review stringlengths 5 16.6k | score stringclasses 5
values | area stringclasses 12
values | text stringlengths 31 5.65k |
|---|---|---|---|
The paper proposed a method to handle arbitrary clothing in Virtual try on models. The papers argue that recent approaches in virtual try-on lack coverage in cases where the clothing styles change vastly, for example from skirt to pants or long sleeve to short sleeves. To cover such cases they proposed a limb predictio... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The paper proposed a method to handle arbitrary clothing in Virtual try on models. The papers argue that recent approaches in virtual try-on lack coverage in cases where the clothing styles change vastly, for example from skirt to pants or long sleeve to short sleeves. To cover such cases they proposed a limb p... |
The paper proposes to use temperature scaling during the training process, thus improving accuracy and reducing the need for learning rate schedules. In the theoretical part of the paper, the gradients have been calculated analytically, comparing the gradients with and without the proposed modifications. In the experim... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes to use temperature scaling during the training process, thus improving accuracy and reducing the need for learning rate schedules. In the theoretical part of the paper, the gradients have been calculated analytically, comparing the gradients with and without the proposed modifications. In the... |
Traditional graph data augmentation (GDA) strategies regard augmentations as independent processes. This paper considers the GDA as a Markov decision process and develops a reinforced method, named GA2C, to achieve continuous and learnable GDA. Experiments on 17 datasets verify the effectiveness of GA2C.
Strength:
(1... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
Traditional graph data augmentation (GDA) strategies regard augmentations as independent processes. This paper considers the GDA as a Markov decision process and develops a reinforced method, named GA2C, to achieve continuous and learnable GDA. Experiments on 17 datasets verify the effectiveness of GA2C.
Streng... |
Paper uses the DP synthetic data, generated by adding Gaussian noise (Gaussian mechanism) to the sensitive data, to model the analysis using the measurement error approach. Simulations show that the proposed method works well for linear regression.
The proposed method draws interesting parallels between the measurement... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
Paper uses the DP synthetic data, generated by adding Gaussian noise (Gaussian mechanism) to the sensitive data, to model the analysis using the measurement error approach. Simulations show that the proposed method works well for linear regression.
The proposed method draws interesting parallels between the mea... |
This paper studies the problem of personalized bilevel optimization in a decentralized distributed learning system over stochastic and directed underlying communication networks. The major contribution of this paper is to propose an effective scheme termed HGP to compute the hyper-gradient, i.e., the gradient of hyperp... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
This paper studies the problem of personalized bilevel optimization in a decentralized distributed learning system over stochastic and directed underlying communication networks. The major contribution of this paper is to propose an effective scheme termed HGP to compute the hyper-gradient, i.e., the gradient o... |
The last layer of a neural network trained to classify instances usually linearly projects the feature vector from the previous layer to compute log odds of a class. This paper proposes replacement of that linear projection with a norm of the projection of the feature vector into a subspace. The paper shows how to opti... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The last layer of a neural network trained to classify instances usually linearly projects the feature vector from the previous layer to compute log odds of a class. This paper proposes replacement of that linear projection with a norm of the projection of the feature vector into a subspace. The paper shows how... |
The paper considers the case of incomplete physics for numerically solving Partial Differential Equations. It uses a neural network to complete the missing terms and more accurately forecast the evolution of the dynamical system. The proposed approach is evaluated on different variants of a reactive flow problem govern... | Recommendation: 3: reject, not good enough | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper considers the case of incomplete physics for numerically solving Partial Differential Equations. It uses a neural network to complete the missing terms and more accurately forecast the evolution of the dynamical system. The proposed approach is evaluated on different variants of a reactive flow proble... |
The paper proposes an improved graph learning algorithm based on a robust Gromov-Wasserstein discrepancy (RGWD). Following the previous GDL by Vincent-Cuaz et al. 2021, the new method replaces GWD by a new RGWD which uses a minimax formulation. Moreover, robust graph dictionary learning algorithm is also proposed based... | Recommendation: 8: accept, good paper | Area: Unsupervised and Self-supervised learning | Review:
The paper proposes an improved graph learning algorithm based on a robust Gromov-Wasserstein discrepancy (RGWD). Following the previous GDL by Vincent-Cuaz et al. 2021, the new method replaces GWD by a new RGWD which uses a minimax formulation. Moreover, robust graph dictionary learning algorithm is also propos... |
Test-time adaptation updates a model during testing to improve its generalization to different data.
A popular choice of test-time optimization is entropy minimization or pseudo-labeling, which trains on the model's own predictions.
This work seeks to prepare a model for such test-time optimization by meta-learning (1)... | Recommendation: 5: marginally below the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
Test-time adaptation updates a model during testing to improve its generalization to different data.
A popular choice of test-time optimization is entropy minimization or pseudo-labeling, which trains on the model's own predictions.
This work seeks to prepare a model for such test-time optimization by meta-lear... |
This paper proposes Contrast-Consistent Search (CCS), that tries to identify latent knowledge in language models in an unsupervised manner. The process is to convert each question into binary-answer pairs, then train the LM in an unsupervised way via optimizing a consistency loss (probabilities of opposite answers shou... | 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 Contrast-Consistent Search (CCS), that tries to identify latent knowledge in language models in an unsupervised manner. The process is to convert each question into binary-answer pairs, then train the LM in an unsupervised way via optimizing a consistency loss (probabilities of opposite answ... |
The paper examines the use of color in data augmentation for training deep networks and how it impacts the embedding representation of the input. In particular, they compare the traditional random color augmentation strategy with a color augmentation based on simulating changing illuminants. The illuminants used are dr... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The paper examines the use of color in data augmentation for training deep networks and how it impacts the embedding representation of the input. In particular, they compare the traditional random color augmentation strategy with a color augmentation based on simulating changing illuminants. The illuminants use... |
The paper suggest a k-subset sampling gradient estimator, namely SIMPLE. The proposed SIMPLE estimator can replace the reparameterization by relaxation with exact gradient estimation, and utilizes the gradient with respect to the exact conditional marginals in the backpropagtion. The authors conducted various experimen... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The paper suggest a k-subset sampling gradient estimator, namely SIMPLE. The proposed SIMPLE estimator can replace the reparameterization by relaxation with exact gradient estimation, and utilizes the gradient with respect to the exact conditional marginals in the backpropagtion. The authors conducted various e... |
This work studies the problem of learning a causal network from temporal data in the presence of under sampling. In particular, the authors propose a solver based approach for learning a graph from under sampled data, which improves upon prior art significantly in terms of running time. The core of the approach is conv... | Recommendation: 8: accept, good paper | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
This work studies the problem of learning a causal network from temporal data in the presence of under sampling. In particular, the authors propose a solver based approach for learning a graph from under sampled data, which improves upon prior art significantly in terms of running time. The core of the approach... |
This paper examines the convergence properties of existing RL algorithms for the average reward setting when the MDP satisfies a weaker-than-standard assumption on state visitation. It requires only weak communication, which means that a subset of states is communicating while others may be transient. For average-rew... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper examines the convergence properties of existing RL algorithms for the average reward setting when the MDP satisfies a weaker-than-standard assumption on state visitation. It requires only weak communication, which means that a subset of states is communicating while others may be transient. For ave... |
This algorithm proposes an algorithm for iterated batch reinforcement learning. The algorithm uses model-free RL to learn a guide policy, and then uses decision-time planning to improve the policy. The decision-time planning uses some exploration method and a rollout procedure to get a good action.
