review stringlengths 5 16.6k | score stringclasses 5
values | area stringclasses 12
values | text stringlengths 31 5.65k |
|---|---|---|---|
This paper proposes Peaks2Image as a neural-network approach to reconstruct continuous spatial representations of brain activity from peak activation tables. This study aims to learn a decoding model by using TF-IDF features as labels so that the number of decoded terms will be much broader than in current image-based ... | Recommendation: 5: marginally below the acceptance threshold | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
This paper proposes Peaks2Image as a neural-network approach to reconstruct continuous spatial representations of brain activity from peak activation tables. This study aims to learn a decoding model by using TF-IDF features as labels so that the number of decoded terms will be much broader than in current imag... |
This paper proposes to use diffusion model to adapt selected layers of a trained model conditioned on an input sample. Authors conducted experiments onclassification, 3D construction, tablular data and speech separation.
Stength
Using diffusion model for conditional parameter generation is novel.
Weakness
1. I find t... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes to use diffusion model to adapt selected layers of a trained model conditioned on an input sample. Authors conducted experiments onclassification, 3D construction, tablular data and speech separation.
Stength
Using diffusion model for conditional parameter generation is novel.
Weakness
1. ... |
CogVideo proposes a combination of models that can generate videos from a given description. The system uses two model that generate frames in a hierarchical way: the first model generates the "key-frames" while the other model interpolates between the generated key-frames. The authors also proposed "a dual-attention m... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
CogVideo proposes a combination of models that can generate videos from a given description. The system uses two model that generate frames in a hierarchical way: the first model generates the "key-frames" while the other model interpolates between the generated key-frames. The authors also proposed "a dual-att... |
The paper considers applying bias-term fine-tuning (BiTFiT) to DP large model training. They describe efficient implementations of (DP)-BiTFiT and demonstrate that the time and space complexity of these implementations improves both over non-DP and DP full model training. They then give a deep empirical study of BiTFiT... | 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 considers applying bias-term fine-tuning (BiTFiT) to DP large model training. They describe efficient implementations of (DP)-BiTFiT and demonstrate that the time and space complexity of these implementations improves both over non-DP and DP full model training. They then give a deep empirical study o... |
This paper makes a systematic study in terms of the critical factors in temporal stereo matching for camera-only 3D object detection and proposes corresponding solutions with a framework, SOLOFusion, to address the problems. It concludes that the limited history usage and low granularity of matching resolution are the ... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper makes a systematic study in terms of the critical factors in temporal stereo matching for camera-only 3D object detection and proposes corresponding solutions with a framework, SOLOFusion, to address the problems. It concludes that the limited history usage and low granularity of matching resolution ... |
This paper aims to solve the fundamental problem --- long-tailed recognition. It analyzes SWA/SWAG with long-tailed data.
（1) Classifier re-training can make better use of SWA for long-tailed data.
(2) The positive correlation between NLL and dispersion is observed. Based on this phenomenon, a self-d... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper aims to solve the fundamental problem --- long-tailed recognition. It analyzes SWA/SWAG with long-tailed data.
（1) Classifier re-training can make better use of SWA for long-tailed data.
(2) The positive correlation between NLL and dispersion is observed. Based on this phenomenon, ... |
The paper extends Decision Transformers and Generalized Decision Transformers for skill discovery. They learn primitive skills based on unsupervised learning (i.e., without rewards information). Technically, they first use a VQ-VAE to encode discrete skills, then they utilize a causal transformer just like the original... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper extends Decision Transformers and Generalized Decision Transformers for skill discovery. They learn primitive skills based on unsupervised learning (i.e., without rewards information). Technically, they first use a VQ-VAE to encode discrete skills, then they utilize a causal transformer just like the ... |
This paper provides convergence analyses of gradient descent for learning deep, over-parameterized operator networks. The authors show the guarantees two types of networks: (1) with smooth activations and (2) with ReLU activation using two different analysis techniques.
Regarding the strengths, this paper
+ establishe... | Recommendation: 5: marginally below the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper provides convergence analyses of gradient descent for learning deep, over-parameterized operator networks. The authors show the guarantees two types of networks: (1) with smooth activations and (2) with ReLU activation using two different analysis techniques.
Regarding the strengths, this paper
+ es... |
This paper elaborates on a specific simplicity bias called LD-SB. Theoretical proofs of LD-SB are provided for 1-hidden layer neural networks on the IFM distribution under different situations. Experiments on four real datasets demonstrate that LD-SB is of practical importance. The authors further propose a new robust ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper elaborates on a specific simplicity bias called LD-SB. Theoretical proofs of LD-SB are provided for 1-hidden layer neural networks on the IFM distribution under different situations. Experiments on four real datasets demonstrate that LD-SB is of practical importance. The authors further propose a new... |
The paper presents a solution to the OOD detection problem. Starting from approaches that train the model on auxiliary data, the authors suggest that the distribution of auxiliary data used for training might differ from that to be encountered in practice limiting the detection capabilities. As such the authors propose... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The paper presents a solution to the OOD detection problem. Starting from approaches that train the model on auxiliary data, the authors suggest that the distribution of auxiliary data used for training might differ from that to be encountered in practice limiting the detection capabilities. As such the authors... |
The authors propose a solution to the problem of active domain adaptation, where limited target data is annotated to maximally benefit the model adaptation. The proposed solution called "Dirichlet-based Uncertainty Calibration (DUC)" achieves both target representativeness and mitigation of miscalibration. The authors ... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
The authors propose a solution to the problem of active domain adaptation, where limited target data is annotated to maximally benefit the model adaptation. The proposed solution called "Dirichlet-based Uncertainty Calibration (DUC)" achieves both target representativeness and mitigation of miscalibration. The ... |
This work studies the problem of finetuning policies that were previously trained on offline datasets, a challenging problem since offline pre-training often leverages pessimistic objectives, which would hurt online fine-tuning in searching for new behaviors. To approach this, it proposes O3F, a method that works on to... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This work studies the problem of finetuning policies that were previously trained on offline datasets, a challenging problem since offline pre-training often leverages pessimistic objectives, which would hurt online fine-tuning in searching for new behaviors. To approach this, it proposes O3F, a method that wor... |
This paper proposes to perform few-shot recognition, and introduces two innovations. The first is that instead of averaging support image features to get a prototype, the proposed approach meta-learns a model that produces a prototype from concatenated features. The second innovation is to perform self-training with we... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes to perform few-shot recognition, and introduces two innovations. The first is that instead of averaging support image features to get a prototype, the proposed approach meta-learns a model that produces a prototype from concatenated features. The second innovation is to perform self-training... |
The Paper studies temporal domain generalization with dynamic neural networks.
