review stringlengths 5 16.6k | score stringclasses 5
values | area stringclasses 12
values | text stringlengths 31 5.65k |
|---|---|---|---|
This paper addresses the problem of catastrophic forgetting in deep neural networks by introducing a complementary network for gating the neurons. This second network is trained based on the same input to produce task-aware gates which are sparse and regularized to be similar within each task while orthogonal between t... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
This paper addresses the problem of catastrophic forgetting in deep neural networks by introducing a complementary network for gating the neurons. This second network is trained based on the same input to produce task-aware gates which are sparse and regularized to be similar within each task while orthogonal b... |
This paper proposes UniFormerV2, which could arm the readily available and well-pretrained image ViT with efficient Uniformer designs. Extensive experiments demonstrate the effectiveness of the proposed method.
Pros:
1. The idea of arming rich well-pretrained image ViTs with efficient Uniformer design for video perform... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes UniFormerV2, which could arm the readily available and well-pretrained image ViT with efficient Uniformer designs. Extensive experiments demonstrate the effectiveness of the proposed method.
Pros:
1. The idea of arming rich well-pretrained image ViTs with efficient Uniformer design for video... |
The authors propose an attention-based contrastive learning model (CVT) for OOD detection in image classification tasks based on self-supervised methods by using a ViT as a feature extractor. The proposed model outperforms several SOTA methods based on CIFAR-10/100.
Strengths:
Identifying OOD data is an important top... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
The authors propose an attention-based contrastive learning model (CVT) for OOD detection in image classification tasks based on self-supervised methods by using a ViT as a feature extractor. The proposed model outperforms several SOTA methods based on CIFAR-10/100.
Strengths:
Identifying OOD data is an impor... |
The paper has several contributions:
1. it derives $(\mathcal{O}\frac{1}{\sqrt{T}})$ convergence rate of the best iterate of an extension of Optimistic gradient (OG). The setup is weak minty VIs (MVI)---which is a setting that includes non-monotone VIs as well; Sec. $3$.
2. proposes an accelerated version of Reflected ... | Recommendation: 5: marginally below the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The paper has several contributions:
1. it derives $(\mathcal{O}\frac{1}{\sqrt{T}})$ convergence rate of the best iterate of an extension of Optimistic gradient (OG). The setup is weak minty VIs (MVI)---which is a setting that includes non-monotone VIs as well; Sec. $3$.
2. proposes an accelerated version of Re... |
The paper proposes an algorithm to identify "meta-factor" from subjects' answers to multiple questionnaires with different lengths. Specific constraints are introduced and the problem is formulated a constrained optimization problem. An ADMM algorithm is derived to compute the factorization solution. Performance compar... | Recommendation: 3: reject, not good enough | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper proposes an algorithm to identify "meta-factor" from subjects' answers to multiple questionnaires with different lengths. Specific constraints are introduced and the problem is formulated a constrained optimization problem. An ADMM algorithm is derived to compute the factorization solution. Performanc... |
The paper draws attention to the problem of Bias Analysis in text summarization tasks which has not been addressed before. It defines the different types of biases present in summarisation task. It performs Bias Analysis using both abstractive and extractive automatic summarising models. The experiments suggest biasnes... | Recommendation: 3: reject, not good enough | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper draws attention to the problem of Bias Analysis in text summarization tasks which has not been addressed before. It defines the different types of biases present in summarisation task. It performs Bias Analysis using both abstractive and extractive automatic summarising models. The experiments suggest... |
This paper presents a method for dynamic parameter sharing over individual and group levels.
Results show an improvement over standard baselines on two challenging benchmarks (Google Research Football and StarCraft).
# Strengths
* The method addresses a challenging and important problem.
* The results seem to indicate ... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper presents a method for dynamic parameter sharing over individual and group levels.
Results show an improvement over standard baselines on two challenging benchmarks (Google Research Football and StarCraft).
# Strengths
* The method addresses a challenging and important problem.
* The results seem to i... |
This paper aims to explore the suitability of various Knowledge Graph Embeddings (KGE) for graph structure prediction tasks such as entity and relation neighborhood prediction, type/range of a given relation, and relation prediction. This text also investigates the performance of these models on several downstream task... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper aims to explore the suitability of various Knowledge Graph Embeddings (KGE) for graph structure prediction tasks such as entity and relation neighborhood prediction, type/range of a given relation, and relation prediction. This text also investigates the performance of these models on several downstr... |
The author proposes to replace the random mask in FedBABU with a masked optimized for the device heterogeneity. The proposed masking is layer-wise so that all parameters in a layer will be frozen or not, based on the formulation Eq(10). Empirical evaluation shows better performance, less trainable parameters and FLOPS.... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The author proposes to replace the random mask in FedBABU with a masked optimized for the device heterogeneity. The proposed masking is layer-wise so that all parameters in a layer will be frozen or not, based on the formulation Eq(10). Empirical evaluation shows better performance, less trainable parameters an... |
This paper presents a method for continual learning (Cognitive Continual Learner, CCL) with three components: a ‘working model’ that learns to classify from raw input, an ‘inductive bias learner’ that learns to classify from a processed version of the input in the form of shape information, and a ’semantic learner’, wh... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper presents a method for continual learning (Cognitive Continual Learner, CCL) with three components: a ‘working model’ that learns to classify from raw input, an ‘inductive bias learner’ that learns to classify from a processed version of the input in the form of shape information, and a ’semantic lear... |
This paper is a theoretical analysis of a recent work [a]. They use the Gaussian model to explain and understand the gradient descent with hard-label and conjugate labels. For convenience analysis, they only consider the binary classification problem. The theoretical results and analysis are interesting.
[a] Test-Tim... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper is a theoretical analysis of a recent work [a]. They use the Gaussian model to explain and understand the gradient descent with hard-label and conjugate labels. For convenience analysis, they only consider the binary classification problem. The theoretical results and analysis are interesting.