Strength:
The paper... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This algorithm proposes an algorithm for iterated batch reinforcement learning. The algorithm uses model-free RL to learn a guide policy, and then uses decision-time planning to improve the policy. The decision-time planning uses some exploration method and a rollout procedure to get a good action.
Strength:
T... |
This paper considers contextual bandits with concave rewards where the rewards are defined by a concave objective function with stochastic reward vectors. This setting removes the restrictions on the policy space compared to prior works. Under this setting, this paper proposes an (first) algorithm with the vanishing re... | Recommendation: 8: accept, good paper | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper considers contextual bandits with concave rewards where the rewards are defined by a concave objective function with stochastic reward vectors. This setting removes the restrictions on the policy space compared to prior works. Under this setting, this paper proposes an (first) algorithm with the vani... |
The paper under review studies the clustering problem on directed graphs using random walks. First, let us recall the clustering problem on undirected graphs using random walks. The setup here is rather well-studied. You just perform a suitably long walk at a random vertex within the cluster. Using spectral methods, yo... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
The paper under review studies the clustering problem on directed graphs using random walks. First, let us recall the clustering problem on undirected graphs using random walks. The setup here is rather well-studied. You just perform a suitably long walk at a random vertex within the cluster. Using spectral met... |
This paper presents a meta-learning formulation for learning to adapt a deep time-index model to the look-back window. INR was the choice of time-index models, and meta-learning was formulated by using samples from look-back window as context set and forecasting horizon as query set. The goal of the meta-learning was m... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper presents a meta-learning formulation for learning to adapt a deep time-index model to the look-back window. INR was the choice of time-index models, and meta-learning was formulated by using samples from look-back window as context set and forecasting horizon as query set. The goal of the meta-learni... |
This paper proposes a network based method, in the realm of physics informed methods, to solve a class of PDEs. The proposed methods mainly extend the general framework of having PDE as constraints in the loss function by introducing a partial density and locality of particle interactions. It also proposes to use MCMC ... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper proposes a network based method, in the realm of physics informed methods, to solve a class of PDEs. The proposed methods mainly extend the general framework of having PDE as constraints in the loss function by introducing a partial density and locality of particle interactions. It also proposes to u... |
This paper proposes an architecture to infer causal link from irregularly sampled time series data. The authors embrace the Granger causality paradigm. The method works by interleaving missing data imputation scheme and causal graph learning. The causal graph is used to structure the imputation predictions by using onl... | Recommendation: 8: accept, good paper | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
This paper proposes an architecture to infer causal link from irregularly sampled time series data. The authors embrace the Granger causality paradigm. The method works by interleaving missing data imputation scheme and causal graph learning. The causal graph is used to structure the imputation predictions by u... |
This paper presents a negative sampling strategy for contrastive learning, which many contrastive learning-related frameworks can incorporate to improve performance. Instead of the uniform sampler and kNN sampler, the proposed proximity graph could well capture the similarity relationships among instances. Random walks... | Recommendation: 3: reject, not good enough | Area: Unsupervised and Self-supervised learning | Review:
This paper presents a negative sampling strategy for contrastive learning, which many contrastive learning-related frameworks can incorporate to improve performance. Instead of the uniform sampler and kNN sampler, the proposed proximity graph could well capture the similarity relationships among instances. Rand... |
The paper proposes GradientMix, a regularization method where minibatch gradients are computed as a linear combination of per-example gradients with randomized rather than uniform example weights. The primary application it explores is improving generalization at large batch sizes.
The paper demonstrates GradientMix i... | Recommendation: 6: marginally above the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
The paper proposes GradientMix, a regularization method where minibatch gradients are computed as a linear combination of per-example gradients with randomized rather than uniform example weights. The primary application it explores is improving generalization at large batch sizes.
The paper demonstrates Gradi... |
The paper describes 1) five antibody prediction benchmark tasks, and 2) two loss functions for pre-training antibody language proteins to incorporate the evolutionary relationship of antibodies during pre-training.
## Strengths
* I am not aware of an existing benchmark specifically for antibodies
* The described loss... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper describes 1) five antibody prediction benchmark tasks, and 2) two loss functions for pre-training antibody language proteins to incorporate the evolutionary relationship of antibodies during pre-training.
## Strengths
* I am not aware of an existing benchmark specifically for antibodies
* The descri... |
The paper proposes a novel certification mechanism for certification in the multiple-output setting (e.g., image segmentation) based on randomized smoothing (RS), referred to as a collective certificate.
RS typically is used to create a robust-by-construction classifier in the (single-output) classification setting, by... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper proposes a novel certification mechanism for certification in the multiple-output setting (e.g., image segmentation) based on randomized smoothing (RS), referred to as a collective certificate.