For a discrete set of datasets: D1, ..., D_T at different timestamps: t_1, ..., t_T, goal is to build a model that captures concept drift.
In particular, model learns how to predicts parameters of NN for timestamp T+1 for the mapping X_{T+... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
The Paper studies temporal domain generalization with dynamic neural networks.
For a discrete set of datasets: D1, ..., D_T at different timestamps: t_1, ..., t_T, goal is to build a model that captures concept drift.
In particular, model learns how to predicts parameters of NN for timestamp T+1 for the mappi... |
This paper proposes a text embedding that deeply represents semantic meaning,
and calls it a neural embedding.
While the conventional text embeddings use the vector output of a pre-trained language model,
the neural embedding learns the text and literally pick its brain, and takes the actual weights of the model’s neur... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a text embedding that deeply represents semantic meaning,
and calls it a neural embedding.
While the conventional text embeddings use the vector output of a pre-trained language model,
the neural embedding learns the text and literally pick its brain, and takes the actual weights of the mode... |
This paper firstly proposes a metric called SynExp that is an expectation of the SynFlow values over the weights and mask distributions. According to the authors, SynExp is a constant and independent of the pruning granularity so the pruning granularity may not have an influence on the trained models' accuracy. Later o... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper firstly proposes a metric called SynExp that is an expectation of the SynFlow values over the weights and mask distributions. According to the authors, SynExp is a constant and independent of the pruning granularity so the pruning granularity may not have an influence on the trained models' accuracy.... |
In this paper, authors proposed a new framework for human grasps generation for holding objects in hand. Instead of directly learning mapping from object point cloud to joint angles of hand, authors divide this whole problem into 2 steps. In the first stage, authors have tried to learn ContactCVAE, which given point cl... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
In this paper, authors proposed a new framework for human grasps generation for holding objects in hand. Instead of directly learning mapping from object point cloud to joint angles of hand, authors divide this whole problem into 2 steps. In the first stage, authors have tried to learn ContactCVAE, which given ... |
Maximum Manifold Capacity Representation (MMCR) is proposed as a novel self-supervised learning framework by maximize the number of linearly separable object manifolds, which is interesting to see, and different from many SSL methods are inspired by information & entropy criterions.
Experimental results on several co... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
Maximum Manifold Capacity Representation (MMCR) is proposed as a novel self-supervised learning framework by maximize the number of linearly separable object manifolds, which is interesting to see, and different from many SSL methods are inspired by information & entropy criterions.
Experimental results on se... |
Most of the prior works for blind image super-resolution design explicit degradation estimation, which is often infeasible due to lack of proper degradation labels. Also, there might not exist a universal degradation model to capture various practical degradation types, such as blur, noise, compression, and platform in... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
Most of the prior works for blind image super-resolution design explicit degradation estimation, which is often infeasible due to lack of proper degradation labels. Also, there might not exist a universal degradation model to capture various practical degradation types, such as blur, noise, compression, and pla... |
The paper proposes a method for automated prompt generation using a combination of Language Models. It shows that certain source LM combinations are more succesful than others in outperforming autoprompts trained on single ML. The paper also includes a thorough analysis of potential reasons.
Additionally, the paper... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The paper proposes a method for automated prompt generation using a combination of Language Models. It shows that certain source LM combinations are more succesful than others in outperforming autoprompts trained on single ML. The paper also includes a thorough analysis of potential reasons.
Additionally, t... |
This paper proposes a method for growing neural networks during training. The authors formulate the problem as an optimization problem. Specifically, when the network lacks expressivity, there will be no gradient updates in the weight space that will lead to a change in the outputs that align with the desired functiona... | Recommendation: 3: reject, not good enough | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper proposes a method for growing neural networks during training. The authors formulate the problem as an optimization problem. Specifically, when the network lacks expressivity, there will be no gradient updates in the weight space that will lead to a change in the outputs that align with the desired f... |
This paper studied active learning under out-of-distribution data, which could be unreliable with the existence of OOD data. This paper proposed a multi-objective loss to simultaneously control the data uncertainty and simultaneously filter the OOD data. Empirical results justified the proposed approach
### Pros
- Thi... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
This paper studied active learning under out-of-distribution data, which could be unreliable with the existence of OOD data. This paper proposed a multi-objective loss to simultaneously control the data uncertainty and simultaneously filter the OOD data. Empirical results justified the proposed approach
### Pr... |
The paper introduces Factorized-FNO, where they consider separable Fourier representation (by taking the fourier transform of each dimension separately and independently) the authors achieve more “stable” models, whose performance increases as the networks are made deeper (something that does not happen with baseline F... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper introduces Factorized-FNO, where they consider separable Fourier representation (by taking the fourier transform of each dimension separately and independently) the authors achieve more “stable” models, whose performance increases as the networks are made deeper (something that does not happen with ba... |
Given the complete point cloud of an object, how can a vision model infer the right grasp pose for a human hand such that the object can be grasped stably? The paper tries to address this issue via a staged pipeline. First, it trains a generative model (conditional VAE) to reconstruct the ground-truth contact map of th... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
Given the complete point cloud of an object, how can a vision model infer the right grasp pose for a human hand such that the object can be grasped stably? The paper tries to address this issue via a staged pipeline. First, it trains a generative model (conditional VAE) to reconstruct the ground-truth contact m... |
In this paper the authors propose a neural model to deform a template based on the single-view 3D observations while correspondences between both states are implicitly computed. To that end, the authors exploit a feature constraint deformation network on the single-view point cloud as input. The method uses well-known ... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
In this paper the authors propose a neural model to deform a template based on the single-view 3D observations while correspondences between both states are implicitly computed. To that end, the authors exploit a feature constraint deformation network on the single-view point cloud as input. The method uses wel... |
This paper presents a knowledge distillation framework for link prediction, Linkless Link Prediction (LLP). Unlike simple knowledge distillation methods that match outputs or representations of models, LLP distills relational knowledge of each link to the student MLP using proposed rank-based and distribution-based mat... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper presents a knowledge distillation framework for link prediction, Linkless Link Prediction (LLP). Unlike simple knowledge distillation methods that match outputs or representations of models, LLP distills relational knowledge of each link to the student MLP using proposed rank-based and distribution-b... |
The authors propose a saliency based detector TextShield to identify adversarial samples in NLP. They use existing methods (vanilla gradient, guided backpropagation, layerwise relevance propagation, and integration gradient) to obtain saliency maps and then combine the outputs of all the methods to produce a final pred... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The authors propose a saliency based detector TextShield to identify adversarial samples in NLP. They use existing methods (vanilla gradient, guided backpropagation, layerwise relevance propagation, and integration gradient) to obtain saliency maps and then combine the outputs of all the methods to produce a fi... |
