes an unsupervised domain adaptation model for time series combining nearest-neighbourhood contrastive learning and adversarial learning.
S
+ extensive experimentation and ablation tests
+ strong results against baselines
W
+ lack of comparisons with already cited works (NCL)
The paper proposes a c... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes an unsupervised domain adaptation model for time series combining nearest-neighbourhood contrastive learning and adversarial learning.
S
+ extensive experimentation and ablation tests
+ strong results against baselines
W
+ lack of comparisons with already cited works (NCL)
The paper prop... |
This paper studies accuracy estimation from unlabeled data, in the presence of distribution shifts. They show that "dispersity" that captures the marginal distribution of the predicted labels correlates well with the accuracy. The paper assumes that the true label distribution is uniform and proposes a measure of dispe... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper studies accuracy estimation from unlabeled data, in the presence of distribution shifts. They show that "dispersity" that captures the marginal distribution of the predicted labels correlates well with the accuracy. The paper assumes that the true label distribution is uniform and proposes a measure ... |
This paper proposes a method to estimate the MSE (and PSNR) in image denoising using only noisy images (no reference). The main idea is to use three noisy images to produce an unbiased estimator of the MSE. The method is sound and well motivated. Several results on synthetic cases and a real example on denoising Electr... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper proposes a method to estimate the MSE (and PSNR) in image denoising using only noisy images (no reference). The main idea is to use three noisy images to produce an unbiased estimator of the MSE. The method is sound and well motivated. Several results on synthetic cases and a real example on denoisin... |
This paper presents a model for Time Series Classification (TSC).
The model is made of two parts: first an encoder is trained to extract interpretable features using a VQ-VAE strategy and second, a classification head builds n-grams (in practice unigrams and bigrams) and performs regularized logistic regression on the ... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper presents a model for Time Series Classification (TSC).
The model is made of two parts: first an encoder is trained to extract interpretable features using a VQ-VAE strategy and second, a classification head builds n-grams (in practice unigrams and bigrams) and performs regularized logistic regression... |
This paper proposes a novel, simple and intuitive Aggregation Separation Loss (ASLoss), which aggregates the representations of the same class samples as near as possible and separates the representations of different classes as far as possible. The authors conduct extensive experiments on diffirent scenarios i.e. data... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a novel, simple and intuitive Aggregation Separation Loss (ASLoss), which aggregates the representations of the same class samples as near as possible and separates the representations of different classes as far as possible. The authors conduct extensive experiments on diffirent scenarios i... |
The paper reports a set of experimental results based on protein sequence masked language modeling using a convolutional architecture, which they name "CARP", comparing to ESM-1b, a previously published transformer based model. The authors show that
* CARP performs similarly to ESM-1b in terms of MLM loss, for differ... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper reports a set of experimental results based on protein sequence masked language modeling using a convolutional architecture, which they name "CARP", comparing to ESM-1b, a previously published transformer based model. The authors show that
* CARP performs similarly to ESM-1b in terms of MLM loss, fo... |
The author(s) study the model aggregation strategy in federated learning. The author(s) conduct a theoretial analysis on the impact of the model aggregation strategy. In particular, the author(s) show that the model aggregation can lead to a trade-off between convergence rate and convergence error (a non-vanishing bias... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The author(s) study the model aggregation strategy in federated learning. The author(s) conduct a theoretial analysis on the impact of the model aggregation strategy. In particular, the author(s) show that the model aggregation can lead to a trade-off between convergence rate and convergence error (a non-vanish... |
This paper studies systematic generalization and the extent to which modular networks are better at learning systematic mappings relative to non-systematic mappings. Leveraging prior results on deep linear networks, the authors demonstrate that while non-modular networks learn a combined systematic and non-systematic m... | Recommendation: 8: accept, good paper | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper studies systematic generalization and the extent to which modular networks are better at learning systematic mappings relative to non-systematic mappings. Leveraging prior results on deep linear networks, the authors demonstrate that while non-modular networks learn a combined systematic and non-syst... |
This paper addresses the decoder improvements in Chinese Word Segmentation (CWS). The authors state the contribution of the previous CWS models are limited in the encoder.
They proposed the optimization of the decoder of Boundary-Enhanced Decoder (BED). Based on the conventional CRF decoder, the BED model introduces B... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper addresses the decoder improvements in Chinese Word Segmentation (CWS). The authors state the contribution of the previous CWS models are limited in the encoder.