RS typically is used to create a robust-by-construction classifier in the (single-output) classification set... |
This paper introduces a new approach that uses programmatic weak supervision. First, labeling functions are used to generate weak labels for each example. Then, a hyper label model is used to infer the ground-truth labels from the weak labels. The hyper label model is Graph Neural Network (GNN) that is trained on synth... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
This paper introduces a new approach that uses programmatic weak supervision. First, labeling functions are used to generate weak labels for each example. Then, a hyper label model is used to infer the ground-truth labels from the weak labels. The hyper label model is Graph Neural Network (GNN) that is trained ... |
In this manuscript, time-series forecasting has been studied in an online scenario. For online training of deep neural predictors one has to take into account both the new coming knowledge as well as retaining the learned patterns in the past (the so called stability-plasticity dilemma). The authors reportedly are insp... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
In this manuscript, time-series forecasting has been studied in an online scenario. For online training of deep neural predictors one has to take into account both the new coming knowledge as well as retaining the learned patterns in the past (the so called stability-plasticity dilemma). The authors reportedly ... |
The paper addresses deep neural network explainability and specifically proposes a method for generating textual description of neurons in a pre-trained network. The method operates as follows: Given a pretrained network f, a "probe set" D (a set of images), a "concept set" S (a set of words/phrases) and a neuron in f,... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper addresses deep neural network explainability and specifically proposes a method for generating textual description of neurons in a pre-trained network. The method operates as follows: Given a pretrained network f, a "probe set" D (a set of images), a "concept set" S (a set of words/phrases) and a neur... |
The paper investigates the possibility of masked frequency modeling for representation learning. It covers several research topics centered on MFM: a method that does MFM; connecting previous pre-training methods to MFM (e.g., super resolution); architecture change (from ViT to ConvNets) as an advantage of MFM over mas... | Recommendation: 8: accept, good paper | Area: Unsupervised and Self-supervised learning | Review:
The paper investigates the possibility of masked frequency modeling for representation learning. It covers several research topics centered on MFM: a method that does MFM; connecting previous pre-training methods to MFM (e.g., super resolution); architecture change (from ViT to ConvNets) as an advantage of MFM ... |
This paper studies the optimal precision of GANs trained on disconnected distributions under some simplifying assumptions about the distributions. The paper derives lower and upper bounds for the precision of GANs under these assumptions, and also studies what partitioning of the latent space that can give rise to the ... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
This paper studies the optimal precision of GANs trained on disconnected distributions under some simplifying assumptions about the distributions. The paper derives lower and upper bounds for the precision of GANs under these assumptions, and also studies what partitioning of the latent space that can give rise... |
This paper proposes a new objective function for mixed sample data augmentation (MSDA). Technically, the proposed Decoupled Mixup (DM) loss is equivalent to the summation of the original mixup loss and the additional losses that only consider one mixed target label while the other mixed target labels are ignored. For e... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a new objective function for mixed sample data augmentation (MSDA). Technically, the proposed Decoupled Mixup (DM) loss is equivalent to the summation of the original mixup loss and the additional losses that only consider one mixed target label while the other mixed target labels are ignore... |
The paper discusses a non-Markovian stochastic optimal control problem, with path-dependent dynamics and costs. It discusses the modeling of such problems and presents a numerical method based on a policy optimization procedure. For unknown dynamics, a model-based RL setup is considered, where the algorithm simultaneou... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper discusses a non-Markovian stochastic optimal control problem, with path-dependent dynamics and costs. It discusses the modeling of such problems and presents a numerical method based on a policy optimization procedure. For unknown dynamics, a model-based RL setup is considered, where the algorithm sim... |
I found this paper to be very interesting. The idea of having two 'simulators' one for vorticity for physics tracking and one for RGB video tracking is clever, and the coupling between vorticity/velocity in Equation (3) is something I was unfamiliar with. Although limited to 2D, in its current formulation, the idea of ... | Recommendation: 8: accept, good paper | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
I found this paper to be very interesting. The idea of having two 'simulators' one for vorticity for physics tracking and one for RGB video tracking is clever, and the coupling between vorticity/velocity in Equation (3) is something I was unfamiliar with. Although limited to 2D, in its current formulation, the ... |
The paper focuses on improving features to be more generalizable to achieve better performance in Long-Tailed Recognition (LTR). In particular, the paper adopts Stochastic Weight Averaging (SWA) for improving the generalization of deep neural networks. Moreover, the paper proposes a classifier re-training algorithm bas... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper focuses on improving features to be more generalizable to achieve better performance in Long-Tailed Recognition (LTR). In particular, the paper adopts Stochastic Weight Averaging (SWA) for improving the generalization of deep neural networks. Moreover, the paper proposes a classifier re-training algor... |
This paper evaluates the performance of state-of-the-art deep-learning methods such as RNNs, GNNs, and Transformers as well as the classic statistical ARIMA method to predict the tourism volume and tourism flow. Granular limited data supplied by a tourism region is extended by exogenous data such as geolocation traject... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper evaluates the performance of state-of-the-art deep-learning methods such as RNNs, GNNs, and Transformers as well as the classic statistical ARIMA method to predict the tourism volume and tourism flow. Granular limited data supplied by a tourism region is extended by exogenous data such as geolocation... |
This paper focuses on temporal variation modeling and proposes a representation method intended for incorporating multiple intraperiod- and interperiod-variations.
For each selected frequency, the proposed method generates a two-dimensional representation of the original time series and applies a convolutional network... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper focuses on temporal variation modeling and proposes a representation method intended for incorporating multiple intraperiod- and interperiod-variations.
For each selected frequency, the proposed method generates a two-dimensional representation of the original time series and applies a convolutional... |
The authors propose a color quantization transformer with the goal of reduding the number of bits needed to allocate the color information, but, at the same time, trying to look at the question: Do machine learning methods make evolve the different color terms as human civilizations?
The paper has a positive side, on t... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The authors propose a color quantization transformer with the goal of reduding the number of bits needed to allocate the color information, but, at the same time, trying to look at the question: Do machine learning methods make evolve the different color terms as human civilizations?
The paper has a positive si... |
This work proposes an autoencoder based multi-channel speech enhancement framework. Overall the work is applied and mostly empirical in nature. The proposed framework is based on existing well-established ideas from empirical deep learning. There are no theoretical insights and the contribution in terms of the combinat... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This work proposes an autoencoder based multi-channel speech enhancement framework. Overall the work is applied and mostly empirical in nature. The proposed framework is based on existing well-established ideas from empirical deep learning. There are no theoretical insights and the contribution in terms of the ... |
The paper presents a new method that learns a 3D-aware generative model on unaligned image datasets like ImageNet.