In this paper, the authors study what makes hierarchical ViTs (in particular SwinTransformers) work, and based on the insights, propose a new vision transformer architecture called HiViT. The authors show that HiViT works better than Swin and vanilla ViT on ImageNet and ADE20K. Moreover, HiViT doesn't use window attent... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
In this paper, the authors study what makes hierarchical ViTs (in particular SwinTransformers) work, and based on the insights, propose a new vision transformer architecture called HiViT. The authors show that HiViT works better than Swin and vanilla ViT on ImageNet and ADE20K. Moreover, HiViT doesn't use windo... |
This paper provides a theoretical analysis of oversmoothing in linear GCNs in an SBM setting where the features and labels are assumed to be Gaussian. It provides bounds for shallow GNNs rather than convergence rates in the infinite-layer limit. The theoretical results are supported with empirical data showing the effe... | Recommendation: 8: accept, good paper | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper provides a theoretical analysis of oversmoothing in linear GCNs in an SBM setting where the features and labels are assumed to be Gaussian. It provides bounds for shallow GNNs rather than convergence rates in the infinite-layer limit. The theoretical results are supported with empirical data showing ... |
The paper studies empirical risk minimization with differential privacy and has two main results:
* A simple reduction from constrained to unconstrained minimization that preserves, convexity, the Lipszhitzness constant (with respect to $\ell_2$, and optimal solutions. This reduction allows extending privacy lower bou... | Recommendation: 5: marginally below the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper studies empirical risk minimization with differential privacy and has two main results:
* A simple reduction from constrained to unconstrained minimization that preserves, convexity, the Lipszhitzness constant (with respect to $\ell_2$, and optimal solutions. This reduction allows extending privacy l... |
A memory-augmented representation learning model, EventFormer, is proposed for asynchronous event-based perception. To achieve this goal, the EventFormer learns to store, retrieve and update its memory representation in the latent form of higher-order spatiotemporal dynamics of the events. The idea is interesting and a... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
A memory-augmented representation learning model, EventFormer, is proposed for asynchronous event-based perception. To achieve this goal, the EventFormer learns to store, retrieve and update its memory representation in the latent form of higher-order spatiotemporal dynamics of the events. The idea is interesti... |
In this paper, the authors proposed a physical-driven model and regularization scheme to predict energy for molecular systems. Compared to other baseline models, their full model preforms much better on force prediction, while the energy MAE is slightly lagging behind the state-of-the art MXMNet model. When applying to... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
In this paper, the authors proposed a physical-driven model and regularization scheme to predict energy for molecular systems. Compared to other baseline models, their full model preforms much better on force prediction, while the energy MAE is slightly lagging behind the state-of-the art MXMNet model. When app... |
This paper studies regret minimization for reinforcement learning (RL) in Latent Markov Decision Processes (LMDPs) with context in hindsight. Episodic undiscounted case is studied, and regret guarantees are found. The regret bound only scales logarithmically with the planning horizon. Further, the state term dependence... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper studies regret minimization for reinforcement learning (RL) in Latent Markov Decision Processes (LMDPs) with context in hindsight. Episodic undiscounted case is studied, and regret guarantees are found. The regret bound only scales logarithmically with the planning horizon. Further, the state term de... |
This paper studies the expressive power of graph neural networks from a different angle than the classical $k$-WL hierarchy. Authors argue that $k$-WL is not very well-suited for developing new graph neural network architectures since (i) these algorithms are already computationally demanding with $k\geq 2$, (ii) this ... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper studies the expressive power of graph neural networks from a different angle than the classical $k$-WL hierarchy. Authors argue that $k$-WL is not very well-suited for developing new graph neural network architectures since (i) these algorithms are already computationally demanding with $k\geq 2$, (i... |
This paper proposes a targeted adversarial training method for non-contrastive self-supervised learning (SSL) to improve the performance. Using the proposed algorithm, non-contrastive SSL gets an improvement in the robustness of the downstream tasks.
Strength:
The paper explains the intuition and numerical experiments... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper proposes a targeted adversarial training method for non-contrastive self-supervised learning (SSL) to improve the performance. Using the proposed algorithm, non-contrastive SSL gets an improvement in the robustness of the downstream tasks.
Strength:
The paper explains the intuition and numerical exp... |
This work proposes to consider the generalization ability into the acquisition process of new labeled instances in active learning. The work selects instances with a high value of loss sharpness on pseudo-labels, which can be used for improving the generalization ability of the model based on the finding of Sharpness-A... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This work proposes to consider the generalization ability into the acquisition process of new labeled instances in active learning. The work selects instances with a high value of loss sharpness on pseudo-labels, which can be used for improving the generalization ability of the model based on the finding of Sha... |
This paper raises an interesting problem, which is the miscalibrated distributions caused by using $minMPJPE$ to choose the best estimates in multi-hypothesis 3D pose estimation. To resolve the miscalibration, the paper uses the existing ECE metric proposed by Naeini et al., and proposes cGNF to learn the conditional d... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper raises an interesting problem, which is the miscalibrated distributions caused by using $minMPJPE$ to choose the best estimates in multi-hypothesis 3D pose estimation. To resolve the miscalibration, the paper uses the existing ECE metric proposed by Naeini et al., and proposes cGNF to learn the condi... |
This paper studies the effect of label error on the model’s disparity metrics (e.g., calibration, FPR, FNR) on both the training and test set. Empirically, the authors have found that label errors have a larger influence on minority groups than on majority groups. To mitigate the impact of label errors, The authors hav... | Recommendation: 5: marginally below the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper studies the effect of label error on the model’s disparity metrics (e.g., calibration, FPR, FNR) on both the training and test set. Empirically, the authors have found that label errors have a larger influence on minority groups than on majority groups. To mitigate the impact of label errors, The aut... |
This paper studies the relationship between label memorization and membership inference. It showed that examples with high memorization are more easily attacked. Based on this, it formulated membership inference attack where the attacker choose highly memorized examples, and show that the attack performance improved si... | Recommendation: 3: reject, not good enough | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper studies the relationship between label memorization and membership inference. It showed that examples with high memorization are more easily attacked. Based on this, it formulated membership inference attack where the attacker choose highly memorized examples, and show that the attack performance imp... |
The work is motivated by a problem that current continual learning metrics are mainly coarse-grained (e.g., task-based), which can lead to information loss. For instance, monitoring task-based metrics cannot show the stability gap (substantial but temporary forgetting upon learning a new task). The stability gap is def... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The work is motivated by a problem that current continual learning metrics are mainly coarse-grained (e.g., task-based), which can lead to information loss. For instance, monitoring task-based metrics cannot show the stability gap (substantial but temporary forgetting upon learning a new task). The stability ga... |
Paper proposes differentially private dataset condensation, and improvement over the ICML'22 work [1], and shows that the proposed method provides good utility.