[a] ... |
The paper presents and extension of the HFGI method for GAN-Inversion, named WaGI. The author propose to replace the StyleGAN generator with a SWAGAN generator, and apply Wavelet loss on the HFGI ADA's module. In addition, they propose a Wavelet fusion method to fuse the into the latent \w. The authors suggest that pre... | Recommendation: 8: accept, good paper | Area: Generative models | Review:
The paper presents and extension of the HFGI method for GAN-Inversion, named WaGI. The author propose to replace the StyleGAN generator with a SWAGAN generator, and apply Wavelet loss on the HFGI ADA's module. In addition, they propose a Wavelet fusion method to fuse the into the latent \w. The authors suggest ... |
The authors introduce a new problem of mobile construction, where an autonomous agent must construct a grid world according to a design (target grid world state) that is fed as input to the agent. To tackle this problem, the authors de-couple the approach into (1) learning to localize the agent via an L-Net, and (2) zo... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The authors introduce a new problem of mobile construction, where an autonomous agent must construct a grid world according to a design (target grid world state) that is fed as input to the agent. To tackle this problem, the authors de-couple the approach into (1) learning to localize the agent via an L-Net, an... |
This paper improves BEIT by introducing the Vector-Quantized Knowledge Distillation (VQ-KD) algorithm for better visual tokenizer training. A patch aggregation strategy is also introduced into the Masked-Image-Modeling (MIM) pretraining framework. Experimental results show that BEITv2 (this work) significantly outperfo... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper improves BEIT by introducing the Vector-Quantized Knowledge Distillation (VQ-KD) algorithm for better visual tokenizer training. A patch aggregation strategy is also introduced into the Masked-Image-Modeling (MIM) pretraining framework. Experimental results show that BEITv2 (this work) significantly ... |
The paper proposes a modification to the standard normalizing flow training loss to account for certain dependencies in data. Specifically, the authors consider the special case where datapoints are dependent in latent space, but are transformed pointwise to the observed space. In addition, the authors specialize to th... | Recommendation: 3: reject, not good enough | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The paper proposes a modification to the standard normalizing flow training loss to account for certain dependencies in data. Specifically, the authors consider the special case where datapoints are dependent in latent space, but are transformed pointwise to the observed space. In addition, the authors speciali... |
This paper proposes model ensemble for few-shot prompt tuning in knowledge transfer tasks. The paper use an attention module to do the sample-specific ensemble for different tasks' soft prompt and apply the new ensemble of logits to the new task. Numerical experiments are presented to show the effectiveness of the idea... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes model ensemble for few-shot prompt tuning in knowledge transfer tasks. The paper use an attention module to do the sample-specific ensemble for different tasks' soft prompt and apply the new ensemble of logits to the new task. Numerical experiments are presented to show the effectiveness of ... |
This work devised a novel framework called *Analogical Networks* for *few-shot* 3D object part parsing. The framework works as follows: (1) given an input point cloud (of novel object type), the top k closest labeled object instances are retrieved from memory set (2) these objects and their labels are used as context i... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This work devised a novel framework called *Analogical Networks* for *few-shot* 3D object part parsing. The framework works as follows: (1) given an input point cloud (of novel object type), the top k closest labeled object instances are retrieved from memory set (2) these objects and their labels are used as c... |
The paper proposes using Fourier Neural Network (FNO) to combine Graph Shift Operator (GSO), to enable multivariate forecasting.
The architecture first obtains a Fourier transform of embedding of input and then applies it to the proposed operator FGSO, which mainly is consisted of FNO. The authors compare the proposed ... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper proposes using Fourier Neural Network (FNO) to combine Graph Shift Operator (GSO), to enable multivariate forecasting.
The architecture first obtains a Fourier transform of embedding of input and then applies it to the proposed operator FGSO, which mainly is consisted of FNO. The authors compare the p... |
This paper proposes a new method to approximate the hypervolume function with a deep neural network The network is built by using specialized layers to incorporate some symmetry properties of the hypervolume function, such as permutation and scale equivariance.
Strengths
1. A new deep neural network-based approximatio... | Recommendation: 6: marginally above the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper proposes a new method to approximate the hypervolume function with a deep neural network The network is built by using specialized layers to incorporate some symmetry properties of the hypervolume function, such as permutation and scale equivariance.
Strengths
1. A new deep neural network-based appr... |
The paper introduces a notion of curvature on hypergraphs. It is related to Ollivier's notion, which works for metric spaces with Markov chains on them. In that case, curvature is (basically) the curvature describes the contraction of the map from points x to the transition probabilities from x.
There is a "canonical... | Recommendation: 8: accept, good paper | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The paper introduces a notion of curvature on hypergraphs. It is related to Ollivier's notion, which works for metric spaces with Markov chains on them. In that case, curvature is (basically) the curvature describes the contraction of the map from points x to the transition probabilities from x.
There is a "c... |
The authors propose a recurrent spiking network-based reinforcement learning algorithm, in which the network consists of two major parts (an agent and a model). The agent is trained to act using policy gradient, while the model is trained to predict the effect's of the agent's actions (both the reward and the state tra... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The authors propose a recurrent spiking network-based reinforcement learning algorithm, in which the network consists of two major parts (an agent and a model). The agent is trained to act using policy gradient, while the model is trained to predict the effect's of the agent's actions (both the reward and the s... |
The authors of this paper ran a large amount of AlphaZero-based training runs, with many different sizes of neural networks, for two different board games: Connect Four and Pentago. Based on these experiments, the paper demonstrates for both games, very similar power laws exist that predict how the playing strength (me... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The authors of this paper ran a large amount of AlphaZero-based training runs, with many different sizes of neural networks, for two different board games: Connect Four and Pentago. Based on these experiments, the paper demonstrates for both games, very similar power laws exist that predict how the playing stre... |
This paper extends CellOT to the case of unbalanced optimal transport, which is essential to modelling most cell systems. This extension is accomplished through the formalism of semi-couplings, learning reweighted distributions which are then transported with balanced transport, in this case using the ICNN framework. A... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper extends CellOT to the case of unbalanced optimal transport, which is essential to modelling most cell systems. This extension is accomplished through the formalism of semi-couplings, learning reweighted distributions which are then transported with balanced transport, in this case using the ICNN fram... |
This work presents a universal diffusion model for speech enhancement for a wide variety of distortions (Table 3 in the appendix), in contrast to prior studies which focused on a limited set of distortion types (cf Table 2). The authors first presented a series of exploration of model architectures and auxiliary losses... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This work presents a universal diffusion model for speech enhancement for a wide variety of distortions (Table 3 in the appendix), in contrast to prior studies which focused on a limited set of distortion types (cf Table 2). The authors first presented a series of exploration of model architectures and auxiliar... |
This paper proposes a simple method, Iterative Patch Selection (IPS), which can select the image patches from a high-resolution image and meet the memory constraint. Two-stage DNN is used to produce the final result, the first part of the DNN is used to select the input patches in autoregressive fashion. The selected ... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a simple method, Iterative Patch Selection (IPS), which can select the image patches from a high-resolution image and meet the memory constraint. Two-stage DNN is used to produce the final result, the first part of the DNN is used to select the input patches in autoregressive fashion. The s... |
This paper presents a tree structured model for regression for applications in finance. The authors describe an inference procedure for fitting the structure and parameters. The authors provide empirical analysis of proposed method and its variants on their own corporate data and academic datasets.