The authors propose to use an off-the-shelf depth estimator to guide 3D generator. The estimated depth is concatenated to the real image and feed to discriminator to guide the generated image and depth fr... | Recommendation: 3: reject, not good enough | Area: Generative models | Review:
The paper presents a new method that learns a 3D-aware generative model on unaligned image datasets like ImageNet.
The authors propose to use an off-the-shelf depth estimator to guide 3D generator. The estimated depth is concatenated to the real image and feed to discriminator to guide the generated image and ... |
The work shows a way of generating brain signals that can be useful in augmentation tasks. The current work shows that DPM-based brain signal generation is a very feasible task and can be used to create datasets that help improve deep learning models in classification tasks.
*Strength
-This work explored the applica... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The work shows a way of generating brain signals that can be useful in augmentation tasks. The current work shows that DPM-based brain signal generation is a very feasible task and can be used to create datasets that help improve deep learning models in classification tasks.
*Strength
-This work explored the... |
This paper proposes a new framework to jointly optimize the model size and inference acceleration. Compared to previous methods, This paper can train the sparse model considering the model size and computation efficiency simultaneously using an end-to-end manner. The unified framework is novel and elegant.
Strength:
... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a new framework to jointly optimize the model size and inference acceleration. Compared to previous methods, This paper can train the sparse model considering the model size and computation efficiency simultaneously using an end-to-end manner. The unified framework is novel and elegant.
Str... |
This paper points out the existing limitations of current contrastive-based VDA methods, including limited variations of positives within the same domain, inaccurate cross-domain positive pairs, and inevitable false negatives. To address the above limitations, the authors introduce target-domain nearest neighbors and s... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper points out the existing limitations of current contrastive-based VDA methods, including limited variations of positives within the same domain, inaccurate cross-domain positive pairs, and inevitable false negatives. To address the above limitations, the authors introduce target-domain nearest neighbo... |
This paper studies the different pre-training models for the antibody understanding tasks, propose new methods with biological information for pre-training, and a new antibody understanding benchmark is created. With different study experiments, the authors conclude several observations from different perspectives.
St... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper studies the different pre-training models for the antibody understanding tasks, propose new methods with biological information for pre-training, and a new antibody understanding benchmark is created. With different study experiments, the authors conclude several observations from different perspecti... |
This paper proposes the problem of minimax bias estimation of the value of a policy using offline data. The algorithm is based on marginal-importance-sampling and minimizing the Bellman residual error.
Strength:
1) The problem of estimating the confidence interval of the value of a policy using offline data is very im... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes the problem of minimax bias estimation of the value of a policy using offline data. The algorithm is based on marginal-importance-sampling and minimizing the Bellman residual error.
Strength:
1) The problem of estimating the confidence interval of the value of a policy using offline data is... |
The paper proposes to practical method called FairCOCCO that can incorporate multitype/multivariation sensitive attributes for fairness-aware learning. The method simply develops the kernel-based measure of fairness and uses it as a regularizer for learning a predictive model. The experimental results seem promising, b... | 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 to practical method called FairCOCCO that can incorporate multitype/multivariation sensitive attributes for fairness-aware learning. The method simply develops the kernel-based measure of fairness and uses it as a regularizer for learning a predictive model. The experimental results seem prom... |
This paper focuses on the referring expression comprehension (REC) problem, ie, localizing the description of a natural language expression in an image. Specifically, they argue that existing Transformer-based methods don't use any location priors, which may result in inaccuracies in practice. To this end, they propose... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper focuses on the referring expression comprehension (REC) problem, ie, localizing the description of a natural language expression in an image. Specifically, they argue that existing Transformer-based methods don't use any location priors, which may result in inaccuracies in practice. To this end, they... |
The paper proposed two methods to improve auto curriculum learning strategies in the RL settings where multiple goals are available with no change to the environment dynamics:
1. Continuous goal sampling, where the goal of the agent changes every R < episode length steps
2. A curriculum based on "learning progress", w... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper proposed two methods to improve auto curriculum learning strategies in the RL settings where multiple goals are available with no change to the environment dynamics:
1. Continuous goal sampling, where the goal of the agent changes every R < episode length steps
2. A curriculum based on "learning prog... |
This paper studies policy evaluation with nonlinear function approximation. It proposes and analyzed two algorithms called VRPD and VRPD+ to optimize the primal-dual form of the MSBPE. The VRPD algorithm utilizes variance reduction techniques to achieve $O(1/K)$ convergence rate.
Strength:
1. Paper is well-written and ... | Recommendation: 6: marginally above the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper studies policy evaluation with nonlinear function approximation. It proposes and analyzed two algorithms called VRPD and VRPD+ to optimize the primal-dual form of the MSBPE. The VRPD algorithm utilizes variance reduction techniques to achieve $O(1/K)$ convergence rate.
Strength:
1. Paper is well-writ... |
This paper proposes a novel semi-parametric language model architecture dubbed Knowledge-in-Context (KiC) which utilizes external knowledge memories of K (K=6 in the experiments) different structured knowledge types (dictionary, commonsense, Entity, event, script, causality) to make the prediction. When a query is give... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a novel semi-parametric language model architecture dubbed Knowledge-in-Context (KiC) which utilizes external knowledge memories of K (K=6 in the experiments) different structured knowledge types (dictionary, commonsense, Entity, event, script, causality) to make the prediction. When a query... |
The paper aims to understand Mixup training, and specifically the training dynamics if mixup is continued to run for much longer than is usual. The paper shows an interesting empirical phenomenon, that overtraining with mixup leads to a U-shaped curve for training error, i.e. training error increases after a point (and... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The paper aims to understand Mixup training, and specifically the training dynamics if mixup is continued to run for much longer than is usual. The paper shows an interesting empirical phenomenon, that overtraining with mixup leads to a U-shaped curve for training error, i.e. training error increases after a po... |
Motivated by the fact that data diversity plays a key role in representation learning, this paper proposes "Multi-View Masked Autoencoder" to leverage the cross-view and cross-frame information to learn the visual representation for control tasks. Experiments validate that the proposed multi-view representation learnin... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
Motivated by the fact that data diversity plays a key role in representation learning, this paper proposes "Multi-View Masked Autoencoder" to leverage the cross-view and cross-frame information to learn the visual representation for control tasks. Experiments validate that the proposed multi-view representation... |
The paper empirically groups sequence to sequence neural architectures models according to the Chomsky Hierarchy. They achieve this by evaluating 10250 neural architectures models across 15 transduction tasks spanning the entire Chomsky Hierarchy.