They proposed the optimization of the decoder of Boundary-Enhanced Decoder (BED). Based on the conventional CRF decoder, the BED model intr... |
This work aims to train Gausssian-Bermoulli restricted Boltzmann machines (GRBMs) for generative modeling. For inference, it proposes to use a hybrid sampling that combines Gibbs sampling steps with Langevin samplers for GRBMs instead of the vanilla Gibbs sampling such that the gradient information of the log density c... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
This work aims to train Gausssian-Bermoulli restricted Boltzmann machines (GRBMs) for generative modeling. For inference, it proposes to use a hybrid sampling that combines Gibbs sampling steps with Langevin samplers for GRBMs instead of the vanilla Gibbs sampling such that the gradient information of the log d... |
This paper studies the effect of random pruning at initialization on the generalization of a neural network. The author theoretically shows that pruning of some two-layer convolutional neural networks under small initialization could improve their generalization if the probability of dropping a neural is mildly large, ... | Recommendation: 5: marginally below the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper studies the effect of random pruning at initialization on the generalization of a neural network. The author theoretically shows that pruning of some two-layer convolutional neural networks under small initialization could improve their generalization if the probability of dropping a neural is mildly... |
This paper presents a method to self-train ASR systems directly, without requiring an initial pre-training stage, which is especially useful in low-resource setups. Similarly to an existing method, slimIPL, the proposed approach generates pseudo-labels as training progresses and maintains a cache of pseudo-labels, regu... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper presents a method to self-train ASR systems directly, without requiring an initial pre-training stage, which is especially useful in low-resource setups. Similarly to an existing method, slimIPL, the proposed approach generates pseudo-labels as training progresses and maintains a cache of pseudo-labe... |
In this paper, the authors introduce a novel layer, called Gated Relational Message Passing (GRMP), for efficient multi-relational modelling over graphs. Importantly, GRMP scales better than existing multi-relational models when increasing the number of considered relations.
The authors use GRMP as building block for ... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
In this paper, the authors introduce a novel layer, called Gated Relational Message Passing (GRMP), for efficient multi-relational modelling over graphs. Importantly, GRMP scales better than existing multi-relational models when increasing the number of considered relations.
The authors use GRMP as building bl... |
The paper focuses on two player zero-sum stochastic games. It is shown that Optimistic FTRL (together with value update step) converges to a Nash equilibrium in $O(1/T)$ improving the result of Zhang et al.
The result is important and adds value to the current growing literature. The techniques are interesting, exploi... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper focuses on two player zero-sum stochastic games. It is shown that Optimistic FTRL (together with value update step) converges to a Nash equilibrium in $O(1/T)$ improving the result of Zhang et al.
The result is important and adds value to the current growing literature. The techniques are interesting... |
This paper builds the basic theoretical result for neural counterfactual inference. Based on the theoretical results, the authors present the algorithm to determine the identification of counterfacual and give the algorithm to estimate it if it is identifiable.
Strength:
1. Counterfactual inference is an interesting to... | Recommendation: 8: accept, good paper | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
This paper builds the basic theoretical result for neural counterfactual inference. Based on the theoretical results, the authors present the algorithm to determine the identification of counterfacual and give the algorithm to estimate it if it is identifiable.
Strength:
1. Counterfactual inference is an intere... |
The paper proposes an adversarially robust deepfake detection approach. The approaches divided the frequency information into multiple blocks based on random bisection or saliency information. This divided information is provided to the different models and later combined for robust detection accuracy. The proposed app... | 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 proposes an adversarially robust deepfake detection approach. The approaches divided the frequency information into multiple blocks based on random bisection or saliency information. This divided information is provided to the different models and later combined for robust detection accuracy. The prop... |
The paper presents an align-and-filter structure for neural network-based multi-channel speech enhancement.
The paper claims that the proposed decomposition of a filter into an element-wise product of alignment gain and a filtering gain is critical for an improved performance. Furthermore, it is claimed that training w... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper presents an align-and-filter structure for neural network-based multi-channel speech enhancement.
The paper claims that the proposed decomposition of a filter into an element-wise product of alignment gain and a filtering gain is critical for an improved performance. Furthermore, it is claimed that tr... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.