This paper presents ... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper presents a tree structured model for regression for applications in finance. The authors describe an inference procedure for fitting the structure and parameters. The authors provide empirical analysis of proposed method and its variants on their own corporate data and academic datasets.
This paper p... |
Authors present a new method for estimating the uncertainty in
image-to-image problems. Given a model that is already trained to do
image-to-image prediction, authors train a model that predicts a mask that
only retains the accurate areas in the prediction. The inaccurate areas are
masked out, ... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
Authors present a new method for estimating the uncertainty in
image-to-image problems. Given a model that is already trained to do
image-to-image prediction, authors train a model that predicts a mask that
only retains the accurate areas in the prediction. The inaccurate areas are
mask... |
This paper introduces an approach for GNNs to jointly prune neurons (magnitude pruning) and sub-sample neighbouring nodes (importance sampling) and provides theoretical and empirical evaluation to show the improvements in sample complexity (less samples required for learning) and converge rate (less epochs required). R... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper introduces an approach for GNNs to jointly prune neurons (magnitude pruning) and sub-sample neighbouring nodes (importance sampling) and provides theoretical and empirical evaluation to show the improvements in sample complexity (less samples required for learning) and converge rate (less epochs requ... |
The authors introduce a novel approach for Preference-based Reinforcement Learning in an offline setting. In contrast to other methods, they do not learn an explicit reward function, but use information matching against a latent representation of an (approximate) optimal trajectory. The latent representation is learn v... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The authors introduce a novel approach for Preference-based Reinforcement Learning in an offline setting. In contrast to other methods, they do not learn an explicit reward function, but use information matching against a latent representation of an (approximate) optimal trajectory. The latent representation is... |
The authors present a method which employs an instantiation of Clone-structured cognitive graphs to learn structured representations of environments, which can be used downstream for transfer, inference and planning. The performance of the proposed approach is evaluated using grid-world style navigation tasks in primar... | Recommendation: 3: reject, not good enough | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
The authors present a method which employs an instantiation of Clone-structured cognitive graphs to learn structured representations of environments, which can be used downstream for transfer, inference and planning. The performance of the proposed approach is evaluated using grid-world style navigation tasks i... |
The paper tackles the problem of using Large Pre-trained Language Models (LLMs) for multiple-choice question-answering. Instead of using the standard cloze formulation, the paper suggests presenting the question and answer choices to the model and have the model output the answer symbol (e.g., A, B,C, ...etc.). The aut... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper tackles the problem of using Large Pre-trained Language Models (LLMs) for multiple-choice question-answering. Instead of using the standard cloze formulation, the paper suggests presenting the question and answer choices to the model and have the model output the answer symbol (e.g., A, B,C, ...etc.).... |
This paper presents a robust multiclass classification "add on" scheme for deep neural networks called TAC (total activation classifiers) in which multiple hidden layers of a network are trained to output error correcting codes (ECOCs) that are used to identify the proper class of an example and guard against spurious... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper presents a robust multiclass classification "add on" scheme for deep neural networks called TAC (total activation classifiers) in which multiple hidden layers of a network are trained to output error correcting codes (ECOCs) that are used to identify the proper class of an example and guard against ... |
The paper proposed a differentiable optical flow data generation pipeline and a loss function to drive the pipeline. The proposed modules enable automatic and efficient synthesis of a dataset effectively to a target domain, given a snippet of target data. This distinctiveness is achieved by proposing an efficient data ... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper proposed a differentiable optical flow data generation pipeline and a loss function to drive the pipeline. The proposed modules enable automatic and efficient synthesis of a dataset effectively to a target domain, given a snippet of target data. This distinctiveness is achieved by proposing an efficie... |
This paper proposes GLGExplainer (Global Logic-based GNN Explainer), which is aimed at capturing the behavior of the model as a whole, abstracting individual noisy local explanations in favor of a single robust overview of the GNN model by generating explanations as arbitrary Boolean combinations of learned human-under... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper proposes GLGExplainer (Global Logic-based GNN Explainer), which is aimed at capturing the behavior of the model as a whole, abstracting individual noisy local explanations in favor of a single robust overview of the GNN model by generating explanations as arbitrary Boolean combinations of learned hum... |
The paper finds that when backdoors exist in transformer models, particularly BERT, attention concentrates more on trigger tokens compared to clean tokens. They leverage this observation to improve the sample efficiency of backdoor attacks by adding an explicit loss term that encourages this observed behavior. Empirica... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper finds that when backdoors exist in transformer models, particularly BERT, attention concentrates more on trigger tokens compared to clean tokens. They leverage this observation to improve the sample efficiency of backdoor attacks by adding an explicit loss term that encourages this observed behavior. ... |
The paper makes two observations regarding training large language models:
1. During fine-tuning large transformers, several authors have previously shown that it is necessary to train a small number of parameters. This paper proposes using K-FAC to fine tune such models. The authors say that second order methods are v... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper makes two observations regarding training large language models:
1. During fine-tuning large transformers, several authors have previously shown that it is necessary to train a small number of parameters. This paper proposes using K-FAC to fine tune such models. The authors say that second order metho... |
This paper proposes MECTA in order to improve out-of-distribution model accuracy via computation-efficient online test-time gradient descents in a memory economic manner. The key idea behind MECTA is to reduce batch sizes, adopt an adaptive normalization layer to maintain stable and accurate predictions, and stop the b... | Recommendation: 8: accept, good paper | Area: Unsupervised and Self-supervised learning | Review:
This paper proposes MECTA in order to improve out-of-distribution model accuracy via computation-efficient online test-time gradient descents in a memory economic manner. The key idea behind MECTA is to reduce batch sizes, adopt an adaptive normalization layer to maintain stable and accurate predictions, and st... |
This work introduces curriculum learning in training selective neural networks.