They demonstrate that memory augmented architectures tend to fall highe... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The paper empirically groups sequence to sequence neural architectures models according to the Chomsky Hierarchy. They achieve this by evaluating 10250 neural architectures models across 15 transduction tasks spanning the entire Chomsky Hierarchy.
They demonstrate that memory augmented architectures tend to fa... |
The authors claim to find a relationship between compute, model size and performance (here measured by ELO). Studies are conducted over 2 games (Connect Four and Pentago) and the AlphaZero algorithm. Finally the authors combine these relationships to make recommendations based on optimal model size given the amount of ... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The authors claim to find a relationship between compute, model size and performance (here measured by ELO). Studies are conducted over 2 games (Connect Four and Pentago) and the AlphaZero algorithm. Finally the authors combine these relationships to make recommendations based on optimal model size given the am... |
The paper introduces NIERT which is a new framework for numerical interpolation for scattered data. The key idea of the framework is to use Transformers as a representation encoder and treats observed and target points in a unified fashion by embedding them into the same representation space. In the paper, the authors ... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper introduces NIERT which is a new framework for numerical interpolation for scattered data. The key idea of the framework is to use Transformers as a representation encoder and treats observed and target points in a unified fashion by embedding them into the same representation space. In the paper, the ... |
This paper studies prediction bias amplification of graph neural networks using contextualized stochastic block model. Specifically, their analysis utilizes a distribution distance as bias measurement and discusses when $\Delta$ bias is enhanced regarding homophily edge ratio, graph size and density. Based on it, the a... | Recommendation: 3: reject, not good enough | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper studies prediction bias amplification of graph neural networks using contextualized stochastic block model. Specifically, their analysis utilizes a distribution distance as bias measurement and discusses when $\Delta$ bias is enhanced regarding homophily edge ratio, graph size and density. Based on i... |
This is a practical paper providing empirical insights on how to better backdoor attack a DNN model via poisoning samples (without interfering with the training process). Three observations/ideas are provided: 1) using shared adversarial perturbation as the trigger; 2) using easiness of being misclassified through a tr... | 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 is a practical paper providing empirical insights on how to better backdoor attack a DNN model via poisoning samples (without interfering with the training process). Three observations/ideas are provided: 1) using shared adversarial perturbation as the trigger; 2) using easiness of being misclassified thro... |
This paper proposes LAtFormer, which incorporates lattice geometry and topology priors in attention masks. The proposed architecture is a modification of the standard attention mechanism where attention weights are scaled using soft attention masks generated by a convolution neural net. Experiments on ARC and synthetic... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes LAtFormer, which incorporates lattice geometry and topology priors in attention masks. The proposed architecture is a modification of the standard attention mechanism where attention weights are scaled using soft attention masks generated by a convolution neural net. Experiments on ARC and s... |
This paper presents an auto-encoder based pipeline to achieve versatile neural process, which consists of a bottleneck encoder and modulated MLP based decoder. The bottleneck encoder aims to produce fewer and informative context tokens while the decoder hierarchically learns multiple global latent variables and the unc... | Recommendation: 6: marginally above the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
This paper presents an auto-encoder based pipeline to achieve versatile neural process, which consists of a bottleneck encoder and modulated MLP based decoder. The bottleneck encoder aims to produce fewer and informative context tokens while the decoder hierarchically learns multiple global latent variables and... |
This paper provides theoretical and experimental support to the following hypothesis: maps induced by denoising diffusion processes are optimal transport maps between the base and target distributions. For the case of multidimensional Gaussians, the paper provides a complete analytical result. For the general case, it ... | Recommendation: 8: accept, good paper | Area: Generative models | Review:
This paper provides theoretical and experimental support to the following hypothesis: maps induced by denoising diffusion processes are optimal transport maps between the base and target distributions. For the case of multidimensional Gaussians, the paper provides a complete analytical result. For the general c... |
This paper proposed a framework, named MultiWave, which augments any deep learning time series model with components operating at the intrinsic frequencies of the signals. It applies discrete wavelet decomposition on each signal to obtain subsignals of different frequencies and groups all subsignals in the same frequen... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper proposed a framework, named MultiWave, which augments any deep learning time series model with components operating at the intrinsic frequencies of the signals. It applies discrete wavelet decomposition on each signal to obtain subsignals of different frequencies and groups all subsignals in the same... |
The authors propose to formulate sketch generation as the reversal process of sketch deformation, allowing the use of diffusion process that's more appropriate to stroke based sketch drawing. This is quite similar to [Luhman and Luhman, 2020] which use the same approach but on handwriting generation. The authors also p... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
The authors propose to formulate sketch generation as the reversal process of sketch deformation, allowing the use of diffusion process that's more appropriate to stroke based sketch drawing. This is quite similar to [Luhman and Luhman, 2020] which use the same approach but on handwriting generation. The author... |
This paper presents a theoretically-motivated framework that reformulates a single-task RL problem as a multi-task RL problem defined by a curriculum for computationally-efficient policy learning. Their framework and other baseline are demonstrated in a goal-reaching task with a Mujoco dynamical system. The results sho... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper presents a theoretically-motivated framework that reformulates a single-task RL problem as a multi-task RL problem defined by a curriculum for computationally-efficient policy learning. Their framework and other baseline are demonstrated in a goal-reaching task with a Mujoco dynamical system. The res... |
The authors propose a new benchmark set for symbolic regression. Starting from the Feynman symbolic regression dataset (FSRD, Udrescu et al), they propose a new dataset of 120 problems, together with rules for sampling variables and parameters. They introduce a new metric for comparing symbolic functions, based on the ... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The authors propose a new benchmark set for symbolic regression. Starting from the Feynman symbolic regression dataset (FSRD, Udrescu et al), they propose a new dataset of 120 problems, together with rules for sampling variables and parameters. They introduce a new metric for comparing symbolic functions, based... |
This paper sets out to analyze the influence of attributed graph elements on the changes of parameters in a Graph Convolutional Network under Simple Graph Convolution, without having to go through retraining. The derived influence functions are capable to estimate changes caused by the removal of nodes or edges from th... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
This paper sets out to analyze the influence of attributed graph elements on the changes of parameters in a Graph Convolutional Network under Simple Graph Convolution, without having to go through retraining. The derived influence functions are capable to estimate changes caused by the removal of nodes or edges... |
The paper studies top-k classification problem. Observing existing top-k classification methods neglect the ranking of the ground truth label among the predicted k labels, the authors proposed to take label ranking into consideration. Specifically, a new 3-stage approach was proposed to tackle this problem. In the firs... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
The paper studies top-k classification problem. Observing existing top-k classification methods neglect the ranking of the ground truth label among the predicted k labels, the authors proposed to take label ranking into consideration. Specifically, a new 3-stage approach was proposed to tackle this problem. In ... |
In the paper, the authors propose a continuous and reversible memory transformation method to prevent overfitting on the memory in continual learning. The main idea is to increase the diversity of the data in the memory buffer while maintaining hardness. The authors propose Deterministic Continuous Memory Transformer t... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
In the paper, the authors propose a continuous and reversible memory transformation method to prevent overfitting on the memory in continual learning. The main idea is to increase the diversity of the data in the memory buffer while maintaining hardness. The authors propose Deterministic Continuous Memory Trans... |
This paper perform an exhaustive investigation of the role of batch size and the steps used in temporal difference learning. Unlike the conventional intuition in supervised learning that increasing batch size would lead to variance reduction, thus increase the overall performance. In RL (Deep Q Learning based algorithm... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper perform an exhaustive investigation of the role of batch size and the steps used in temporal difference learning. Unlike the conventional intuition in supervised learning that increasing batch size would lead to variance reduction, thus increase the overall performance. In RL (Deep Q Learning based a... |
The paper introduces a novel specification for neural network verification,
where instead of checking whether sets of inputs are classified equivalently by
a network, it concerns checking the robustness of networks when their behaviour
is restricted to certain neural activation patterns.