[1] Dong, T., Zhao, B., & Lyu, L. (2022). Privacy for Free: How does Dataset Condensation Help Privacy?. arXiv preprint arXiv:2206.00240.
Paper proposes a di... | Recommendation: 6: marginally above the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
Paper proposes differentially private dataset condensation, and improvement over the ICML'22 work [1], and shows that the proposed method provides good utility.
[1] Dong, T., Zhao, B., & Lyu, L. (2022). Privacy for Free: How does Dataset Condensation Help Privacy?. arXiv preprint arXiv:2206.00240.
Paper propo... |
### Problem
The paper tackles the problem of neural network verification, which aims to give guarantees of whether a trained NN follows certain specifications.
### Proposed method
The authors claim that the current adversarial robustness specification used by the verification community is improper. Their claim is th... | Recommendation: 3: reject, not good enough | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
### Problem
The paper tackles the problem of neural network verification, which aims to give guarantees of whether a trained NN follows certain specifications.
### Proposed method
The authors claim that the current adversarial robustness specification used by the verification community is improper. Their cla... |
This paper gives a new algorithm for computing the Earth Mover's Distance between two discrete distributions in the hypercube. This is an extensively studied problem with several theoretical algorithms papers studying it. This present paper exploits some of their efficient-in-practice techniques in order to design a pr... | Recommendation: 8: accept, good paper | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper gives a new algorithm for computing the Earth Mover's Distance between two discrete distributions in the hypercube. This is an extensively studied problem with several theoretical algorithms papers studying it. This present paper exploits some of their efficient-in-practice techniques in order to des... |
This paper proposes a scalable differentially private distributed learning framework for protecting privacy over distributed dataset. The scalability of the framework is based on a scalable secure summation protocol. The privacy of the framework is achieved by aggregating the noise of non colluded users in an obliviou... | 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 proposes a scalable differentially private distributed learning framework for protecting privacy over distributed dataset. The scalability of the framework is based on a scalable secure summation protocol. The privacy of the framework is achieved by aggregating the noise of non colluded users in an ... |
This paper tries to adjust hyper parameters in two types of uncertain dataset information: 1) dataset labels are postponed to be obtained so hyper parameters need to be adjusted without complete dataset information. 2) hyper parameters are adjusted with a subset training dataset since training models with complete trai... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper tries to adjust hyper parameters in two types of uncertain dataset information: 1) dataset labels are postponed to be obtained so hyper parameters need to be adjusted without complete dataset information. 2) hyper parameters are adjusted with a subset training dataset since training models with compl... |
This paper studies the saturation effect of KRR. Under some additional assumptions, the paper provides a lower bound of the rate of generalization error for the case where $f\in[\mathcal{H}]^\alpha$ for some $\alpha>2$. The proposed lower bound matches the existing upper bound when $\alpha>2$ and verifies the saturatio... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper studies the saturation effect of KRR. Under some additional assumptions, the paper provides a lower bound of the rate of generalization error for the case where $f\in[\mathcal{H}]^\alpha$ for some $\alpha>2$. The proposed lower bound matches the existing upper bound when $\alpha>2$ and verifies the s... |
The paper proposes DepthFL, an approach that enables resource-constrained clients to participate in Federated Learning (FL) of a global model that is larger than these devices. DepthFL defines shallower versions of the global model for simultaneously training, whereby each client trains a local model of a depth appropr... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The paper proposes DepthFL, an approach that enables resource-constrained clients to participate in Federated Learning (FL) of a global model that is larger than these devices. DepthFL defines shallower versions of the global model for simultaneously training, whereby each client trains a local model of a depth... |
The authors introduce the task of Sequential Model Editing (SME) where a model needs to adapt to and fix its mistakes from a stream of incoming input-output examples of a certain machine learning task. The setting extends Model Editing (ME), where people focus on editing the model to fix a single occurrence of an obser... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The authors introduce the task of Sequential Model Editing (SME) where a model needs to adapt to and fix its mistakes from a stream of incoming input-output examples of a certain machine learning task. The setting extends Model Editing (ME), where people focus on editing the model to fix a single occurrence of ... |
The paper studies the weight pruning and weight sharing for image-text retrieval model. Contrary to previous works that applies the pruning and sharing on the layer level, the main idea of this paper is to apply weight sharing and weight pruning on the weight level. Therefore, the pruned model could achieve better comp... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper studies the weight pruning and weight sharing for image-text retrieval model. Contrary to previous works that applies the pruning and sharing on the layer level, the main idea of this paper is to apply weight sharing and weight pruning on the weight level. Therefore, the pruned model could achieve bet... |
The paper aims to efficiently replace the text encoder of the text-to-image generation model to improve the generation quality (or to make a multilingual text-to-image generative model). To this end, the authors proposed Model Translation Network (MTN) in which training loss is motivated by domain adaptation and cross-... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper aims to efficiently replace the text encoder of the text-to-image generation model to improve the generation quality (or to make a multilingual text-to-image generative model). To this end, the authors proposed Model Translation Network (MTN) in which training loss is motivated by domain adaptation an... |
The paper under consideration proposes a normalizing flow model (called JKO-Flow) which combines JKO scheme and Continuous Normalizing Flows. The previous approaches which dealt with JKO scheme parameterized the pushforward transform $T(x)$ arising in JKO objective either by ICNNs [1, 2] or general NNs of the form $F ... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
The paper under consideration proposes a normalizing flow model (called JKO-Flow) which combines JKO scheme and Continuous Normalizing Flows. The previous approaches which dealt with JKO scheme parameterized the pushforward transform $T(x)$ arising in JKO objective either by ICNNs [1, 2] or general NNs of the ... |
This paper proposes the setting of ``continual unbiased learning" (CUL), where for each task of continual learning, labels are strongly correlated with specific but different input features. It also proposes to use self-supervised learning methods to address the model bias issue during CUL.