- Selective Neural Network, also referred to as the abstaining classifier, have the option to reject an input, i.e., abstain from prediction. Typically, this abstention mechanism is learnt to be a measure of the difficulty of the input.
-... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This work introduces curriculum learning in training selective neural networks.
- Selective Neural Network, also referred to as the abstaining classifier, have the option to reject an input, i.e., abstain from prediction. Typically, this abstention mechanism is learnt to be a measure of the difficulty of the i... |
This paper studies the deep regression problem and proposes a new automated label encoding learning framework along with two regularizers. In specific, the authors relax the assumption of binarized label encodings of the BEL method and propose to search for continuous label embeddings. Moreover, two regularizers are in... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper studies the deep regression problem and proposes a new automated label encoding learning framework along with two regularizers. In specific, the authors relax the assumption of binarized label encodings of the BEL method and propose to search for continuous label embeddings. Moreover, two regularizer... |
This paper proposes a semi-parametric approach which trains neural network to recognize in-scope and out-scope data, and generating prompts that enables editing the response of the base model. The result is on par with state-of-the-art in most cases, while the training time and inference memory seem better.
The propose... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a semi-parametric approach which trains neural network to recognize in-scope and out-scope data, and generating prompts that enables editing the response of the base model. The result is on par with state-of-the-art in most cases, while the training time and inference memory seem better.
The... |
The authors propose a technique to learn local and global visual representation via Persistent Homology.
The overall structure of the proposed method is to perform self-supervised learning (regression) on synthetic images for Persistent Homology as a pre-training, and then natural images for fine-tuning.
The synthetic... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
The authors propose a technique to learn local and global visual representation via Persistent Homology.
The overall structure of the proposed method is to perform self-supervised learning (regression) on synthetic images for Persistent Homology as a pre-training, and then natural images for fine-tuning.
The s... |
The authors used a hierarchical VAE to learn representations from natural images and leverage their latent space hierarchy to learn voxel-to-image mappings. They showed that mapping V1/V2 responses to the early layer and higher visual areas to the deep layer of the latent hierarchy allow them to achieve superior image ... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The authors used a hierarchical VAE to learn representations from natural images and leverage their latent space hierarchy to learn voxel-to-image mappings. They showed that mapping V1/V2 responses to the early layer and higher visual areas to the deep layer of the latent hierarchy allow them to achieve superio... |
This paper provides a scalable way to learn OT maps via a deep neural network. It is especially relevant due to the recent approaches demonstrating the use of OT maps for generative purposes, as opposed to the previous methods where OT was used as a loss when training generators.
The paper is very well-written and ped... | Recommendation: 8: accept, good paper | Area: Generative models | Review:
This paper provides a scalable way to learn OT maps via a deep neural network. It is especially relevant due to the recent approaches demonstrating the use of OT maps for generative purposes, as opposed to the previous methods where OT was used as a loss when training generators.
The paper is very well-written... |
The paper presents various ways of using Self-Supervised pre-training (SSP) using public data in the context of differentially private learning. In particular, the paper claims to provide suggestions on what is the best approach when varying **amounts** and **types** of public data are available for SSP.
## Strength
*... | Recommendation: 3: reject, not good enough | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper presents various ways of using Self-Supervised pre-training (SSP) using public data in the context of differentially private learning. In particular, the paper claims to provide suggestions on what is the best approach when varying **amounts** and **types** of public data are available for SSP.
## Str... |
This paper proposes a novel NTK-based GAN named GA-NTK. Specifically, the authors use an infinite-wide linear neural network (i.e., NTK model) as the discriminator for GAN, which thus enables one to directly obtain a closed-form solution for the inner maximization problem in GAN training. Compared with vanilla GAN, GA-... | Recommendation: 8: accept, good paper | Area: Generative models | Review:
This paper proposes a novel NTK-based GAN named GA-NTK. Specifically, the authors use an infinite-wide linear neural network (i.e., NTK model) as the discriminator for GAN, which thus enables one to directly obtain a closed-form solution for the inner maximization problem in GAN training. Compared with vanilla ... |
The authors propose a new image harmonization method. Image harmonization aims at compositing regions of different images in a way the observer can not tell they were coming from different original images. The main novelty of this work is the allowance of interaction by the user. In the proposed model the user can sele... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
The authors propose a new image harmonization method. Image harmonization aims at compositing regions of different images in a way the observer can not tell they were coming from different original images. The main novelty of this work is the allowance of interaction by the user. In the proposed model the user ... |
This manuscript proposes Fourier PINNs, a variant of PINN that exactly satisfies a simple boundary condition. In addition to the reinforcement of boundary conditions, this manuscript also did a Fourier analysis of the proposed neural architecture and proposed an efficient implementation of Fourier PINNs that has a bett... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This manuscript proposes Fourier PINNs, a variant of PINN that exactly satisfies a simple boundary condition. In addition to the reinforcement of boundary conditions, this manuscript also did a Fourier analysis of the proposed neural architecture and proposed an efficient implementation of Fourier PINNs that ha... |
This paper proposes RDM theory, which is applicable to different asymmetric designs (with and without the predictor), and can serve as a unified understanding of existing non-contrastive learning methods. Besides, the RDM theory also provides practical guidelines for designing many new non-contrastive variants. RDM ach... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper proposes RDM theory, which is applicable to different asymmetric designs (with and without the predictor), and can serve as a unified understanding of existing non-contrastive learning methods. Besides, the RDM theory also provides practical guidelines for designing many new non-contrastive variants.... |
This paper proposed reduced-precision quantization methods for efficient deep neural network training. One of the methods mainly proposed in this paper parameterized the number of mantissa bits so that the number of mantissa bits can be automatically adjusted during training (Quantum-Mantissa). The authors also propose... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper proposed reduced-precision quantization methods for efficient deep neural network training. One of the methods mainly proposed in this paper parameterized the number of mantissa bits so that the number of mantissa bits can be automatically adjusted during training (Quantum-Mantissa). The authors also... |
This paper presents a latent dynamics learning (or system identification) scenario from very high dimensional observations (images). Inspired by the Koopman operator theory, the authors try to find a latent space where the dynamics is linear. and use this dynamics to control two simple dynamical systems (pendulum toss-... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper presents a latent dynamics learning (or system identification) scenario from very high dimensional observations (images). Inspired by the Koopman operator theory, the authors try to find a latent space where the dynamics is linear. and use this dynamics to control two simple dynamical systems (pendul... |