+ The specifications introduced... | 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 novel specification for neural network verification,
where instead of checking whether sets of inputs are classified equivalently by
a network, it concerns checking the robustness of networks when their behaviour
is restricted to certain neural activation patterns.
+ The specifications in... |
This paper investigates the neural collapse phenomenon (minimal intra-class variance and maximally separated class means in learned representations of neural networks) in the context of transfer learning. They find that this type of collapse in the representations of the pretrained model hurts transfer performance (tho... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper investigates the neural collapse phenomenon (minimal intra-class variance and maximally separated class means in learned representations of neural networks) in the context of transfer learning. They find that this type of collapse in the representations of the pretrained model hurts transfer performa... |
The paper is trying to deal with non-linear relationships in time series data and proposes a structural equation model called Rhino, which combines vector auto-regression, deep learning as well as variational inference techniques.
Strength
1. The problem causal discovery the paper is trying to tackle is quite important... | Recommendation: 6: marginally above the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The paper is trying to deal with non-linear relationships in time series data and proposes a structural equation model called Rhino, which combines vector auto-regression, deep learning as well as variational inference techniques.
Strength
1. The problem causal discovery the paper is trying to tackle is quite i... |
This is a theoretical paper studying a method for regression in high dimensions. In particular, it studies the use of PCA followed by linear regression. Their motivation for doing so is to better understand overparameterization in more general encoder-decoder models. The authors provide an explicit formula for the expe... | Recommendation: 3: reject, not good enough | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This is a theoretical paper studying a method for regression in high dimensions. In particular, it studies the use of PCA followed by linear regression. Their motivation for doing so is to better understand overparameterization in more general encoder-decoder models. The authors provide an explicit formula for ... |
This paper presents a novel method for modeling neural correlations
based on sum-product networks. The idea is that inference of the
statistical structure of neural population activity is made easier by
the assumptions encoded in the choice of this type of network, which
enforce a hierarchical structure of alternating ... | Recommendation: 8: accept, good paper | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
This paper presents a novel method for modeling neural correlations
based on sum-product networks. The idea is that inference of the
statistical structure of neural population activity is made easier by
the assumptions encoded in the choice of this type of network, which
enforce a hierarchical structure of alte... |
Paper proposes a membership inference method based on learning query vectors, and shows that it improves the current state-of-the-art methods.
Membership inference attacks are of great interest to the privacy and security community, and the paper proposes a method to improve LiRA by learning the query vectors. I find ... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
Paper proposes a membership inference method based on learning query vectors, and shows that it improves the current state-of-the-art methods.
Membership inference attacks are of great interest to the privacy and security community, and the paper proposes a method to improve LiRA by learning the query vectors.... |
This paper proposes a new video-language pre-training (VLP) method which is computationally efficient. The key idea is to reduce (1) temporal visual redundancy by frame sampling and (2) spatial visual redundancy by using region features that are extracted by a pre-learned object detector. To align region features and t... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a new video-language pre-training (VLP) method which is computationally efficient. The key idea is to reduce (1) temporal visual redundancy by frame sampling and (2) spatial visual redundancy by using region features that are extracted by a pre-learned object detector. To align region featur... |
This is a work on rotationally equivariant feed forward networks for applications in protein folding and related tasks.
A number of equivariant operations which appear useful for these applications are introduced, such as a cross product operator.
Experiments are done on several useful tasks using protein structure dat... | Recommendation: 8: accept, good paper | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This is a work on rotationally equivariant feed forward networks for applications in protein folding and related tasks.
A number of equivariant operations which appear useful for these applications are introduced, such as a cross product operator.
Experiments are done on several useful tasks using protein struc... |
The paper studies random features ridge regression where the data $\{(x_i, y_i)\}_{i=1}^n$ is generated from a noisy nonlinear model $y_i=f_d(x_i)+\epsilon_i$ with the covariates $x_i$ i.i.d. uniformly sampled from $\mathbb{S}^{d-1}(\sqrt{d})$.
The estimator is given by $f_{a,\Theta}(x)=\sum_{i=1}^{N}a_i\sigma(\langl... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The paper studies random features ridge regression where the data $\{(x_i, y_i)\}_{i=1}^n$ is generated from a noisy nonlinear model $y_i=f_d(x_i)+\epsilon_i$ with the covariates $x_i$ i.i.d. uniformly sampled from $\mathbb{S}^{d-1}(\sqrt{d})$.