Strength:
The setting and me... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes the setting of ``continual unbiased learning" (CUL), where for each task of continual learning, labels are strongly correlated with specific but different input features. It also proposes to use self-supervised learning methods to address the model bias issue during CUL.
Strength:
The settin... |
The paper presents a learning strategy to grow smaller pre-trained networks into larger networks with the hope that model growth can save compute compared to a network that is trained from scratch. While mostly in the literature, growth strategies are based on heuristics, e.g., copying the layers of the network for exp... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper presents a learning strategy to grow smaller pre-trained networks into larger networks with the hope that model growth can save compute compared to a network that is trained from scratch. While mostly in the literature, growth strategies are based on heuristics, e.g., copying the layers of the network... |
The authors propose an adaptive algorithm for sharpness aware minimization, a special type of regularized loss function that has a min-max structure. The adaptive algorithm is reminiscent of Adam. A convergence analysis to a first-order stationary value of the original unregularized loss is presented. The convergence o... | Recommendation: 3: reject, not good enough | Area: Optimization (eg, convex and non-convex optimization) | Review:
The authors propose an adaptive algorithm for sharpness aware minimization, a special type of regularized loss function that has a min-max structure. The adaptive algorithm is reminiscent of Adam. A convergence analysis to a first-order stationary value of the original unregularized loss is presented. The conve... |
This paper addresses the Boundary-caused Weights Confusion problem in semantic segmentation. The authors propose an E-CRF model by integrating CRF into CNN and pursuing gradient descent not only from scale but direction aspects. The method improves baseline for about 1% mIoU.
Strength:
+ This paper is well-written and ... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
This paper addresses the Boundary-caused Weights Confusion problem in semantic segmentation. The authors propose an E-CRF model by integrating CRF into CNN and pursuing gradient descent not only from scale but direction aspects. The method improves baseline for about 1% mIoU.
Strength:
+ This paper is well-writ... |
Authors study how the inclusion of nodes' coordinates into deep learning-based routing models affects the performance and robustness of these models. They evaluate different methods of embedding coordinates on synthetic traveling salesman problem task and FPD (Food Pickup and Delivery? Authors don't seem to spell out t... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
Authors study how the inclusion of nodes' coordinates into deep learning-based routing models affects the performance and robustness of these models. They evaluate different methods of embedding coordinates on synthetic traveling salesman problem task and FPD (Food Pickup and Delivery? Authors don't seem to spe... |
This paper proposes a new algorithm to learn policies for stochastic path problems that is called stochastic hierarchical abstraction-guided robot planner (SHARP). The algorithm consists of four parts: identify critical regions, synthesize option endpoints, generate a pseudo-reward function, and finally learn an option... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes a new algorithm to learn policies for stochastic path problems that is called stochastic hierarchical abstraction-guided robot planner (SHARP). The algorithm consists of four parts: identify critical regions, synthesize option endpoints, generate a pseudo-reward function, and finally learn a... |
This paper proposes a more time efficient active learning method based upon epistemic uncertainty by incorporating snapshot ensembles rather than using the traditional method of deep ensembles. The experimental results demonstrate comparable recognition performance with a significantly lower runtime.
The strength of th... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a more time efficient active learning method based upon epistemic uncertainty by incorporating snapshot ensembles rather than using the traditional method of deep ensembles. The experimental results demonstrate comparable recognition performance with a significantly lower runtime.
The streng... |
This paper proposed to solve the 3D semantic segmentation problem by distilling the knowledge embedded in the latent space of powerful 2D models. Unlike traditional knowledge distillation approaches, where student and teacher models take the same input，the inputs of the 2D teacher model in this paper are the panorama i... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposed to solve the 3D semantic segmentation problem by distilling the knowledge embedded in the latent space of powerful 2D models. Unlike traditional knowledge distillation approaches, where student and teacher models take the same input，the inputs of the 2D teacher model in this paper are the pa... |
The paper proposes an approach for learning policies that can enable better generalization across tasks. This is done using an ensemble of learners, each of which uses attention to focus on a different part of the feature space, and then using bandit-exploration to select a particular policy. Further, the authors inclu... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper proposes an approach for learning policies that can enable better generalization across tasks. This is done using an ensemble of learners, each of which uses attention to focus on a different part of the feature space, and then using bandit-exploration to select a particular policy. Further, the autho... |
The authors formulate the task of new drug recommendation as a few-shot learning task and propose a framework to solve it. They first utilize the drug ontology to learn the drug representation. Then they represent each patient as a set of phenotype vectors. Given a new drug, they make recommendations by measuring the s... | Recommendation: 3: reject, not good enough | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The authors formulate the task of new drug recommendation as a few-shot learning task and propose a framework to solve it. They first utilize the drug ontology to learn the drug representation. Then they represent each patient as a set of phenotype vectors. Given a new drug, they make recommendations by measuri... |
This paper proposes a discriminative method for pre-training Graph Neural Networks. The main idea is to simultaneously train a generator to recover identities of the masked edges, and train a discriminator to distinguish the generated edges from the original graph’s edges.
Strength
- The proposed method makes sense.
... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a discriminative method for pre-training Graph Neural Networks. The main idea is to simultaneously train a generator to recover identities of the masked edges, and train a discriminator to distinguish the generated edges from the original graph’s edges.
Strength
- The proposed method makes... |
The authors present a gradient-free framework for quantum deep learning using known techniques for the optimization, in particular Grover's algorithm. The techniques are not really new and the scheme is not at all near-term as is shown from the lack of experiments.
+ new framework for quantum neural networks that avoid... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
The authors present a gradient-free framework for quantum deep learning using known techniques for the optimization, in particular Grover's algorithm. The techniques are not really new and the scheme is not at all near-term as is shown from the lack of experiments.