Message Passing Neural Network (MPNN) is a simple yet efficient class of Graph Neural Networks (GNNs), which have been widely adopted in many previous works. However, it is also known that MPNN has limited representational power due to its simple structure. This paper, in particular, considers subgraph methods. The aut... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
Message Passing Neural Network (MPNN) is a simple yet efficient class of Graph Neural Networks (GNNs), which have been widely adopted in many previous works. However, it is also known that MPNN has limited representational power due to its simple structure. This paper, in particular, considers subgraph methods.... |
This paper extends the notion of conditional coding in DMC and Sheng et al., 2021 by additionally introducing a masked image transformer-based entropy model (MIMT) for learned video compression. The MIMT appears to be inspired by VCT and improves on VCT by a scheduling-based transmission. The gain of the proposed metho... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper extends the notion of conditional coding in DMC and Sheng et al., 2021 by additionally introducing a masked image transformer-based entropy model (MIMT) for learned video compression. The MIMT appears to be inspired by VCT and improves on VCT by a scheduling-based transmission. The gain of the propos... |
The authors propose to represent first-order logical queries as query graphs and then design a novel message passing-based framework LMPNN on the graphs to answer given queries. This is a successful attempt to combine the power of pretrained knowledge graph embeddings and message passing to perform logical reasoning. T... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The authors propose to represent first-order logical queries as query graphs and then design a novel message passing-based framework LMPNN on the graphs to answer given queries. This is a successful attempt to combine the power of pretrained knowledge graph embeddings and message passing to perform logical reas... |
This paper questions the common sense that correlation between pre-training validation loss such as masked language modeling loss and generalization performance of downstream tasks after fine-tuning. It empirically shows that the models with the same pre-training loss can achieve varying test accuracy on downstream tas... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper questions the common sense that correlation between pre-training validation loss such as masked language modeling loss and generalization performance of downstream tasks after fine-tuning. It empirically shows that the models with the same pre-training loss can achieve varying test accuracy on downst... |
The authors propose a variant of interval bound propagation (IBP) for certified training using a small region around a preliminary center, e.g. from a PGD attack. Following a theoretical analysis of the employed box propagation, paying special attention to the role of ReLU activations, the authors present a set of expe... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The authors propose a variant of interval bound propagation (IBP) for certified training using a small region around a preliminary center, e.g. from a PGD attack. Following a theoretical analysis of the employed box propagation, paying special attention to the role of ReLU activations, the authors present a set... |
This paper proposes an efficient transformer architecture, a skeleton transformer (SKTformer), for modeling long sequence data. It contains two main components: a smoothing block to mix information over long sequences through Fourier convolution, and a matrix sketch method that simultaneously selects columns and rows f... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes an efficient transformer architecture, a skeleton transformer (SKTformer), for modeling long sequence data. It contains two main components: a smoothing block to mix information over long sequences through Fourier convolution, and a matrix sketch method that simultaneously selects columns an... |
This paper studies how deep networks generalize the concept of similarity in the presence of noise. Specifically, two phenomena are studied (1) double descent behavior and (2) online/offline relations. The double descent behavior also exhibits in the cases of contrastive learning with noise. Besides, the equivalence be... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
This paper studies how deep networks generalize the concept of similarity in the presence of noise. Specifically, two phenomena are studied (1) double descent behavior and (2) online/offline relations. The double descent behavior also exhibits in the cases of contrastive learning with noise. Besides, the equiva... |
This paper considers the problem of reinforcement learning in the presence of unmeasured confounders. In particular, the authors consider a setting where the data is collected according to a behavioral policy $\pi^b: S \times U \rightarrow \Delta(A)$ but the variables in $U$ are unobserved. The goal is to learn a "supe... | Recommendation: 3: reject, not good enough | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper considers the problem of reinforcement learning in the presence of unmeasured confounders. In particular, the authors consider a setting where the data is collected according to a behavioral policy $\pi^b: S \times U \rightarrow \Delta(A)$ but the variables in $U$ are unobserved. The goal is to learn... |
This paper proposed a method to learn hierarchical information for image classification. The proposed method introduces prompt tokens into the intermediate layers in ViT, and these prompt tokens are used to make a prediction about the coarse class of the input image. This allows the prompts to learn coarse class inform... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposed a method to learn hierarchical information for image classification. The proposed method introduces prompt tokens into the intermediate layers in ViT, and these prompt tokens are used to make a prediction about the coarse class of the input image. This allows the prompts to learn coarse clas... |
This paper investigated pretraining a mask language model in a resource-constrained setting, i.e. a single GPU for one day. The authors empirically tested various architectural and data changes in order to maximize performance. There are some interesting findings, such as per-gradient efficiency only depends on model s... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper investigated pretraining a mask language model in a resource-constrained setting, i.e. a single GPU for one day. The authors empirically tested various architectural and data changes in order to maximize performance. There are some interesting findings, such as per-gradient efficiency only depends on... |
The paper proposes a transformer-based dynamics model that can be trained on top of pre-trained slot representation of a video. The model first pre-trains SAVi/STEVE to obtain slots from video frames. It then trains an auto-regressive transformer to learn to predict future slots given the past slots.
### Pros
1. The m... | Recommendation: 8: accept, good paper | Area: Unsupervised and Self-supervised learning | Review:
The paper proposes a transformer-based dynamics model that can be trained on top of pre-trained slot representation of a video. The model first pre-trains SAVi/STEVE to obtain slots from video frames. It then trains an auto-regressive transformer to learn to predict future slots given the past slots.
### Pros
... |
SP (Stochastic Polyak step size) can be interpreted as a method specialized to interpolated models since it solves the interpolation equations. SP can be interpreted as a projection into a stochastic-constrained linearization of the objective function. The main idea of this paper is that it extends this constrained lin... | Recommendation: 5: marginally below the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
SP (Stochastic Polyak step size) can be interpreted as a method specialized to interpolated models since it solves the interpolation equations. SP can be interpreted as a projection into a stochastic-constrained linearization of the objective function. The main idea of this paper is that it extends this constra... |
The authors develop a deep learning method on molecules to better leverage training data and contextual information to make property predictions.
I really appreciate the implementation of the baselines, as well as the authors’ interpretation of them (Frequent Hitters and Similarity Search). Upon release of code, it is ... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The authors develop a deep learning method on molecules to better leverage training data and contextual information to make property predictions.
