review stringlengths 5 16.6k | score stringclasses 5
values | area stringclasses 12
values | text stringlengths 31 5.65k |
|---|---|---|---|
In this work, motivated by the limitation that there are no fast or scalable solvers for vector quantile regression (VQR), the authors provide a highly-scalable solver for VQR that relies on solving its relaxed dual formulation. In addition, the authors propose vector monotone rearrangement (VMR), which serves as a ref... | Recommendation: 8: accept, good paper | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
In this work, motivated by the limitation that there are no fast or scalable solvers for vector quantile regression (VQR), the authors provide a highly-scalable solver for VQR that relies on solving its relaxed dual formulation. In addition, the authors propose vector monotone rearrangement (VMR), which serves ... |
This paper focuses on the interpretability of deep learning in sequential scenarios. To this end, the authors propose to explain sequential predictions at the subsequence-level, and a distribution-based segmentation method.
Strong points:
1. The motivation of subsequence-level explanation is clear.
Weak points:
1. The... | 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 focuses on the interpretability of deep learning in sequential scenarios. To this end, the authors propose to explain sequential predictions at the subsequence-level, and a distribution-based segmentation method.
Strong points:
1. The motivation of subsequence-level explanation is clear.
Weak points... |
This paper proposes a graph variational Bayesian causal inference framework to predict a cell's gene expression under counterfactual perturbations, gene regulatory networks were used to aid the individualized cellular response predictions. A robust estimator for the asymptotically efficient estimation of the marginal p... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper proposes a graph variational Bayesian causal inference framework to predict a cell's gene expression under counterfactual perturbations, gene regulatory networks were used to aid the individualized cellular response predictions. A robust estimator for the asymptotically efficient estimation of the ma... |
This paper tries to learn domain-general representation by achieving the target conditioned representation independence (TCRI). Specifically, the proposed TCRI method not only addresses the domain invariant property but also assumes that the domain-general feature and domain-specific feature are conditionally independe... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper tries to learn domain-general representation by achieving the target conditioned representation independence (TCRI). Specifically, the proposed TCRI method not only addresses the domain invariant property but also assumes that the domain-general feature and domain-specific feature are conditionally i... |
This paper proposes to enhance the DeBERTa model with the more sample-efficient replaced token detection (RTD) pre-training task.
To further improve the training efficiency and performance of the vanilla embedding-sharing strategy (ELECTRA) for the DeBERTa + RTD setting, a new gradient-entangled embedding-sharing (GDE... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes to enhance the DeBERTa model with the more sample-efficient replaced token detection (RTD) pre-training task.
To further improve the training efficiency and performance of the vanilla embedding-sharing strategy (ELECTRA) for the DeBERTa + RTD setting, a new gradient-entangled embedding-shar... |
The paper introduces a methodology to train concept bottleneck models, which automates the process of concept definition and it is empirically shown to outperform existing baselines making use of user-defined concepts. The methodology consists of four stages:
In stage 1, concepts are defined by generating lists of wor... | 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 introduces a methodology to train concept bottleneck models, which automates the process of concept definition and it is empirically shown to outperform existing baselines making use of user-defined concepts. The methodology consists of four stages:
In stage 1, concepts are defined by generating list... |
The authors propose to conduct psychophysical experiments using a combination of in-lab and
crowdsourced subjects to derive an 'objective' (that is, one decoupled from any particular model)
measure of the distribution of sample-difficulty in two widely-used object recognition datasets,
ImageNet and ObjectNet. This diff... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The authors propose to conduct psychophysical experiments using a combination of in-lab and
crowdsourced subjects to derive an 'objective' (that is, one decoupled from any particular model)
measure of the distribution of sample-difficulty in two widely-used object recognition datasets,
ImageNet and ObjectNet. T... |
This is a very cool paper, considering the strategic incentives of multiple content creators when producing content for a YouTube-like recommender system. The article addresses questions of whether the setup of the recommending algorithm has an effect on the type of aggregate content-producer behaviour we would expect ... | Recommendation: 8: accept, good paper | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This is a very cool paper, considering the strategic incentives of multiple content creators when producing content for a YouTube-like recommender system. The article addresses questions of whether the setup of the recommending algorithm has an effect on the type of aggregate content-producer behaviour we would... |
In this paper, the authors propose to directly solve the Monge's optimal transport problem under $L^2$ cost function, where the existence and uniqueness of the solution is guaranteed. Specifically, they use a neural network to parameterize the OT map, and try to solve the KL-divergence regularized OT problem proposed i... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
In this paper, the authors propose to directly solve the Monge's optimal transport problem under $L^2$ cost function, where the existence and uniqueness of the solution is guaranteed. Specifically, they use a neural network to parameterize the OT map, and try to solve the KL-divergence regularized OT problem pr... |
The paper studies the value function back-up estimation problem in data efficient reinforcement learning. The key idea is to the Graph Backup, with uses creates a tree-structured update of the value function based on multiple trajectories which have overlapping states. Because of the overlapping trajectories, it has ad... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper studies the value function back-up estimation problem in data efficient reinforcement learning. The key idea is to the Graph Backup, with uses creates a tree-structured update of the value function based on multiple trajectories which have overlapping states. Because of the overlapping trajectories, i... |
The paper provides empirical evidences of generalizability of modern neural architectures over synthetic tasks following the Chomsky hierarchy. The results generally agree with theoretical results, highlighting (again) the need for external memory.
Strength
======
- A nice empirical study with clear designs and message... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper provides empirical evidences of generalizability of modern neural architectures over synthetic tasks following the Chomsky hierarchy. The results generally agree with theoretical results, highlighting (again) the need for external memory.
Strength
======
- A nice empirical study with clear designs and... |
This paper proposes a hierarchical FL framework with a new label-driven distillation method to handle non-iid FL scenarios
S1. The proposal of a hierarchical structure for FL is reasonable.
W1. The paper is very hard to follow, as a lot of design considerations are proposed, but which parts are the most novel ones are... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper proposes a hierarchical FL framework with a new label-driven distillation method to handle non-iid FL scenarios
S1. The proposal of a hierarchical structure for FL is reasonable.