+ new framework for quantum neural networks th... |
This paper looks at the problem of bias in automatic summarization algorithms. The authors design a method to quantitatively measure the extent to which an article's attributes, such as structure, paraphrasing, and article content influence _bias_ in generated summaries. They also study the causes of _bias_ by varying ... | Recommendation: 5: marginally below the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper looks at the problem of bias in automatic summarization algorithms. The authors design a method to quantitatively measure the extent to which an article's attributes, such as structure, paraphrasing, and article content influence _bias_ in generated summaries. They also study the causes of _bias_ by ... |
The paper looks at the task of online density estimation and classification. This is motivated by the problem of memory constraints, concept drifts, and temporal correlations. To this end, they propose Recurrent Real-valued Neural Autoregressive Density Estimator (RRNADE) which is an extension of the previous work RNAD... | Recommendation: 3: reject, not good enough | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The paper looks at the task of online density estimation and classification. This is motivated by the problem of memory constraints, concept drifts, and temporal correlations. To this end, they propose Recurrent Real-valued Neural Autoregressive Density Estimator (RRNADE) which is an extension of the previous w... |
This work proposes K-SAM as an alternative to SAM to improve algorithmic efficiency. The basic idea is to reduce the number of training samples used to approximate the stochastic gradients in both inner and outer optimization steps based on their loss values. This idea is originally proposed in Ordered-SGD, a biased st... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This work proposes K-SAM as an alternative to SAM to improve algorithmic efficiency. The basic idea is to reduce the number of training samples used to approximate the stochastic gradients in both inner and outer optimization steps based on their loss values. This idea is originally proposed in Ordered-SGD, a b... |
This work studies the generalization performance of momentum methods (GD+M, SGD+M) through the lens of implicit gradient regularization (Barrett & Dherin, 2020). The authors show that the effect of momentum can be understood as stronger implicit gradient regularization + stronger variance reduction (compared with GD/SG... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This work studies the generalization performance of momentum methods (GD+M, SGD+M) through the lens of implicit gradient regularization (Barrett & Dherin, 2020). The authors show that the effect of momentum can be understood as stronger implicit gradient regularization + stronger variance reduction (compared wi... |
The paper is about deep metric learning for fine-grained image retrieval. The task consists in learning a parameterized embedding function modelled as a neural network to project images into a deep space where images from the same categories are similar and images from different categories are not similar. The authors ... | Recommendation: 1: strong reject | Area: Deep Learning and representational learning | Review:
The paper is about deep metric learning for fine-grained image retrieval. The task consists in learning a parameterized embedding function modelled as a neural network to project images into a deep space where images from the same categories are similar and images from different categories are not similar. The ... |
This paper introduces a method called "FedorAS" which incorporates ideas from Neural Architecture Search into the domain of Federated Learning. In the scenario where the devices participating in a Federated Learning session has varying capabilities, and heterogeneous data distribution across devices, the goal of FedorA... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper introduces a method called "FedorAS" which incorporates ideas from Neural Architecture Search into the domain of Federated Learning. In the scenario where the devices participating in a Federated Learning session has varying capabilities, and heterogeneous data distribution across devices, the goal o... |
This paper studies transfer learning in the context of Reinforcement Learning. Using 3 diverse domains as a testbed (a toy platform game, some tasks from the DeepMind Control Suite, and some navigation tasks from habitat), it makes the following contributions:
- Contribution #1: Shows that when transferring representat... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper studies transfer learning in the context of Reinforcement Learning. Using 3 diverse domains as a testbed (a toy platform game, some tasks from the DeepMind Control Suite, and some navigation tasks from habitat), it makes the following contributions:
- Contribution #1: Shows that when transferring rep... |
This paper focuses on an emerging and interesting research topic, Physics-informed neural networks (PINNs) for PDE constrained optimization. Different from existing works that use the regularization-based paradigm which is hard to set a proper weights to balance the optimization targets and regularization terms, this p... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper focuses on an emerging and interesting research topic, Physics-informed neural networks (PINNs) for PDE constrained optimization. Different from existing works that use the regularization-based paradigm which is hard to set a proper weights to balance the optimization targets and regularization terms... |
This paper presents a novel method for photorealistic video style transfer that conducts color style transfer in videos without undesirable painterly spatial distortions and temporally inconsistent flickering artifacts. The proposed style removal-and-restoration framework, namely ColoristaNet, is capable of learning st... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper presents a novel method for photorealistic video style transfer that conducts color style transfer in videos without undesirable painterly spatial distortions and temporally inconsistent flickering artifacts. The proposed style removal-and-restoration framework, namely ColoristaNet, is capable of lea... |
This paper aims to theoretically investigate the optimization and generalization of deep PAC-Bayesian learning in the setting of wide neural networks. Technically, based on
the analysis ideas of the neural tangent kernel (NTK) for the deterministic models under the overparameterized settings, this paper characterizes... | Recommendation: 5: marginally below the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper aims to theoretically investigate the optimization and generalization of deep PAC-Bayesian learning in the setting of wide neural networks. Technically, based on
the analysis ideas of the neural tangent kernel (NTK) for the deterministic models under the overparameterized settings, this paper chara... |
This paper introduces a novel data augmentation model and technique for improving performance of few shot and low resource NLP tasks. Based on a seq-to-seq pre-trained language model (T5 large), this paper first present a multi-task pre-finetuning on 100+ NLP task using a new Knowledge Mixture Training (KoMT) framework... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper introduces a novel data augmentation model and technique for improving performance of few shot and low resource NLP tasks. Based on a seq-to-seq pre-trained language model (T5 large), this paper first present a multi-task pre-finetuning on 100+ NLP task using a new Knowledge Mixture Training (KoMT) f... |
This paper begins with an argument that the robustness of a model should also be conditioned on the training process. They then focus on the one specific aspect of the training process -- learning rate (LR) decays. They found that after LR decays, the confusion matrix becomes more symmetric. The author claimed that th... | 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 begins with an argument that the robustness of a model should also be conditioned on the training process. They then focus on the one specific aspect of the training process -- learning rate (LR) decays. They found that after LR decays, the confusion matrix becomes more symmetric. The author claimed... |
This paper argues that current Graph Neural Networks (GNNs) suffer from high time cost when trained on massive data, and class-imbalance issue. The authors then propose GraphDec to address the problems by pruning both GNN and the training data.