I really appreciate the implementation of the baselines, as well as the authors’ interpretation of them (Frequent Hitters and Similarity Search). Upon release of code... |
The paper presents a framework for scalable 3D object-centric learning. Unlike existing works that are limited to the bounded scene, this work allows to model 3D objects present in the large-scale 3D scenes. Specifically, the method mains a cognitive Map that allows the registration and querying of objects on a global ... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper presents a framework for scalable 3D object-centric learning. Unlike existing works that are limited to the bounded scene, this work allows to model 3D objects present in the large-scale 3D scenes. Specifically, the method mains a cognitive Map that allows the registration and querying of objects on a... |
This paper proposes diffuseq – a diffusion model for sequence-to-sequence text generation tasks. The x and y pairs are concatenated together and sent to the forward diffusion process. Different with diffusion-lm’s classifier-guided diffusion, diffuseq here uses classifier-free diffusion guided by points in space. Simi... | Recommendation: 8: accept, good paper | Area: Generative models | Review:
This paper proposes diffuseq – a diffusion model for sequence-to-sequence text generation tasks. The x and y pairs are concatenated together and sent to the forward diffusion process. Different with diffusion-lm’s classifier-guided diffusion, diffuseq here uses classifier-free diffusion guided by points in spac... |
The paper presents an approach for tabular anomaly detection which is based on Partial Identification (PID) and on Generalized Additive Models (GAM) and extensions. The method works also in semisupervised settings and compares well with alternatives, as shown in a quite extensive experimental evaluation.
Positive point... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
The paper presents an approach for tabular anomaly detection which is based on Partial Identification (PID) and on Generalized Additive Models (GAM) and extensions. The method works also in semisupervised settings and compares well with alternatives, as shown in a quite extensive experimental evaluation.
Positi... |
The paper presents a new method (SPG) for task-incremental learning. SPG uses soft-masks to condition the full model for each task. Experiment results on benchmark datasets demonstrate the effectiveness of the proposed method.
The main strength of the paper is that it is relatively easy to follow and understand the hig... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The paper presents a new method (SPG) for task-incremental learning. SPG uses soft-masks to condition the full model for each task. Experiment results on benchmark datasets demonstrate the effectiveness of the proposed method.
The main strength of the paper is that it is relatively easy to follow and understand... |
The paper considers optimization and generalization properties one-hidden-layer ReLU networks trained in the NTK regime, with an initialization of the bias that is fixed to a non-zero constant $B \geq 0$. This leads to networks with sparse activation patterns, which may be useful for sparsifying network parameters afte... | Recommendation: 5: marginally below the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The paper considers optimization and generalization properties one-hidden-layer ReLU networks trained in the NTK regime, with an initialization of the bias that is fixed to a non-zero constant $B \geq 0$. This leads to networks with sparse activation patterns, which may be useful for sparsifying network paramet... |
This paper proposes an interesting approach to design autoencoders based on convolutional sparse coding. In this framework, each layer-wise operation is replaced by an approximate solution to a Lasso problem, with a convolutional dictionary. The reconstruction ('decoding') from the so-obtained sparse codes is given sim... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
This paper proposes an interesting approach to design autoencoders based on convolutional sparse coding. In this framework, each layer-wise operation is replaced by an approximate solution to a Lasso problem, with a convolutional dictionary. The reconstruction ('decoding') from the so-obtained sparse codes is g... |
This paper proposes Adversarial Counterfactual Attention (ACAT), which first generates regions of interest (ROI) in (medical) images with saliency (heat)map techniques, and then uses these saliency map ROIs to generate soft spatial attention masks at various scales, that are integrated into a deep learning classifier. ... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes Adversarial Counterfactual Attention (ACAT), which first generates regions of interest (ROI) in (medical) images with saliency (heat)map techniques, and then uses these saliency map ROIs to generate soft spatial attention masks at various scales, that are integrated into a deep learning clas... |
The paper presents an interesting study on disentangled representation learning for domain generalization. Technically, the main idea is to build a two-branch network, one for target classification while the other for domain classification. The paper uses carefully designed experiments to demonstrate that separating th... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The paper presents an interesting study on disentangled representation learning for domain generalization. Technically, the main idea is to build a two-branch network, one for target classification while the other for domain classification. The paper uses carefully designed experiments to demonstrate that separ... |
This work explored source-free domain adaptation setting (SF-UDA). Different from conventional UDA scenario, SF-UDA assumes that model adaptation of target domain only accesses well-trained source model without well-labeled source data.
This work explored source-free domain adaptation setting (SF-UDA). Different from c... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This work explored source-free domain adaptation setting (SF-UDA). Different from conventional UDA scenario, SF-UDA assumes that model adaptation of target domain only accesses well-trained source model without well-labeled source data.
This work explored source-free domain adaptation setting (SF-UDA). Differen... |
This paper presents an approach to detect out-of-distribution (ODD) samples by an ensemble of pre-trained models. For a given test sample, OOD scores are computed based on a set of pre-trained models and a suitable p-value is determined whether the sample is OOD while maintaining a given true positive rate. experiments... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper presents an approach to detect out-of-distribution (ODD) samples by an ensemble of pre-trained models. For a given test sample, OOD scores are computed based on a set of pre-trained models and a suitable p-value is determined whether the sample is OOD while maintaining a given true positive rate. exp... |
This study addresses Multivariate time series forecasting using a transformer-based model. Namely, the authors presented Crossformer, a transformer based model which not only takes into account the temporal sequence, but also the cross dimension dependencies among the variables.
The inputs for Crossformer, is embedded ... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This study addresses Multivariate time series forecasting using a transformer-based model. Namely, the authors presented Crossformer, a transformer based model which not only takes into account the temporal sequence, but also the cross dimension dependencies among the variables.