W1. The paper is very hard to follow, as a lot of design considerations are proposed, but which parts are the most novel ... |
This paper discusses a method for detecting the evaluation Parkinson’s disease and the patients’ use of medication by monitoring the motion of patients at home. This is done by indoor localization at room level fusing Bluetooth RSSI measurements and accelerometer (or actually whole IMU?) measurements. The paper builds ... | Recommendation: 3: reject, not good enough | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper discusses a method for detecting the evaluation Parkinson’s disease and the patients’ use of medication by monitoring the motion of patients at home. This is done by indoor localization at room level fusing Bluetooth RSSI measurements and accelerometer (or actually whole IMU?) measurements. The paper... |
This paper presents the first NeRF reconstruction method, Generalizable NeRF Transformer (GNT), based on transformer. It introduces the view transformer which predicts coordinate-aligned features by aggregating information from epipolar lines on the neighboring views with attention and the ray transformer which renders... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper presents the first NeRF reconstruction method, Generalizable NeRF Transformer (GNT), based on transformer. It introduces the view transformer which predicts coordinate-aligned features by aggregating information from epipolar lines on the neighboring views with attention and the ray transformer which... |
The manuscript is describing how to estimate the flow of air around an airfoil only from the pressure and velocity distribution on the airfoil for high Reynolds number flows. Graph based structure is used to define a "mesh" has some intelligence, i.e. features at each vertex. The paper combines far-field information wi... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The manuscript is describing how to estimate the flow of air around an airfoil only from the pressure and velocity distribution on the airfoil for high Reynolds number flows. Graph based structure is used to define a "mesh" has some intelligence, i.e. features at each vertex. The paper combines far-field inform... |
In this work, the authors investigate the the ability of agents to engage in turn taking through a simple language based game. In this game, agents can emit information through a shared channel and using the information the agents needs to solve the task. The authors find that agents that develop turn taking achieve hi... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
In this work, the authors investigate the the ability of agents to engage in turn taking through a simple language based game. In this game, agents can emit information through a shared channel and using the information the agents needs to solve the task. The authors find that agents that develop turn taking ac... |
The reviewed work proposes "identical initialization" as a novel initialization scheme for neural networks. The guiding principle of this initialization scheme is to ensure that each layer, at initialization, propagates activations identically.
The motivation behind this choice is that the resulting isommetry property... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The reviewed work proposes "identical initialization" as a novel initialization scheme for neural networks. The guiding principle of this initialization scheme is to ensure that each layer, at initialization, propagates activations identically.
The motivation behind this choice is that the resulting isommetry ... |
[REVISED]
Task: The paper considers the relatively unexplored task of making multi-label text classification interpretable by learning a model that induces implicit labeled spans given only sentence-level labels as supervision.
Model: The paper assumes a "backbone" model for unsupervised parsing that gives an embeddi... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
[REVISED]
Task: The paper considers the relatively unexplored task of making multi-label text classification interpretable by learning a model that induces implicit labeled spans given only sentence-level labels as supervision.
Model: The paper assumes a "backbone" model for unsupervised parsing that gives an... |
The paper introduces a method that learns a loss landscape
where gradient descent is efficient, bringing massive improvement and acceleration
to the inversion process of learning a CNN. Advantages are demonstrated using a number of methods
for both generative and discriminative tasks.
The method helps to accelerate op... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper introduces a method that learns a loss landscape
where gradient descent is efficient, bringing massive improvement and acceleration
to the inversion process of learning a CNN. Advantages are demonstrated using a number of methods
for both generative and discriminative tasks.
The method helps to accel... |
This paper presents a review-based explainable recommedation model that predicts user preference and gives explanations based on user/item clusters. The core idea is to learn two sets of clusters, one based on text reviews and the other based on user/item interaction information. This explicitly ensures that informatio... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper presents a review-based explainable recommedation model that predicts user preference and gives explanations based on user/item clusters. The core idea is to learn two sets of clusters, one based on text reviews and the other based on user/item interaction information. This explicitly ensures that in... |
The authors present MISA, a framework for offline RL based on mutual information based regularization over states and actions. The authors motivate their approach step by step and demonstrate how TD3+Behavior cloning and CQL are related to their proposed approach. Finally, the authors provide empirical results relative... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The authors present MISA, a framework for offline RL based on mutual information based regularization over states and actions. The authors motivate their approach step by step and demonstrate how TD3+Behavior cloning and CQL are related to their proposed approach. Finally, the authors provide empirical results ... |
This paper extends the concept of curriculum learning and investigates the question of teaching an existing human-like model to be stronger using a data-efficient curriculum, while maintaining the model’s human similarity. The authors then applied the method to chess and experiments show that the choice of teacher has ... | 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 extends the concept of curriculum learning and investigates the question of teaching an existing human-like model to be stronger using a data-efficient curriculum, while maintaining the model’s human similarity. The authors then applied the method to chess and experiments show that the choice of teac... |
CAN is a method of self-supervised learning for visual representations. The method builds on three families of self-supervised learning methods: Contrastive learning, Masked Autoencoding, and Noise prediction. The paper claims that the learning mechanisms of each family are complementary to one another, and includes so... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
CAN is a method of self-supervised learning for visual representations. The method builds on three families of self-supervised learning methods: Contrastive learning, Masked Autoencoding, and Noise prediction. The paper claims that the learning mechanisms of each family are complementary to one another, and inc... |
This manuscript considers the task of vision-language manipulation; the proposed approach leverages Combinatory Categorical Grammar (CCG) to parse natural language instructions into a manipulation program (task plan from a domain-specific language), consisting of functional modules.
(Strengths)
Provides a framework fo... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This manuscript considers the task of vision-language manipulation; the proposed approach leverages Combinatory Categorical Grammar (CCG) to parse natural language instructions into a manipulation program (task plan from a domain-specific language), consisting of functional modules.