## Strength
(+) The authors provide plentiful empirical results to show the... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper argues that current Graph Neural Networks (GNNs) suffer from high time cost when trained on massive data, and class-imbalance issue. The authors then propose GraphDec to address the problems by pruning both GNN and the training data.
## Strength
(+) The authors provide plentiful empirical results to ... |
The authors analyze multiple deep learning prediction cases with continuous targets where discretizing the target into a classification problem results in better results compared to directly employing regression loss functions. The authors argue that the root cause of the problem is that the regression objective only o... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The authors analyze multiple deep learning prediction cases with continuous targets where discretizing the target into a classification problem results in better results compared to directly employing regression loss functions. The authors argue that the root cause of the problem is that the regression objectiv... |
This work considers the effect of large stepsize for GD in non-convex optimization. It shows theoretically that, under certain conditions, with a large stepsize GD can escape/avoid undesirable minimum, and in comparison, GD/SGD with small stepsize might converge into the undesirable minimum.
Then the paper turns to p... | Recommendation: 5: marginally below the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
This work considers the effect of large stepsize for GD in non-convex optimization. It shows theoretically that, under certain conditions, with a large stepsize GD can escape/avoid undesirable minimum, and in comparison, GD/SGD with small stepsize might converge into the undesirable minimum.
Then the paper tu... |
This paper explores efficient packing methods for training sequences of BERT, which avoids the padding tokens to speed up the training.
The authors introduce shortest-pack-first histogram-packing (SPFHP) and non-negative least-squares histogram-packing (NNLSHP) algorithms, which are shown to be straightforward to impl... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper explores efficient packing methods for training sequences of BERT, which avoids the padding tokens to speed up the training.
The authors introduce shortest-pack-first histogram-packing (SPFHP) and non-negative least-squares histogram-packing (NNLSHP) algorithms, which are shown to be straightforward... |
In this paper, the authors study the extensively-studied iterative hard thresholding (IHT) algorithm. They establish a new and critical gradient descent property of the hard thresholding (HT) operator, which is related to the distance between sparse points. In addition, they introduce the notion of HT-stable/unstable s... | Recommendation: 3: reject, not good enough | Area: Optimization (eg, convex and non-convex optimization) | Review:
In this paper, the authors study the extensively-studied iterative hard thresholding (IHT) algorithm. They establish a new and critical gradient descent property of the hard thresholding (HT) operator, which is related to the distance between sparse points. In addition, they introduce the notion of HT-stable/un... |
The work presents a deep GAN model to predict the 3d motion of humans in a social interaction. It introduces a dynamic intent model that models the interaction between the representation of the individual motions.
Introduction level discussion:
“stochastic forecasting of human trajectories in crowds …”
Advanced a lot... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The work presents a deep GAN model to predict the 3d motion of humans in a social interaction. It introduces a dynamic intent model that models the interaction between the representation of the individual motions.
Introduction level discussion:
“stochastic forecasting of human trajectories in crowds …”
Advanc... |
The paper studies adversarial linear mixture MDPs with unknown transition and bandit feedback. The authors propose an algorithm with an upper bound $\tilde{O}(dS^2 \sqrt{K} + \sqrt{HSAK})$, and prove an almost near-matching lower bound (in $d, K, A$). The authors highlight the use of a novel occupancy measure differenc... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper studies adversarial linear mixture MDPs with unknown transition and bandit feedback. The authors propose an algorithm with an upper bound $\tilde{O}(dS^2 \sqrt{K} + \sqrt{HSAK})$, and prove an almost near-matching lower bound (in $d, K, A$). The authors highlight the use of a novel occupancy measure d... |
The paper observes that the performance of a deep learning model might exhibit a high variance between different training instances and hence this might lead to a high variance in the performance gap across different groups. The paper proposes a solution based on the reordering of data, to lower the variance.
---------... | 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 observes that the performance of a deep learning model might exhibit a high variance between different training instances and hence this might lead to a high variance in the performance gap across different groups. The paper proposes a solution based on the reordering of data, to lower the variance.
-... |
In this work, the authors introduce the relational transformer, a modification of the conventional transformer architecture to work on graphs. This is not the first instance of transformer-inspired architectures which handle graph input, however a fundamental difference is that this method allows for full edge usage, n... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
In this work, the authors introduce the relational transformer, a modification of the conventional transformer architecture to work on graphs. This is not the first instance of transformer-inspired architectures which handle graph input, however a fundamental difference is that this method allows for full edge ... |
In this paper, the authors propose a new communication compression method called $\beta$-stochastic sign SGD, which is Byzantine-resilient and differentially private. The authors provide theoretical results about differential privacy and convergence. Besides, the proposed method is empirically compared with SignSGD and... | Recommendation: 3: reject, not good enough | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
In this paper, the authors propose a new communication compression method called $\beta$-stochastic sign SGD, which is Byzantine-resilient and differentially private. The authors provide theoretical results about differential privacy and convergence. Besides, the proposed method is empirically compared with Sig... |
The paper proposes a new approach to few-shot multimodal learning using a meta-learning approach that relies on generating visual prefixes that leverage prior training with other tasks. The visual prefix is produced by the proposed meta-mapper that meta-learns from previous tasks. The prefix approach enables seamless b... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper proposes a new approach to few-shot multimodal learning using a meta-learning approach that relies on generating visual prefixes that leverage prior training with other tasks. The visual prefix is produced by the proposed meta-mapper that meta-learns from previous tasks. The prefix approach enables se... |
This study built a parameterized solution space to help model driving behaviors for planning, which considering the effect of individual differences of driving demonstrations.
Strength: writing and novelty
Weakness: Reproducibility and application potential
1. How about planning efficiency? Planning time consumption... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This study built a parameterized solution space to help model driving behaviors for planning, which considering the effect of individual differences of driving demonstrations.
Strength: writing and novelty
Weakness: Reproducibility and application potential
1. How about planning efficiency? Planning time con... |
This paper proposes methods to distill unimodal teacher models of vision and language onto vision-language (VL) student models for utilizing the benefit of unimodal pre-trained models.
The authors present ideas to improve the performance: (1) VL student models are not changed, but additional learnable components are a... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper proposes methods to distill unimodal teacher models of vision and language onto vision-language (VL) student models for utilizing the benefit of unimodal pre-trained models.