The estimator is given by $f_{a,\Theta}(x)=\sum_{i=1}^{N}a_i\sigm... |
The authors propose a way to make quantum kernels generalizable through a bandwidth parameter that can restrict features in the Bloch sphere. They also provide some evidence on real datasts that these kernels can generalize unlike the ones without the bandwidth parameter.
+ it's a simple but possibly powerful way to i... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
The authors propose a way to make quantum kernels generalizable through a bandwidth parameter that can restrict features in the Bloch sphere. They also provide some evidence on real datasts that these kernels can generalize unlike the ones without the bandwidth parameter.
+ it's a simple but possibly powerful ... |
This paper tackles developing neural machine translation models in the federated learning scenario.
There have been several methods that attempt to solve the identical scenario.
The authors point out that the conventional methods require vast communication overheads and heavy synchronization that makes the methods impr... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper tackles developing neural machine translation models in the federated learning scenario.
There have been several methods that attempt to solve the identical scenario.
The authors point out that the conventional methods require vast communication overheads and heavy synchronization that makes the meth... |
The authors are interested in structured pruning of generative language models. In particular, this work builds on top of [Movement Pruning](https://arxiv.org/abs/2005.07683) and [Block movement pruning](https://arxiv.org/abs/2109.04838) to prune entire structures (and not individual weights) in models similar to GPT2 ... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The authors are interested in structured pruning of generative language models. In particular, this work builds on top of [Movement Pruning](https://arxiv.org/abs/2005.07683) and [Block movement pruning](https://arxiv.org/abs/2109.04838) to prune entire structures (and not individual weights) in models similar ... |
The paper extends previous results on individual privacy accounting from Rényi DP to Gaussian DP. It also considers methods to maintain approximate privacy filters.
## Strengths
- The paper has adequate review of the relevant literature, is mostly self-contained, and clearly states the prior results it builds upon.
- I... | 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 extends previous results on individual privacy accounting from Rényi DP to Gaussian DP. It also considers methods to maintain approximate privacy filters.
## Strengths
- The paper has adequate review of the relevant literature, is mostly self-contained, and clearly states the prior results it builds u... |
The paper proposes a new way of measuring forward transfer. They define forward transfer as how easy it is to learn a new task given a set of representations produced by continual learning on previous tasks. And they are looking for an answer to the question of "whether less forgetful representations are more transfera... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes a new way of measuring forward transfer. They define forward transfer as how easy it is to learn a new task given a set of representations produced by continual learning on previous tasks. And they are looking for an answer to the question of "whether less forgetful representations are more t... |
The paper presented a physics-informed neural network model to estimate and predict 3D rotational dynamics from image sequences. The target is achieved by using a multi-stage prediction pipeline that maps individual images to a latent representation homeomorphic to SO(3), computes angular velocities from latent pairs, ... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper presented a physics-informed neural network model to estimate and predict 3D rotational dynamics from image sequences. The target is achieved by using a multi-stage prediction pipeline that maps individual images to a latent representation homeomorphic to SO(3), computes angular velocities from latent... |
This paper proposes an evolutionary approach, $\mu$2Net, to dynamically generate a large scale multitask network for multiple tasks. The evolutionary approach comprises (1) sampling a parent model, (2) mutations, (3) training and scoring children models. Extensive experiments are performed on numerous datasets.
Strengt... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes an evolutionary approach, $\mu$2Net, to dynamically generate a large scale multitask network for multiple tasks. The evolutionary approach comprises (1) sampling a parent model, (2) mutations, (3) training and scoring children models. Extensive experiments are performed on numerous datasets.... |
The paper considers the problem of optimal filtering and simultaneous parameter estimation for discrete-time systems with exponential family latent dynamics. The authors propose a new variational inference algorithm, by splitting the prediction and update step in the filtering algorithm. The variational approximation f... | Recommendation: 8: accept, good paper | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
The paper considers the problem of optimal filtering and simultaneous parameter estimation for discrete-time systems with exponential family latent dynamics. The authors propose a new variational inference algorithm, by splitting the prediction and update step in the filtering algorithm. The variational approxi... |
This paper proposes a neural representation method of neural networks (NeRN) to represent the weights of neural networks, which maps coordinates to convolutional kernels. The paper also considers spatial smoothness constraints on networks’ weights to help NeRN. Moreover, the paper also points out two applications usin... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper proposes a neural representation method of neural networks (NeRN) to represent the weights of neural networks, which maps coordinates to convolutional kernels. The paper also considers spatial smoothness constraints on networks’ weights to help NeRN. Moreover, the paper also points out two applicati... |
The authors propose the DMITRI framework to generate dynamic embeddings for multi-entity interactions with timestamps. A multi-entity interaction involves potentially many types of entities, e.g. traffic type, severity, latitude, and longitude, and consists of a timestamped observation for those entities, such as the n... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The authors propose the DMITRI framework to generate dynamic embeddings for multi-entity interactions with timestamps. A multi-entity interaction involves potentially many types of entities, e.g. traffic type, severity, latitude, and longitude, and consists of a timestamped observation for those entities, such ... |
This paper tries to address the problem of locating runtime error with neural interpreter model. The contribution of this paper includes: (1) a new dataset for the problem, which consists of competitive programs in multiple languages, input data files, input descriptions, runtime error labels. (2) an improved model bas... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper tries to address the problem of locating runtime error with neural interpreter model. The contribution of this paper includes: (1) a new dataset for the problem, which consists of competitive programs in multiple languages, input data files, input descriptions, runtime error labels. (2) an improved m... |
The authors propose a task-agnostic framework to measure the quality of pretrained representations. This approach uses generated data from a conditional Gaussian mixture to evaluate the learned representation. The usefulness of this metric is evaluated by comparing to results from linear probing on CIFAR 10/10-c
Stre... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The authors propose a task-agnostic framework to measure the quality of pretrained representations. This approach uses generated data from a conditional Gaussian mixture to evaluate the learned representation. The usefulness of this metric is evaluated by comparing to results from linear probing on CIFAR 10/1... |
This work analyzes catastrophic forgetting during different multi-task learning strategies, specifically sequence training and replay. The authors test BERT on a suite of classification tasks and question answering, and use linear probes to test the ability of the model on each task before training, after training and ... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This work analyzes catastrophic forgetting during different multi-task learning strategies, specifically sequence training and replay. The authors test BERT on a suite of classification tasks and question answering, and use linear probes to test the ability of the model on each task before training, after train... |