(Strengths)
Provides a fram... |
The paper proposes a new way of representing a text as a single fixed-size vector. While most text embedding methods use model activations to represent the input text, the proposed method computes the difference in model parameters that is induced by finetuning on a single text input towards. The finetuning task is a c... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper proposes a new way of representing a text as a single fixed-size vector. While most text embedding methods use model activations to represent the input text, the proposed method computes the difference in model parameters that is induced by finetuning on a single text input towards. The finetuning tas... |
The authors propose a method for recalibrating a conformal predictor on a shifted data distribution with only unlabeled data. To this extent, they intend to estimate a confidence level beta that, if used for calibration on the original data distribution, would give empirical coverage 1 – alpha (the target coverage) on ... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
The authors propose a method for recalibrating a conformal predictor on a shifted data distribution with only unlabeled data. To this extent, they intend to estimate a confidence level beta that, if used for calibration on the original data distribution, would give empirical coverage 1 – alpha (the target cover... |
This paper considers the problem of using population-based training to learn more "natural" languages for communicating between agents. When two agents (a "sender" and a "receiver") are trained to solve a communication task together, the communication protocol, or language, that they learn often lacks structure, and s... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper considers the problem of using population-based training to learn more "natural" languages for communicating between agents. When two agents (a "sender" and a "receiver") are trained to solve a communication task together, the communication protocol, or language, that they learn often lacks structur... |
This work proposes an instance-specific variational prompt-tuning framework for image-language downstream tasks. Here, the experiments are performed with a frozen CLIP model while fine-tuning the parameters used for prompt creation. It combines two ideas from existing publications on prompt-tuning with learnable prompt... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This work proposes an instance-specific variational prompt-tuning framework for image-language downstream tasks. Here, the experiments are performed with a frozen CLIP model while fine-tuning the parameters used for prompt creation. It combines two ideas from existing publications on prompt-tuning with learnabl... |
This paper presents random-LTD (Layer Token Dropping), a method that randomly skips the computation of certain tokens during pretraining, and evaluates it on BERT and GPT pretraining, and ViT fine-tuning. The paper includes several ablations to show that the design decisions behind its method are necessary.
Strengths
... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper presents random-LTD (Layer Token Dropping), a method that randomly skips the computation of certain tokens during pretraining, and evaluates it on BERT and GPT pretraining, and ViT fine-tuning. The paper includes several ablations to show that the design decisions behind its method are necessary.
St... |
This paper considers the problem of growing neural networks during training. The authors propose a parameterization and optimization scheme that pays close attention to weight and gradient scaling while reacting to training dynamics. They further refine their original proposal to mitigate the problem of newly grafted... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper considers the problem of growing neural networks during training. The authors propose a parameterization and optimization scheme that pays close attention to weight and gradient scaling while reacting to training dynamics. They further refine their original proposal to mitigate the problem of newly... |
Gradient Coding is a general approach distributed computation (of gradients) in the presence of straggling nodes. The paper introduces two new schemes for GC, where coding is not only done across compute nodes, but also across the time dimension, measured by compute rounds. The first scheme uses the original GC but wit... | Recommendation: 8: accept, good paper | Area: General Machine Learning | Review:
Gradient Coding is a general approach distributed computation (of gradients) in the presence of straggling nodes. The paper introduces two new schemes for GC, where coding is not only done across compute nodes, but also across the time dimension, measured by compute rounds. The first scheme uses the original GC... |
This paper tackles the problem of learning with noisy-labels. By combining different existing approaches, such as noise transition matrix estimation and sample reweighing, with a new set-level self-supervised learning pipeline, the proposed approach achieve strong performance on NNL (noisy label learning) vision benchm... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper tackles the problem of learning with noisy-labels. By combining different existing approaches, such as noise transition matrix estimation and sample reweighing, with a new set-level self-supervised learning pipeline, the proposed approach achieve strong performance on NNL (noisy label learning) visio... |
The paper considers the problem of general functions from a general function class under a bandit feedback model. Specifically, this work considers the setup of heteroscedastic noise where in the noise varies with different actions, possibly depending on the action taken by the learner. Under this scenario, the authors... | Recommendation: 5: marginally below the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The paper considers the problem of general functions from a general function class under a bandit feedback model. Specifically, this work considers the setup of heteroscedastic noise where in the noise varies with different actions, possibly depending on the action taken by the learner. Under this scenario, the... |
This paper tackles with a problem that when distilling a small student model using a large pre-trained transformer, the “capacity gap” between the student and the teacher often hinders the teacher from transferring good performance to the student. The HomoDistill method proposed by this paper investigates to progressiv... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper tackles with a problem that when distilling a small student model using a large pre-trained transformer, the “capacity gap” between the student and the teacher often hinders the teacher from transferring good performance to the student. The HomoDistill method proposed by this paper investigates to pr... |
In this paper, the authors revisited the topic: whether the modality based recommender models (MoRec) can exceed or be on par with the ID-only based models (IDRec) when item modality features are available? They examined this question from several perspective, and performed multiple experiments. Their results show that... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
In this paper, the authors revisited the topic: whether the modality based recommender models (MoRec) can exceed or be on par with the ID-only based models (IDRec) when item modality features are available? They examined this question from several perspective, and performed multiple experiments. Their results s... |
This paper is very interesting. It studies the visualness of text based on a simple example. Some words can incur our imagination for the corresponding image, while the others might not. With this motivation, the author proposed a new dataset called TimeD with both manually labeled data and the automatically model labe... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper is very interesting. It studies the visualness of text based on a simple example. Some words can incur our imagination for the corresponding image, while the others might not. With this motivation, the author proposed a new dataset called TimeD with both manually labeled data and the automatically mo... |
This paper introduces a type of neural network architecture that augments individual units in the network with an additional state variable and reparameterizes them using a 3x3 weight matrix. Each individual unit receives input that is the output of a linear transformation (as in standard networks), but instead of pass... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper introduces a type of neural network architecture that augments individual units in the network with an additional state variable and reparameterizes them using a 3x3 weight matrix. Each individual unit receives input that is the output of a linear transformation (as in standard networks), but instead... |
This article presents a novel system that brings differentiable physics and neural radiance fields for dynamic scenes by augmenting a NeRF with a differentiable continuum dynamics model. A hybrid Eulerian-Lagrangian formulation is defined in this system to permit advecting geometry and appearance. This framework outper... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This article presents a novel system that brings differentiable physics and neural radiance fields for dynamic scenes by augmenting a NeRF with a differentiable continuum dynamics model. A hybrid Eulerian-Lagrangian formulation is defined in this system to permit advecting geometry and appearance. This framewor... |
This paper uses a diffusion Transformer to map a pair (initial parameter, target loss) to a distribution over parameters that achieve the specified performance. The model is trained with a pre-training dataset of neural network checkpoints.