The authors present ideas to improve the performance: (1) VL student models are not changed, but additional learnable componen... |
The paper proposes to investigate the computational efficiency of Active Learning (AL). The authors argue that typical AL methods are computationally intensive, as models are retrained from scratch at every round. They propose to reframe the AL setting as a Continual Learning one, which allows them to leverage the vast... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes to investigate the computational efficiency of Active Learning (AL). The authors argue that typical AL methods are computationally intensive, as models are retrained from scratch at every round. They propose to reframe the AL setting as a Continual Learning one, which allows them to leverage ... |
The submission "Hidden Poison: Machine unlearning enables camouflaged poisoning attacks " describes a novel variant of targeted data poisoning attacks. The submission introduces the notion of "camouflage" data points that neutralize the effects of poisoned data samples. This essentially incorporates an "off switch" int... | 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 submission "Hidden Poison: Machine unlearning enables camouflaged poisoning attacks " describes a novel variant of targeted data poisoning attacks. The submission introduces the notion of "camouflage" data points that neutralize the effects of poisoned data samples. This essentially incorporates an "off swi... |
The paper outline a method to use hyperbolic spaces in a graph neural network. The approach mainly relies on the computation of Laplacian Eigenfunctions: after embedding the graph nodes in some space, they are mapped via a kernel transformation to their hyperbolic features (i.e., Hyperbolic Laplacian eigenfunctions), a... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper outline a method to use hyperbolic spaces in a graph neural network. The approach mainly relies on the computation of Laplacian Eigenfunctions: after embedding the graph nodes in some space, they are mapped via a kernel transformation to their hyperbolic features (i.e., Hyperbolic Laplacian eigenfunct... |
This work introduces an algorithm for incorporating a learned sparse and low-rank penalty for image recovery tasks. A majorization bound of the penality functions (as quadratic functions) are derived for efficient optimization by Iteratively Reweighted Least
Squares (IRLS). Due to the recurrent nature of the algorithm,... | Recommendation: 8: accept, good paper | Area: Optimization (eg, convex and non-convex optimization) | Review:
This work introduces an algorithm for incorporating a learned sparse and low-rank penalty for image recovery tasks. A majorization bound of the penality functions (as quadratic functions) are derived for efficient optimization by Iteratively Reweighted Least
Squares (IRLS). Due to the recurrent nature of the al... |
This paper proposes a method for evaluating unsupervised representation learning in reinforcement learning. Using a linear probe on top of frozen, pretrained representations, the paper suggests learning to predict reward values from various states in downstream tasks. Additionally, the paper uses a linear probe to pr... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a method for evaluating unsupervised representation learning in reinforcement learning. Using a linear probe on top of frozen, pretrained representations, the paper suggests learning to predict reward values from various states in downstream tasks. Additionally, the paper uses a linear pro... |
This paper studies the dynamics of gradient descent with a large step size. In particular, the authors are interested in the study beyond the edge of stability, i.e., the step-size regime above the inverse of the Lipschitz constant.
In that regime, local convergence is not guaranteed anymore. However, they show that ... | Recommendation: 5: marginally below the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper studies the dynamics of gradient descent with a large step size. In particular, the authors are interested in the study beyond the edge of stability, i.e., the step-size regime above the inverse of the Lipschitz constant.
In that regime, local convergence is not guaranteed anymore. However, they sh... |
This paper proposes to analyze the theory of deep neural network from the perspective of non-parametric regression. The authors show that training an $L$ layer neural network with weight decay is basically the same as penalized regression. And in the case of functions in the Besov space and bounded variation space, neu... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes to analyze the theory of deep neural network from the perspective of non-parametric regression. The authors show that training an $L$ layer neural network with weight decay is basically the same as penalized regression. And in the case of functions in the Besov space and bounded variation sp... |
The paper deals with identifying a Riemannian metric for the cross-sectional samples of evolving probability measures on a common manifold. In particular, it makes use of optimal transport theory to compute the Riemannian metric (that minimizes the 1-Wasserstein distance). The usefulness of the learned metric is shown... | Recommendation: 8: accept, good paper | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The paper deals with identifying a Riemannian metric for the cross-sectional samples of evolving probability measures on a common manifold. In particular, it makes use of optimal transport theory to compute the Riemannian metric (that minimizes the 1-Wasserstein distance). The usefulness of the learned metric ... |
This is a theoretical work studying the limitations of piecewise linear functions in robustness certification. Main findings include: 1) Any piecewise linear certification is incomplete for piecewise linear networks; 2) For Lipschitz networks, when the learner can control the network, there exists some implementation s... | 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 theoretical work studying the limitations of piecewise linear functions in robustness certification. Main findings include: 1) Any piecewise linear certification is incomplete for piecewise linear networks; 2) For Lipschitz networks, when the learner can control the network, there exists some implemen... |
As a structured representation, the scene graph bridges computer vision and natural language processing, and the scene graph generation task is vital for the CV community. However, The annotation process of the scene graph is very complex including bounding boxes, labels, and relationships, so it is necessary to explor... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
As a structured representation, the scene graph bridges computer vision and natural language processing, and the scene graph generation task is vital for the CV community. However, The annotation process of the scene graph is very complex including bounding boxes, labels, and relationships, so it is necessary t... |
This paper proposed MERMADE, a deep RL approach to mechanism design that fuses model-based methods and gradient-based meta-learning methods to design a mechanism with fast adaptability. The authors analyze the one-shot adaptation performance of a meta-learned planner in a matrix game setting, under both perfect and noi... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposed MERMADE, a deep RL approach to mechanism design that fuses model-based methods and gradient-based meta-learning methods to design a mechanism with fast adaptability. The authors analyze the one-shot adaptation performance of a meta-learned planner in a matrix game setting, under both perfect... |
The authors propose 4D generative adversarial networks (GANs) to learn the unconditional generation of 3D-aware videos in neural implicit representations. The experimental results show that the proposed method learns a rich embedding of decomposable 3D structures and motions whose quality is comparable to that of exist... | Recommendation: 5: marginally below the acceptance threshold | Area: Generative models | Review:
The authors propose 4D generative adversarial networks (GANs) to learn the unconditional generation of 3D-aware videos in neural implicit representations. The experimental results show that the proposed method learns a rich embedding of decomposable 3D structures and motions whose quality is comparable to that ... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.