The inputs for Crossformer, is e... |
This paper presents an initialization scheme for convolutional filters. The main idea is to first build a filter covariance matrix from pre-trained networks and then sample initial filters based on this covariance. The paper also provides a learning-free closed-form alternative for filter initialization. The experiment... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper presents an initialization scheme for convolutional filters. The main idea is to first build a filter covariance matrix from pre-trained networks and then sample initial filters based on this covariance. The paper also provides a learning-free closed-form alternative for filter initialization. The ex... |
Interesting paper, it uses language and demonstration conditioning to improve learning in an behaviour cloning set up for robot manipulation. The paper proposes:
1. a new architecture for training an eval conditioned on language and demonstrations together
2. some level of generalization with this
Strengths:
- pretrain... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
Interesting paper, it uses language and demonstration conditioning to improve learning in an behaviour cloning set up for robot manipulation. The paper proposes:
1. a new architecture for training an eval conditioned on language and demonstrations together
2. some level of generalization with this
Strengths:
- ... |
This paper proposes Dynamic Latent Hierarchy, a latent-variable video prediction model. This model can learn a hierarchy of latent variables where the latent variable at each level of the hierarchy, operates at a different timescale. Unlike related work, ClockWork VAE’s, (Saxena et al. 2021), whether or not a latent va... | Recommendation: 3: reject, not good enough | Area: Generative models | Review:
This paper proposes Dynamic Latent Hierarchy, a latent-variable video prediction model. This model can learn a hierarchy of latent variables where the latent variable at each level of the hierarchy, operates at a different timescale. Unlike related work, ClockWork VAE’s, (Saxena et al. 2021), whether or not a l... |
The paper aims to propose two architectural contributions and a new task for semantic audio-visual embodied navigation.
The paper first employs a knowledge graph for encoding object-object, object-region, and region-region relations. Then, the paper introduces multiple auxiliary models to facilitate audio-visual embo... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper aims to propose two architectural contributions and a new task for semantic audio-visual embodied navigation.
The paper first employs a knowledge graph for encoding object-object, object-region, and region-region relations. Then, the paper introduces multiple auxiliary models to facilitate audio-vis... |
The authors propose a symbolic-regression encapsulating approach to discovery of partial differential equations and a heuristic approximate solver of the constrained least-squares problem arising from the definition of learning PDEs as an optimization problem. The approach is evaluated and compared with related approac... | Recommendation: 8: accept, good paper | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The authors propose a symbolic-regression encapsulating approach to discovery of partial differential equations and a heuristic approximate solver of the constrained least-squares problem arising from the definition of learning PDEs as an optimization problem. The approach is evaluated and compared with related... |
This paper delivers a method called group mask model learning (GMML). GMML learns a representation of an image by partially corrupting it and reconstructing it with a transformer encoder-decoder structure. The authors conducted experiments that show GMML’s superiority in learning representations from small-scale datase... | Recommendation: 3: reject, not good enough | Area: Unsupervised and Self-supervised learning | Review:
This paper delivers a method called group mask model learning (GMML). GMML learns a representation of an image by partially corrupting it and reconstructing it with a transformer encoder-decoder structure. The authors conducted experiments that show GMML’s superiority in learning representations from small-scal... |
The paper proposes a novel approach Stein Variational Goal Generation (SVGG) to build on recent automatic curriculum learning techniques to address the difficulty of discontinuities in the state or goal spaces for goal-reaching tasks. Experiments show that SVGG achieves state-of-the-art results for hard-exploration RL ... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper proposes a novel approach Stein Variational Goal Generation (SVGG) to build on recent automatic curriculum learning techniques to address the difficulty of discontinuities in the state or goal spaces for goal-reaching tasks. Experiments show that SVGG achieves state-of-the-art results for hard-explora... |
The paper proposes a pretraining method to encode 3D small molecules into atom-level encodings which can be useful for downstream prediction tasks. The method is essentially to start with the 3D coordinates of a molecule, add random Gaussian noise to the atom coordinates, and train a 3D GNN to predict the added noise. ... | Recommendation: 8: accept, good paper | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper proposes a pretraining method to encode 3D small molecules into atom-level encodings which can be useful for downstream prediction tasks. The method is essentially to start with the 3D coordinates of a molecule, add random Gaussian noise to the atom coordinates, and train a 3D GNN to predict the added... |
This paper proposes a new method named sharpness-aware and reliable entropy minimization (SAR) to stabilize wild TTA. Specifically, to mitigate the collapse of group norm or layer norm models, SAR first filters samples with high and noisy gradients according to entropy loss. Furthermore, SAR introduces a sharpness-awar... | Recommendation: 8: accept, good paper | Area: Unsupervised and Self-supervised learning | Review:
This paper proposes a new method named sharpness-aware and reliable entropy minimization (SAR) to stabilize wild TTA. Specifically, to mitigate the collapse of group norm or layer norm models, SAR first filters samples with high and noisy gradients according to entropy loss. Furthermore, SAR introduces a sharpn... |
This work introduces Dateformer for time-series forecasting. The time-series is split into day based patches and the processing is shifted from time-based towards patch based. This work uses time representations as the modeling entity and empirically demonstrates the strength of the proposed approach on multiple datase... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This work introduces Dateformer for time-series forecasting. The time-series is split into day based patches and the processing is shifted from time-based towards patch based. This work uses time representations as the modeling entity and empirically demonstrates the strength of the proposed approach on multipl... |
The paper proposes a Gaussian Process (GP)-based approach to meta-learning. Using GP, the method allows probabilistic predictions for in-distribution tasks and detection for out-of-distribution (OoD) tasks. Several practical algorithm variants are introduced to tackle computational requirements of GP. Empirically, the ... | Recommendation: 5: marginally below the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The paper proposes a Gaussian Process (GP)-based approach to meta-learning. Using GP, the method allows probabilistic predictions for in-distribution tasks and detection for out-of-distribution (OoD) tasks. Several practical algorithm variants are introduced to tackle computational requirements of GP. Empirical... |
The paper proposes to reweight episodes in the replay by their advantage values. The paper claims that doing so is especially useful when there are many more low return episodes in the replay buffer compared to high return episodes.
Strength
The proposed idea is intuitive and the claim that the idea is useful when th... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper proposes to reweight episodes in the replay by their advantage values. The paper claims that doing so is especially useful when there are many more low return episodes in the replay buffer compared to high return episodes.
Strength
The proposed idea is intuitive and the claim that the idea is useful... |
This paper extend conventional topic model methods to contextualized word embeddings from BERT-like pre-trained models. The authors propose several modifications including using attention weights as continuous word counts and modeling and considering a contextualized word embeddings to be drawn from a Gaussian distribu... | Recommendation: 3: reject, not good enough | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
This paper extend conventional topic model methods to contextualized word embeddings from BERT-like pre-trained models. The authors propose several modifications including using attention weights as continuous word counts and modeling and considering a contextualized word embeddings to be drawn from a Gaussian ... |
This paper presents a new method to recover the model after an attack by finetuning only one layer. Moreover, it has assumed that the backdoor model tends to be trapped in the local minimum. Therefore, a new purification technique named NGF is invented based on the loss surface curvature matrix, i.e., Fisher Informatio... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper presents a new method to recover the model after an attack by finetuning only one layer. Moreover, it has assumed that the backdoor model tends to be trapped in the local minimum. Therefore, a new purification technique named NGF is invented based on the loss surface curvature matrix, i.e., Fisher In... |
This paper presents a deep model for learning to predict the sound that would be heard at a given target position, assuming that a reference position and the sound heard at the reference position are given.