Pros
- The overall idea is interesting: posing performance-conditional paramete... | Recommendation: 5: marginally below the acceptance threshold | Area: Generative models | Review:
This paper uses a diffusion Transformer to map a pair (initial parameter, target loss) to a distribution over parameters that achieve the specified performance. The model is trained with a pre-training dataset of neural network checkpoints.
Pros
- The overall idea is interesting: posing performance-conditional ... |
In this paper the authors tackle the problem of case-based decision support. In the problem setting, a ML algorithm assists humans in making decisions on classification tasks by identifying an example from the training set that is similar to the unseen example. In the general case, the similarity between examples is de... | Recommendation: 6: marginally above the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
In this paper the authors tackle the problem of case-based decision support. In the problem setting, a ML algorithm assists humans in making decisions on classification tasks by identifying an example from the training set that is similar to the unseen example. In the general case, the similarity between exampl... |
The paper considers approximating solutions of constrained optimization problems using neural networks. A neural network is trained to approximate a parametric family of such problems, mapping the parameters to the optimal solution (i.e. arg-max). The objective for training the neural net is a combination of the obje... | Recommendation: 3: reject, not good enough | Area: Optimization (eg, convex and non-convex optimization) | Review:
The paper considers approximating solutions of constrained optimization problems using neural networks. A neural network is trained to approximate a parametric family of such problems, mapping the parameters to the optimal solution (i.e. arg-max). The objective for training the neural net is a combination of ... |
This paper studies the two-stage transfer learning approach of head tuning first (a generalization of linear probing where the head can be nonlinear), followed by finetuning both the backbone and continuing to finetune the head. Their work builds upon the setup of Kumar et al, but they claim to relax the assumption mad... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper studies the two-stage transfer learning approach of head tuning first (a generalization of linear probing where the head can be nonlinear), followed by finetuning both the backbone and continuing to finetune the head. Their work builds upon the setup of Kumar et al, but they claim to relax the assump... |
The paper set out to investigate the learning dynamics of contrastive learning (e.g.CLIP) in, investigate how contrastive learning learn to align the representations from different views efficiently.
Strengths:
(a) The paper provided a fresh view that the alignment and balance of representation play a important role in... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
The paper set out to investigate the learning dynamics of contrastive learning (e.g.CLIP) in, investigate how contrastive learning learn to align the representations from different views efficiently.
Strengths:
(a) The paper provided a fresh view that the alignment and balance of representation play a important... |
The authors consider the standard error-in-variables model: Y=X’b+q and tilde{X}=X+U. The authors impose that X, U, and q are homoscedastic Gaussians with fixed dimensions. Then the standard attenuation bias formula can be inverted to correct for the attenuation bias, if the variance of q is known. The authors quote es... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
The authors consider the standard error-in-variables model: Y=X’b+q and tilde{X}=X+U. The authors impose that X, U, and q are homoscedastic Gaussians with fixed dimensions. Then the standard attenuation bias formula can be inverted to correct for the attenuation bias, if the variance of q is known. The authors ... |
The authors proposed an attention retractable Transformer (ART) for accurate image restoration, which consists of dense and sparse attention modules. The experiments are extensive and include several general image restoration tasks, where the proposed method achieves SOTA performance.
Strength:
1. The paper is well-wr... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The authors proposed an attention retractable Transformer (ART) for accurate image restoration, which consists of dense and sparse attention modules. The experiments are extensive and include several general image restoration tasks, where the proposed method achieves SOTA performance.
Strength:
1. The paper is... |
The paper tackles the task of open-world instance segmentation in images (no predefined set of classes). The proposed method extends on an existing state-of-the-art single-stage (without employing a dedicated proposals procedure) approach, such as Mask2former, through the addition of a foreground segmentation branch. T... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper tackles the task of open-world instance segmentation in images (no predefined set of classes). The proposed method extends on an existing state-of-the-art single-stage (without employing a dedicated proposals procedure) approach, such as Mask2former, through the addition of a foreground segmentation b... |
This paper considers a problem setup in which decision makers launch experiments (say over target population of agents) and have to wait a period of time before receiving the outcomes. These outcomes may or may not be conclusive, and accordingly the experiment launched may or may not be useful. If found inconclusive, i... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
This paper considers a problem setup in which decision makers launch experiments (say over target population of agents) and have to wait a period of time before receiving the outcomes. These outcomes may or may not be conclusive, and accordingly the experiment launched may or may not be useful. If found inconcl... |
This paper investigates pruning at initialization through the lens of 'effective nodes' and 'effective paths'. Effective paths are defined as paths that exist from input node to output node and effective nodes as nodes that participate in at least one of these paths. The paper argues that these 2 objectives must be sim... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper investigates pruning at initialization through the lens of 'effective nodes' and 'effective paths'. Effective paths are defined as paths that exist from input node to output node and effective nodes as nodes that participate in at least one of these paths. The paper argues that these 2 objectives mus... |
The paper studied protein generation problem, and specifically, motif-scaffolding conditional protein design problem. This paper first proposed a general protein 3D structure generation framework, which can generate high-quality protein backbones in 3D. Then the author proposes a new sampling algorithm for inpainting p... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper studied protein generation problem, and specifically, motif-scaffolding conditional protein design problem. This paper first proposed a general protein 3D structure generation framework, which can generate high-quality protein backbones in 3D. Then the author proposes a new sampling algorithm for inpa... |
This paper proposes an approach to pretrain a network to learn features for RL from videos of the task without actions. The network uses a reconstruction loss and a cycle-consistency loss to learn this. The approach is evaluated on several atari game tasks.
Efficient RL learning is an important task, and this approach ... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes an approach to pretrain a network to learn features for RL from videos of the task without actions. The network uses a reconstruction loss and a cycle-consistency loss to learn this. The approach is evaluated on several atari game tasks.
Efficient RL learning is an important task, and this a... |
This paper analyzes the gradient mismatching problem in direct training neural networks and proposes a novel method to solve this problem by fusing the learnable relaxation degree into the network with random spike noise and get good results on both static and events datasets.
Strengths:
1. It is a novel and simple app... | Recommendation: 6: marginally above the acceptance threshold | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
This paper analyzes the gradient mismatching problem in direct training neural networks and proposes a novel method to solve this problem by fusing the learnable relaxation degree into the network with random spike noise and get good results on both static and events datasets.