This paper provides a safety-aware control approach, based on neural networks, for stabilizing stochastic delay differential equations (SDDEs). In particular, the authors construct a controller with two components - a deterministic control $u_f$ tied to the time-evolution of the system, and a stochastic control $u_g$ t... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper provides a safety-aware control approach, based on neural networks, for stabilizing stochastic delay differential equations (SDDEs). In particular, the authors construct a controller with two components - a deterministic control $u_f$ tied to the time-evolution of the system, and a stochastic control... |
Detecting overfitting of deep learning networks in an efficient, accurate and non-intrusive manner is a non-trivial problem. The authors propose a method to learn a classifier that uses the history of training and validation losses to identify overfitting regions. The paper evaluates the approach using a number of loss... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
Detecting overfitting of deep learning networks in an efficient, accurate and non-intrusive manner is a non-trivial problem. The authors propose a method to learn a classifier that uses the history of training and validation losses to identify overfitting regions. The paper evaluates the approach using a number... |
This article demonstrates the uncertainty distribution of almost all existing OSR methods is actually closer to the expectation of UOSR than OSR. This article also introduces a new evaluation setting into UOSR, which is few-shot UOSR. Then, the FS-KNNS method is proposed, which achieves state-of-the-art performance und... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This article demonstrates the uncertainty distribution of almost all existing OSR methods is actually closer to the expectation of UOSR than OSR. This article also introduces a new evaluation setting into UOSR, which is few-shot UOSR. Then, the FS-KNNS method is proposed, which achieves state-of-the-art perform... |
The authors follow in the line of recent improvements on the slot-attention architecture for unsupervised object discovery that reconstruct a frame not in the raw pixel space, but rather in a more structured space (e.g. flow for SaVi, depth for SaVi++ or feature space of a network which is (pre-)trained in a self-super... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The authors follow in the line of recent improvements on the slot-attention architecture for unsupervised object discovery that reconstruct a frame not in the raw pixel space, but rather in a more structured space (e.g. flow for SaVi, depth for SaVi++ or feature space of a network which is (pre-)trained in a se... |
This paper studies the separation of two antisymmetric ansätze, i.e. Slater ansätze and Jastrow ansätze.
The topic of this paper is interesting and the writing is mostly clear.
However I have concerns about the soundness of theory and experiments.
1.
After Theorem 3.1, authors claim that "For particular choices of th... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper studies the separation of two antisymmetric ansätze, i.e. Slater ansätze and Jastrow ansätze.
The topic of this paper is interesting and the writing is mostly clear.
However I have concerns about the soundness of theory and experiments.
1.
After Theorem 3.1, authors claim that "For particular choic... |
This paper propose a method for representation learning robust to out of distribution cases for the times series classification cases.
## Strength
- consider worst-case scenario and develop the algorithm and modeling end to end.
- extensive real-data experiments to demonstrate the superiority of the method.
## Weakne... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper propose a method for representation learning robust to out of distribution cases for the times series classification cases.
## Strength
- consider worst-case scenario and develop the algorithm and modeling end to end.
- extensive real-data experiments to demonstrate the superiority of the method.
#... |
The paper analyzes the essence of MIM and introduces a universal MIM method that can be applied to both CNNs and Transformers. There are several designs that together build the introduced method, including the mean RGB replacement, the intermediate mask, the frequency reconstruction targets and the HOG targets. Experim... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
The paper analyzes the essence of MIM and introduces a universal MIM method that can be applied to both CNNs and Transformers. There are several designs that together build the introduced method, including the mean RGB replacement, the intermediate mask, the frequency reconstruction targets and the HOG targets.... |
The paper proposes an (approximation) algorithm to compute Gromov-Wasserstein (GW) distance, and called it Bregman Alternating Projected Gradient (BAPG) method. The author(s) claim that this is the first single-loop algorithm that has provable convergence guarantees. In particular, they provide an approximation bound f... | Recommendation: 8: accept, good paper | Area: Optimization (eg, convex and non-convex optimization) | Review:
The paper proposes an (approximation) algorithm to compute Gromov-Wasserstein (GW) distance, and called it Bregman Alternating Projected Gradient (BAPG) method. The author(s) claim that this is the first single-loop algorithm that has provable convergence guarantees. In particular, they provide an approximation... |
The paper analyzes Deep operator Networks (DeepONets) regarding the optimization convergence guarantees. The author provides optimization analysis on two different networks: smooth activations and ReLU activations. Their analysis shows that the overparameterization on branch and trunk networks allows faster convergence... | Recommendation: 8: accept, good paper | Area: Optimization (eg, convex and non-convex optimization) | Review:
The paper analyzes Deep operator Networks (DeepONets) regarding the optimization convergence guarantees. The author provides optimization analysis on two different networks: smooth activations and ReLU activations. Their analysis shows that the overparameterization on branch and trunk networks allows faster con... |
This paper introduces a trainable way to perform weight averaging and explores it on several image classification datasets.
Strengths:
- I like the generalization of traditional weight averaging by explicitly training for it, especially since it empirically helps.
- Authors also consider the computational cost of TWA a... | Recommendation: 8: accept, good paper | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper introduces a trainable way to perform weight averaging and explores it on several image classification datasets.
Strengths:
- I like the generalization of traditional weight averaging by explicitly training for it, especially since it empirically helps.
- Authors also consider the computational cost ... |
This work aims at improving query-based video segmentation based on domain generalization. Specifically, this work focuses on solving the challenges of cross-modal settings, where domain shift exists at both the text and visual levels. To mitigate these challenges, this work proposes the query-guided feature augmentati... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This work aims at improving query-based video segmentation based on domain generalization. Specifically, this work focuses on solving the challenges of cross-modal settings, where domain shift exists at both the text and visual levels. To mitigate these challenges, this work proposes the query-guided feature au... |
This paper proposes bit-rate constrained group DRO, where the assumption is that group identities can be modeled by a simple function class. This allows the framework to improve the overall utility of the learned model by relying less on arbitrarily chosen mis-labeled points (in a DRO framework), but ensure that only o... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes bit-rate constrained group DRO, where the assumption is that group identities can be modeled by a simple function class. This allows the framework to improve the overall utility of the learned model by relying less on arbitrarily chosen mis-labeled points (in a DRO framework), but ensure tha... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.