The authors presents a MLP for learning forward RIR and inverse RIR, where the reference sound is first project... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper presents a deep model for learning to predict the sound that would be heard at a given target position, assuming that a reference position and the sound heard at the reference position are given.
The authors presents a MLP for learning forward RIR and inverse RIR, where the reference sound is first... |
This paper points out some potential issues and limitations using “pushforward” models for density estimation. The main issue is that the general manifold of dimension m may not be effectively embedded in the latent space of the same dimension m. To resolve that challenge, the authors propose to construct an implicit m... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
This paper points out some potential issues and limitations using “pushforward” models for density estimation. The main issue is that the general manifold of dimension m may not be effectively embedded in the latent space of the same dimension m. To resolve that challenge, the authors propose to construct an im... |
The paper proposes a study on a different flavor of continual learning scenarios than previously considered in the literature. Precisely, one in which data separation is not strict among concepts, and later tasks can revisit one of the previously introduced classes. The authors name this "class incremental with repetit... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The paper proposes a study on a different flavor of continual learning scenarios than previously considered in the literature. Precisely, one in which data separation is not strict among concepts, and later tasks can revisit one of the previously introduced classes. The authors name this "class incremental with... |
This work studies the problem of learning with spurious correlations whose correlation with labels can change between the training and test distributions. They propose to regularize the algorithm with a new regularizer called Conditional Spurious Variation (CSV) which essentially measures the difference in loss values ... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This work studies the problem of learning with spurious correlations whose correlation with labels can change between the training and test distributions. They propose to regularize the algorithm with a new regularizer called Conditional Spurious Variation (CSV) which essentially measures the difference in loss... |
This paper proposes a curriculum learning for language model pretraining: training with high-frequency tokens first, and low-frequency tokens later. They are 4 stages of pretraining:
- 1. First 3 stages: training with the text whose low-frequency tokens are replaced with their syntactic label, for example, NP for noun,... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a curriculum learning for language model pretraining: training with high-frequency tokens first, and low-frequency tokens later. They are 4 stages of pretraining:
- 1. First 3 stages: training with the text whose low-frequency tokens are replaced with their syntactic label, for example, NP f... |
The paper reformulates intent detection as a retrieval task, by treating utterances and intent names as questions and answers respectively. The authors leverage a dual-encoder based retrieval architecture for few-shot intent detection, with late-interaction scores (as used in ColBERT) and batch contrastive training.
S... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper reformulates intent detection as a retrieval task, by treating utterances and intent names as questions and answers respectively. The authors leverage a dual-encoder based retrieval architecture for few-shot intent detection, with late-interaction scores (as used in ColBERT) and batch contrastive trai... |
This paper proposes a new way to make experience replay more efficient. Inspired by episodic learning, the authors treat the replay memory as an empirical replay memory MDP (RM-MDP). With dynamic programming, conservative value esimtate is learned by only considering transitions observed in the replay memory. Based on ... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes a new way to make experience replay more efficient. Inspired by episodic learning, the authors treat the replay memory as an empirical replay memory MDP (RM-MDP). With dynamic programming, conservative value esimtate is learned by only considering transitions observed in the replay memory. B... |
This paper studies how to incorporate uncertainty estimations for offline reinforcement learning to prevent the learner from favoring regions of high uncertainty (which are often over-estimated). The authors extend safe policy improvement with soft baseline bootstrapping (soft-SPIBB) to large state-action space, where ... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper studies how to incorporate uncertainty estimations for offline reinforcement learning to prevent the learner from favoring regions of high uncertainty (which are often over-estimated). The authors extend safe policy improvement with soft baseline bootstrapping (soft-SPIBB) to large state-action space... |
This paper proposes Attention Retractable Transformer (ART) for image restoration, which integrates both dense and sparse attention modules in the transformer-based network. The dense attention block and sparse attention blocks emerge alternately to enable interactions among tokens extracted from a sparse area during r... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes Attention Retractable Transformer (ART) for image restoration, which integrates both dense and sparse attention modules in the transformer-based network. The dense attention block and sparse attention blocks emerge alternately to enable interactions among tokens extracted from a sparse area ... |
This paper discusses representation learning with diffusion models. The key idea is from Abstreiter et al., where a diffusion model is additionally trained on conditioning information given by an encoder. The benefit of this is that is becomes possible to minimize the denoising score matching objective to zero. This pa... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper discusses representation learning with diffusion models. The key idea is from Abstreiter et al., where a diffusion model is additionally trained on conditioning information given by an encoder. The benefit of this is that is becomes possible to minimize the denoising score matching objective to zero.... |
This paper proves stability-based generalization bounds under the assumption that the data is drawn from a mixing sequence (rather than assuming that the data is i.i.d.). The bounds improve upon previous results by a factor which is the square root of the sample size.
The paper makes a good contribution to learning the... | Recommendation: 8: accept, good paper | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper proves stability-based generalization bounds under the assumption that the data is drawn from a mixing sequence (rather than assuming that the data is i.i.d.). The bounds improve upon previous results by a factor which is the square root of the sample size.
The paper makes a good contribution to lear... |
Inspired from Adam, authors propose adaptive bilevel algorithms, and a variance reduced variation, coined "biAdam" and VR-BiAdam. They show respectively $1/\epsilon^4$ and $1/\epsilon^3$ convergence rate. Authors propose experiments on data hypercleaning and hyperrepresentation learning
Major concerns:
- I have doubt ... | Recommendation: 5: marginally below the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
Inspired from Adam, authors propose adaptive bilevel algorithms, and a variance reduced variation, coined "biAdam" and VR-BiAdam. They show respectively $1/\epsilon^4$ and $1/\epsilon^3$ convergence rate. Authors propose experiments on data hypercleaning and hyperrepresentation learning
Major concerns:
- I hav... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.