Strengths:
1. It is a novel and si... |
This paper improves SymDL by searching the Pareto-optimal message-passing flows to learn an additional formula skeleton. Such improvement empowers the method with more generality (requiring no prior knowledge), compactness, while maintaining correctness. Additionally, this paper also extends the applicability to Electr... | Recommendation: 8: accept, good paper | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper improves SymDL by searching the Pareto-optimal message-passing flows to learn an additional formula skeleton. Such improvement empowers the method with more generality (requiring no prior knowledge), compactness, while maintaining correctness. Additionally, this paper also extends the applicability t... |
The paper is part of the recent stream of papers which use large language models to generate code to solve coding contest type problems. Two popular ways of measuring the performance of such models is to (1) evaluate $pass@k$ -- $k$ solutions are generated for a problem and evaluated against the unit tests for that pro... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper is part of the recent stream of papers which use large language models to generate code to solve coding contest type problems. Two popular ways of measuring the performance of such models is to (1) evaluate $pass@k$ -- $k$ solutions are generated for a problem and evaluated against the unit tests for ... |
In this paper, they focus on the template inversion attack against face recognition systems. Within a generative adversarial network (GAN)-based framework, learn a mapping from facial templates to the intermediate latent space of a pre-trained face generation network. their proposed method achieves high success attack... | Recommendation: 6: marginally above the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
In this paper, they focus on the template inversion attack against face recognition systems. Within a generative adversarial network (GAN)-based framework, learn a mapping from facial templates to the intermediate latent space of a pre-trained face generation network. their proposed method achieves high succes... |
This paper presents an approach to incorporate lattice priors into attention mask for deep learning. The theoretical results show this prior can be formulated as convolutions on an identity matrix. The proposed approach improves the sample efficiency of the model to solve the ARC tasks (few-shot geometric transformatio... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper presents an approach to incorporate lattice priors into attention mask for deep learning. The theoretical results show this prior can be formulated as convolutions on an identity matrix. The proposed approach improves the sample efficiency of the model to solve the ARC tasks (few-shot geometric trans... |
This paper presents present MUG, a novel interactive task for multimodal grounding where a user and an agent work collaboratively on an interface screen, essentially bridging multimodal applications with user interfaces in HCI research. Their main novelty is an interactive task as opposed to existing static interaction... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper presents present MUG, a novel interactive task for multimodal grounding where a user and an agent work collaboratively on an interface screen, essentially bridging multimodal applications with user interfaces in HCI research. Their main novelty is an interactive task as opposed to existing static int... |
The paper studies the underlying reason behind why fine tuning pre-trained language models works well. The paper shows that in some cases, those fine tuned models can be described using neural tangent kernel. In those cases, the kernel view provide explanation of the fine-tuning success.
Strength:
The paper tackle on... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The paper studies the underlying reason behind why fine tuning pre-trained language models works well. The paper shows that in some cases, those fine tuned models can be described using neural tangent kernel. In those cases, the kernel view provide explanation of the fine-tuning success.
Strength:
The paper t... |
The paper provides a general derivation of Evidence Lower Bounds in Multivariate Diffusion Models and some parametrization schemes. The paper also introduces MALDA diffusion model under its framework and evaluates it on MNIST and CIFAR10.
## Strength:
The derivation seems to be unfied and generic. The experimental res... | Recommendation: 3: reject, not good enough | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The paper provides a general derivation of Evidence Lower Bounds in Multivariate Diffusion Models and some parametrization schemes. The paper also introduces MALDA diffusion model under its framework and evaluates it on MNIST and CIFAR10.
## Strength:
The derivation seems to be unfied and generic. The experime... |
This paper addresses the problem of quality diversity optimization. In short, this problem is to find a set of which maximize some function f(x) where each x in this set is the optimal value in some region of the space of possible x's. The space where these regions are define is typically lower dimensional than the spa... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
This paper addresses the problem of quality diversity optimization. In short, this problem is to find a set of which maximize some function f(x) where each x in this set is the optimal value in some region of the space of possible x's. The space where these regions are define is typically lower dimensional than... |
The paper studies model-based reinforcement learning in the context of changing environments. More specifically, the algorithms are evaluated in the Local Change Adaptation (LoCA) setting. Well known methods like Dreamer struggle in this setting. The paper proposes a simple method that modifies the first-in-first-out r... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper studies model-based reinforcement learning in the context of changing environments. More specifically, the algorithms are evaluated in the Local Change Adaptation (LoCA) setting. Well known methods like Dreamer struggle in this setting. The paper proposes a simple method that modifies the first-in-fir... |
The paper proposes a Regularized Lottery Ticket Hypothesis inspired network to deal with the well-known catastrophic forgetting problem. By only updating partial weights in the network, the proposed method claims to have a good performance on both the classes in previous and current sessions.
Strengths:
1). The paper... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes a Regularized Lottery Ticket Hypothesis inspired network to deal with the well-known catastrophic forgetting problem. By only updating partial weights in the network, the proposed method claims to have a good performance on both the classes in previous and current sessions.
Strengths:
1). T... |
The paper reports the effects of a hidden prior (i.e., the uniform feature distribution assumption) used in the existing self-supervised pretraining methods for class-balanced datasets and real world class-imbalanced datasets. Moreover, a power-law distribution assumption to perfer long-tail prior is formulated for se... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
The paper reports the effects of a hidden prior (i.e., the uniform feature distribution assumption) used in the existing self-supervised pretraining methods for class-balanced datasets and real world class-imbalanced datasets. Moreover, a power-law distribution assumption to perfer long-tail prior is formulate... |
This paper tackles the setting of unsupervised domain adaptation when the shift between the source and target is large and thus gradual domain adaptation is needed. They propose a gradual domain adaptation algorithm that creates intermediate domains for the gradual adaptation and then applies (standard) gradual self tr... | Recommendation: 8: accept, good paper | Area: General Machine Learning | Review:
This paper tackles the setting of unsupervised domain adaptation when the shift between the source and target is large and thus gradual domain adaptation is needed. They propose a gradual domain adaptation algorithm that creates intermediate domains for the gradual adaptation and then applies (standard) gradual... |
This paper proposes a new unified mathematical framework for neural network (NN) model fusion, which is based on the Wasserstein/Gromov-Wasserstein barycenters (WB/GWB), i.e., formulating as a series of optimal transport (OT) problems. The proposed mathematical framework is universal and can be applied to a broad class... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper proposes a new unified mathematical framework for neural network (NN) model fusion, which is based on the Wasserstein/Gromov-Wasserstein barycenters (WB/GWB), i.e., formulating as a series of optimal transport (OT) problems. The proposed mathematical framework is universal and can be applied to a bro... |
This paper proposes the generate-then-read (GENREAD) approach to solve knowledge-intensive tasks. Compared to the previous retrieve-then-read approach, the proposed method first generates the contextual documents by prompting a large language model. The authors provide two variants of reader, zero-shot and supervised s... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes the generate-then-read (GENREAD) approach to solve knowledge-intensive tasks. Compared to the previous retrieve-then-read approach, the proposed method first generates the contextual documents by prompting a large language model. The authors provide two variants of reader, zero-shot and supe... |
In this work, the authors repurpose the well-known knowledge distillation paradigm, which is typically used for improving model accuracy, for the task of transferring the capability of reducing negative flip rate (NFR) from an ensemble to a single model. To this end, the authors devise a method called Ensemble Logit Di... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
In this work, the authors repurpose the well-known knowledge distillation paradigm, which is typically used for improving model accuracy, for the task of transferring the capability of reducing negative flip rate (NFR) from an ensemble to a single model. To this end, the authors devise a method called Ensemble ... |
The paper adds a patching technique to the DCT-Mask model and adopts a refinement technique for each patch, so that high-resolution masks can be achieved. The patching technique can produce better boundaries compared to the DCT mask model itself as element changes for DCT vectors can be limited to the patch level rathe... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper adds a patching technique to the DCT-Mask model and adopts a refinement technique for each patch, so that high-resolution masks can be achieved. The patching technique can produce better boundaries compared to the DCT mask model itself as element changes for DCT vectors can be limited to the patch lev... |
-This paper considers a general label noise problem (which does not assume any noise models). It proposes a fully heuristic approach: they firstly pre-train the classification model with some clean data, and then update the classification model by alternating between minimizing a loss on learning label confidence score... | Recommendation: 3: reject, not good enough | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
-This paper considers a general label noise problem (which does not assume any noise models). It proposes a fully heuristic approach: they firstly pre-train the classification model with some clean data, and then update the classification model by alternating between minimizing a loss on learning label confiden... |
The paper studies the important problem of learning text-to-speech synthesis from noisy unpaired data. To order to achieve this, the authors propose several practical tricks, including normalizing variable and noisy information in speech data, curriculum learning,
length augmentation and auxiliary supervised learning. ... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper studies the important problem of learning text-to-speech synthesis from noisy unpaired data. To order to achieve this, the authors propose several practical tricks, including normalizing variable and noisy information in speech data, curriculum learning,
length augmentation and auxiliary supervised le... |
This paper studies the reason multi-task prompted fine-tuning promotes zero-shot task generalization (ZSTG) in the context of T0. In particular, the authors hypothesize that only a few tasks are crucial for ZSTG. In order to validate this hypothesis pairwise train/test performance between tasks is used to identify the ... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper studies the reason multi-task prompted fine-tuning promotes zero-shot task generalization (ZSTG) in the context of T0. In particular, the authors hypothesize that only a few tasks are crucial for ZSTG. In order to validate this hypothesis pairwise train/test performance between tasks is used to ident... |
This paper investigates the Class-Conditional Distribution (CCD) shift issue in long-tailed recognition due to scarce instances, which exhibits a significant discrepancy between the empirical CCDs for training and test data, especially for tail classes. To alleviate the issue, this paper presents a data augmentation ap... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper investigates the Class-Conditional Distribution (CCD) shift issue in long-tailed recognition due to scarce instances, which exhibits a significant discrepancy between the empirical CCDs for training and test data, especially for tail classes. To alleviate the issue, this paper presents a data augment... |
This paper proposed a class of efficient adaptive bilevel optimization methods based on momentum techniques to solve the nonconvex-strongly-convex bilevel optimization problems. Moreover, it studied the convergence properties of the proposed methods, and provided the solid convergence analysis. It also conducted the em... | Recommendation: 8: accept, good paper | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper proposed a class of efficient adaptive bilevel optimization methods based on momentum techniques to solve the nonconvex-strongly-convex bilevel optimization problems. Moreover, it studied the convergence properties of the proposed methods, and provided the solid convergence analysis. It also conducte... |
This paper proposes a method for identifying the mismatch between a source domain and a target domain for transfer learning. The method relies on first identifying cluster centers using all the training data, and then comparing the current input data example to these center for determining how much compensation should ... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper proposes a method for identifying the mismatch between a source domain and a target domain for transfer learning. The method relies on first identifying cluster centers using all the training data, and then comparing the current input data example to these center for determining how much compensation... |
This paper investigates the problem of minimizing prequential description lengths for image datasets with neural networks, primarily dealing with the issue of computational overhead. The proposed method (Mini-batch Incremental Training with Replay Streams, MI/RS) is inspired by continual learning and its effectiveness... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper investigates the problem of minimizing prequential description lengths for image datasets with neural networks, primarily dealing with the issue of computational overhead. The proposed method (Mini-batch Incremental Training with Replay Streams, MI/RS) is inspired by continual learning and its effec... |
The paper discusses how implementations of Randomized Smoothing (RS), an algorithm for certified robustness, can become unsound due to floating point arithmetic.
RS performs multiple model evaluations under noise to determine a robust output. The finite precision of IEEE-754 floating point numbers causes the addition o... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper discusses how implementations of Randomized Smoothing (RS), an algorithm for certified robustness, can become unsound due to floating point arithmetic.
RS performs multiple model evaluations under noise to determine a robust output. The finite precision of IEEE-754 floating point numbers causes the ad... |
This paper studies the disentangled representation learning. In particular, authors provide the evidence on the improved generalization when the disentangled representations coupled with sparse base-predictors are learned. The paper then presents a theoretical result on identifiability condition wonder with maximally ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper studies the disentangled representation learning. In particular, authors provide the evidence on the improved generalization when the disentangled representations coupled with sparse base-predictors are learned. The paper then presents a theoretical result on identifiability condition wonder with ma... |
This paper proposed a time-varying integration method for accelerating the mixing of Hamiltonian Monte Carlo. In particular, when the potential $f$ is quadratic function, i.e., the target distribution is Gaussian, the ideal HMC with the proposed time-varying integration enjoy a $O(\sqrt{\kappa}\log(1/\epsilon))$ iterat... | Recommendation: 5: marginally below the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper proposed a time-varying integration method for accelerating the mixing of Hamiltonian Monte Carlo. In particular, when the potential $f$ is quadratic function, i.e., the target distribution is Gaussian, the ideal HMC with the proposed time-varying integration enjoy a $O(\sqrt{\kappa}\log(1/\epsilon))... |
This paper presents an unsupervised method to train deep neural networks using Hebbian/Anti-hebbian learning on 3D object classification from point clouds. The authors highlight (empirically and theoreticalaly) key disadvantages of using just Hebbian or Anti-hebbian learning; they demonstrate the improved expressivity ... | Recommendation: 8: accept, good paper | Area: Unsupervised and Self-supervised learning | Review:
This paper presents an unsupervised method to train deep neural networks using Hebbian/Anti-hebbian learning on 3D object classification from point clouds. The authors highlight (empirically and theoreticalaly) key disadvantages of using just Hebbian or Anti-hebbian learning; they demonstrate the improved expre... |
This work adopts a language modeling approach to tackle program type prediction. Specifically, it fine-tunes CodeT5 on a type annotation task, and extends the standard approach by both including richer sets of context inferred via static analysis, and ordering the generation process to simulate information flow through... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This work adopts a language modeling approach to tackle program type prediction. Specifically, it fine-tunes CodeT5 on a type annotation task, and extends the standard approach by both including richer sets of context inferred via static analysis, and ordering the generation process to simulate information flow... |
This paper proposes a simple yet effective learning module upon the framework of contrastive reinforcement learning. The module could automatically learn a optimal transformation for data augmentation. The algorithm is theoretically founded and also demonstrated by expeiments on DMControl and Atari 100k.
Strength:
- T... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes a simple yet effective learning module upon the framework of contrastive reinforcement learning. The module could automatically learn a optimal transformation for data augmentation. The algorithm is theoretically founded and also demonstrated by expeiments on DMControl and Atari 100k.
Streng... |
This paper proposes targeted and transferable adversarial examples for self-supervised asr models which are in pretraining + fine-tuning architecture. An adversary can make use of the transferability property, that is, an adversarial sample produced for a proxy asr can also fool a different remote asr.
Rich self-superv... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes targeted and transferable adversarial examples for self-supervised asr models which are in pretraining + fine-tuning architecture. An adversary can make use of the transferability property, that is, an adversarial sample produced for a proxy asr can also fool a different remote asr.
Rich sel... |
In this paper, the authors propose the node representation method GLEM for Text-Attributed Graphs (TAG) based on a pseudo-likelihood variational framework.
Specifically, they alternatively update GNN (M-step) and LM (E-step) with pseudo-labels from each other.
In experiments, they validate their method achieves SOTA ... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
In this paper, the authors propose the node representation method GLEM for Text-Attributed Graphs (TAG) based on a pseudo-likelihood variational framework.
Specifically, they alternatively update GNN (M-step) and LM (E-step) with pseudo-labels from each other.
In experiments, they validate their method achiev... |
This paper proposes an exploration strategy for reinforcement learning. This exploration strategy is based on the differential extrinsic plasticity rule in neuroscience which outputs temporal correlated and state-dependent exploration noise. The exploration strategy can also be parameterized as a neural network and thi... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes an exploration strategy for reinforcement learning. This exploration strategy is based on the differential extrinsic plasticity rule in neuroscience which outputs temporal correlated and state-dependent exploration noise. The exploration strategy can also be parameterized as a neural network... |
Authors propose an algorithm called D-CIPHER to elicit differential equations from data. Authors also propose a new optimization procedure called CoLLie, to help D-CIPHER. Empirical examples are provided.
Strengths:
Authors have a good grasp of some of the challenges of eliciting differential equations from data... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
Authors propose an algorithm called D-CIPHER to elicit differential equations from data. Authors also propose a new optimization procedure called CoLLie, to help D-CIPHER. Empirical examples are provided.
Strengths:
Authors have a good grasp of some of the challenges of eliciting differential equations f... |
This work proposes VA-DepthNet to solve single image depth prediction problem by exploiting classical first-order variational constraints. The proposed network disentangles the absolute scale from the metric depth and models unscaled depth map as the optimal solution to the pixel-level depth gradiant. The network focus... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This work proposes VA-DepthNet to solve single image depth prediction problem by exploiting classical first-order variational constraints. The proposed network disentangles the absolute scale from the metric depth and models unscaled depth map as the optimal solution to the pixel-level depth gradiant. The netwo... |
This paper proposes a topology-guided sampling strategy (TGSS) to mitigate the gap between sampled data within a mini-batch and global data. The proposed model, which is called TopoZero, consists of a topology alignment module (TAM) and a distribution alignment module. TAM is capable of preserving multidimensional geom... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a topology-guided sampling strategy (TGSS) to mitigate the gap between sampled data within a mini-batch and global data. The proposed model, which is called TopoZero, consists of a topology alignment module (TAM) and a distribution alignment module. TAM is capable of preserving multidimensio... |
This paper argues that existing doubly robust (DR) estimators usually have a large variance, a large bias, and poor generalization ability given inaccurate error imputation. The paper proposes a novel estimator, called targeted doubly robust (TDR), to simultaneously reduce the bias and variance of existing DR estimator... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper argues that existing doubly robust (DR) estimators usually have a large variance, a large bias, and poor generalization ability given inaccurate error imputation. The paper proposes a novel estimator, called targeted doubly robust (TDR), to simultaneously reduce the bias and variance of existing DR e... |
The paper proposes a model-level explanation method for GNNs. The core idea is, given a class, to maximize a graph that has the highest softmax-activation for that class. In order to obtain graphs that are meaningful, a regularization term is added forcing similarity of the graph embedding with the average embeddings f... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper proposes a model-level explanation method for GNNs. The core idea is, given a class, to maximize a graph that has the highest softmax-activation for that class. In order to obtain graphs that are meaningful, a regularization term is added forcing similarity of the graph embedding with the average embe... |
The paper studies an apparently new problem, called Catalog Problem, which looks for grouping of items and ordering of groups. The paper further proposes a neural network-based model called Neural Ordered Clusters for solving the problem. The empirical results show apparently superior performance of the proposed meth... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The paper studies an apparently new problem, called Catalog Problem, which looks for grouping of items and ordering of groups. The paper further proposes a neural network-based model called Neural Ordered Clusters for solving the problem. The empirical results show apparently superior performance of the propo... |
The paper introduces a novel, efficient method for computing multimarginal optimal transport plans. The approach applies to cases with Monge cost, which while a special case, the authors argue is relatively general. A generalized earth-mover distance is introduced, which is linear in the dimensions and bins in theory, ... | Recommendation: 8: accept, good paper | Area: Optimization (eg, convex and non-convex optimization) | Review:
The paper introduces a novel, efficient method for computing multimarginal optimal transport plans. The approach applies to cases with Monge cost, which while a special case, the authors argue is relatively general. A generalized earth-mover distance is introduced, which is linear in the dimensions and bins in ... |
This paper provides theoretical analysis to guarantee the performance of the CCEM criterion for the first time. Furthermore, to improve the identifiability of the classifiers and the workers, the authors also proposed two variants of the CCEM by introducing two different regularization terms. Experimental results on se... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper provides theoretical analysis to guarantee the performance of the CCEM criterion for the first time. Furthermore, to improve the identifiability of the classifiers and the workers, the authors also proposed two variants of the CCEM by introducing two different regularization terms. Experimental resul... |
This paper proposes a method for finding the subset of a large language model (LLM) training corpus that is most responsible for the 0-shot performance of the LLM on a given task. The method involves finding examples whose gradient is similar to the gradient of task-specific supervised data. This is done in an iterativ... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a method for finding the subset of a large language model (LLM) training corpus that is most responsible for the 0-shot performance of the LLM on a given task. The method involves finding examples whose gradient is similar to the gradient of task-specific supervised data. This is done in an ... |
The authors propose an adaption of transformer attention to graph input. Normally, a transformer takes an unordered set as input, which can not adequately represent the structured nature of a graph. The key change they propose is the inclusion of the directed edge between nodes into the attention mechanism.
The author... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The authors propose an adaption of transformer attention to graph input. Normally, a transformer takes an unordered set as input, which can not adequately represent the structured nature of a graph. The key change they propose is the inclusion of the directed edge between nodes into the attention mechanism.
Th... |
This paper presents highway RL, a novel dynamic programming-based strategy for off-policy RL that incorporates multistep information without requiring importance weights. Highway RL replaces the traditional Bellman backup by the maximum over both $n$ and $\pi$ of the $n$-step return of $\pi$ followed by the maximum ove... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper presents highway RL, a novel dynamic programming-based strategy for off-policy RL that incorporates multistep information without requiring importance weights. Highway RL replaces the traditional Bellman backup by the maximum over both $n$ and $\pi$ of the $n$-step return of $\pi$ followed by the max... |
The paper examines concept bottleneck models and focuses on the case where these models have unexpected behaviour when there is insufficient concept information. The authors propose a strategy for decoupling CBMs into explicit and implicit concepts while retaining high predictive performance and interpretability. They ... | Recommendation: 3: reject, not good enough | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper examines concept bottleneck models and focuses on the case where these models have unexpected behaviour when there is insufficient concept information. The authors propose a strategy for decoupling CBMs into explicit and implicit concepts while retaining high predictive performance and interpretabilit... |
The authors have developed a (non-parametric) algorithm for Differentially Private Linear Regression problem. The model works without hyperparameters and privacy bounds using "Tukey depth". They call this method as TukeyEM. TukeyEM works in four steps: (i) splitting data into subsets, (ii) computing columns using appro... | 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 authors have developed a (non-parametric) algorithm for Differentially Private Linear Regression problem. The model works without hyperparameters and privacy bounds using "Tukey depth". They call this method as TukeyEM. TukeyEM works in four steps: (i) splitting data into subsets, (ii) computing columns usi... |
This paper proposes a topology augmentation method for graph contrastive learning which explores invariance of graphs from the spectral perspective. To realize the spectral invariance, the paper aims to identify sensitive edges whose perturbation leads to a large spectral difference. The conjecture is that the GNN enco... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper proposes a topology augmentation method for graph contrastive learning which explores invariance of graphs from the spectral perspective. To realize the spectral invariance, the paper aims to identify sensitive edges whose perturbation leads to a large spectral difference. The conjecture is that the ... |
This paper proposes a new learning rule, named Auxiliary Activation Learning to reduce memory requirements of deep neural networks
Strengths:
- The proposed strategy seems to reduce memory space (11.8x) for ResNet, while not having a big influence in training time.
- Extensive evaluation on training times and memory sa... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a new learning rule, named Auxiliary Activation Learning to reduce memory requirements of deep neural networks
Strengths:
- The proposed strategy seems to reduce memory space (11.8x) for ResNet, while not having a big influence in training time.
- Extensive evaluation on training times and m... |
This paper presents am empirical comparison between the performance of various neural kernels used in GPs for contextual bandit optimisation. Comparisons are done using a modified wheel dataset, and the use of the student t-process to improve exploration in NK bandits is explored.
The background is interesting and the... | Recommendation: 3: reject, not good enough | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
This paper presents am empirical comparison between the performance of various neural kernels used in GPs for contextual bandit optimisation. Comparisons are done using a modified wheel dataset, and the use of the student t-process to improve exploration in NK bandits is explored.
The background is interesting... |
The paper proposes a self-supervised technique for vessel segmentation based on a combination of a denoising diffusion probabilistic model and adversarial learning. The network uses contrast and non-constrast images of vessels and uses the non-contrast images to learn a distribution of the background that is then used ... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
The paper proposes a self-supervised technique for vessel segmentation based on a combination of a denoising diffusion probabilistic model and adversarial learning. The network uses contrast and non-constrast images of vessels and uses the non-contrast images to learn a distribution of the background that is th... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.