review stringlengths 5 16.6k | score stringclasses 5
values | area stringclasses 12
values | text stringlengths 31 5.65k |
|---|---|---|---|
The main contribution of the paper is a hierarchical architecture for predicting observations at different levels of temporal coarseness, and its evaluation for RL from images.
The individual dynamic models for the different temporal coarseness levels are independent from each other and each consist of an encoder, a l... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The main contribution of the paper is a hierarchical architecture for predicting observations at different levels of temporal coarseness, and its evaluation for RL from images.
The individual dynamic models for the different temporal coarseness levels are independent from each other and each consist of an enco... |
This paper attempts to analyze a binary neural classifier's input-output relationship.
A straightforward, brute-force sampling is forbidden due to the high dimension and continuity of input distribution.
To achieve efficiency, the authors reformulate the problem as density of states (DOS) sampling in physics, and propo... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper attempts to analyze a binary neural classifier's input-output relationship.
A straightforward, brute-force sampling is forbidden due to the high dimension and continuity of input distribution.
To achieve efficiency, the authors reformulate the problem as density of states (DOS) sampling in physics, a... |
The authors propose a 3 steps image editing pipeline that 1) generate segmentation mask from the change in text query, 2) encode the image with diffusion process until a time step ‘r’, and 3) decode it back to the image via reverse diffusion process condition on the text query, and with mask as the guidance. They show ... | Recommendation: 5: marginally below the acceptance threshold | Area: Generative models | Review:
The authors propose a 3 steps image editing pipeline that 1) generate segmentation mask from the change in text query, 2) encode the image with diffusion process until a time step ‘r’, and 3) decode it back to the image via reverse diffusion process condition on the text query, and with mask as the guidance. Th... |
This work extended the Predictor-Corrector sampling algorithm originally developed in the continuous diffusion models to the discrete case, which is termed discrete Predictor-Corrector (DPC). Since, in the discrete diffusion models, there is no direct counterpart to the score function, this work proposed to learn a new... | Recommendation: 8: accept, good paper | Area: Generative models | Review:
This work extended the Predictor-Corrector sampling algorithm originally developed in the continuous diffusion models to the discrete case, which is termed discrete Predictor-Corrector (DPC). Since, in the discrete diffusion models, there is no direct counterpart to the score function, this work proposed to lea... |
This paper proposes and effective strategy to jointly scale vision and language models that works on vision, language and vision & language multimodal tasks. The model is trained on a mixture of several pretraining tasks on a new dataset containing 1.6B samples in over 100 different languages. The model is simple and m... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper proposes and effective strategy to jointly scale vision and language models that works on vision, language and vision & language multimodal tasks. The model is trained on a mixture of several pretraining tasks on a new dataset containing 1.6B samples in over 100 different languages. The model is simp... |
This paper proposes a generative model in terms of image restoration, named restoration-based generative model (RGM). This paper eliminates expensive sampling by performing MAP estimation and incorporating implicit prior information via GAN. Furthermore, a multi-scale training is proposed to alleviate the latent ineffi... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
This paper proposes a generative model in terms of image restoration, named restoration-based generative model (RGM). This paper eliminates expensive sampling by performing MAP estimation and incorporating implicit prior information via GAN. Furthermore, a multi-scale training is proposed to alleviate the laten... |
This paper proposes to learn new tasks in low-coherence subspaces rather than orthogonal subspaces to mitigate catastrophic forgetting in continual learning. The authors believe that Gradient Orthogonal Projection (GOP) (though helps battle catastrophic forgetting) causes learning capacity degradation and its learning... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes to learn new tasks in low-coherence subspaces rather than orthogonal subspaces to mitigate catastrophic forgetting in continual learning. The authors believe that Gradient Orthogonal Projection (GOP) (though helps battle catastrophic forgetting) causes learning capacity degradation and its ... |
This submission deals with DDIMs. It proposes a generalization of DDIMs to general diffusion models beyond the isotropic denoising by modifying the parameterization of score. It links DDIMs to SDEs by a score approximation, and provides an interpretation for the success of DDIMs. It validates the results for BDM and CL... | Recommendation: 8: accept, good paper | Area: Generative models | Review:
This submission deals with DDIMs. It proposes a generalization of DDIMs to general diffusion models beyond the isotropic denoising by modifying the parameterization of score. It links DDIMs to SDEs by a score approximation, and provides an interpretation for the success of DDIMs. It validates the results for BD... |
The authors develop an interpretability framework based on information theory. The advantage/novelty of the approach being that it works directly on the 'informativeness' of the individual features and their joint interactions w.r.t. a given target under potentially a variety of losses and a variety of neural network a... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The authors develop an interpretability framework based on information theory. The advantage/novelty of the approach being that it works directly on the 'informativeness' of the individual features and their joint interactions w.r.t. a given target under potentially a variety of losses and a variety of neural n... |
This paper studies the problem of object rearrangement and focuses particularly on the design of an object-centric planner capable of solving rearrangement tasks requiring multi-step reasoning and systematic generalization. A method, NCS, for combining object-centric representation learning with non-parametric planning... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper studies the problem of object rearrangement and focuses particularly on the design of an object-centric planner capable of solving rearrangement tasks requiring multi-step reasoning and systematic generalization. A method, NCS, for combining object-centric representation learning with non-parametric ... |
This work proposes a meta-optimised mirror descent scheme for large-scale optimization problems, based on tuning of a parametric Bregman divergence on a set of curriculum meta-training task. The expected benefits of this meta-learning procedure is to produce an optimiser with accelerated convergence that is transferabl... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This work proposes a meta-optimised mirror descent scheme for large-scale optimization problems, based on tuning of a parametric Bregman divergence on a set of curriculum meta-training task. The expected benefits of this meta-learning procedure is to produce an optimiser with accelerated convergence that is tra... |
This work focuses on adaptive gradient optimization algorithms and attempts to provide practical insights for speeding up convergence with such methods. In particular, the authors proposed the Win method to increase the speed of convergence for adaptive optimization techniques. In particular, this technique incorporate... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This work focuses on adaptive gradient optimization algorithms and attempts to provide practical insights for speeding up convergence with such methods. In particular, the authors proposed the Win method to increase the speed of convergence for adaptive optimization techniques. In particular, this technique inc... |
This paper is concerned with probabilistic predictions: Instead of predicting a class, the prediction is a distribution over the label space. The paper proceeds by using a probabilistic loss functional, and consider an adversarial training approach. The proposed method is a minimax problem, with perturbations in a mome... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
This paper is concerned with probabilistic predictions: Instead of predicting a class, the prediction is a distribution over the label space. The paper proceeds by using a probabilistic loss functional, and consider an adversarial training approach. The proposed method is a minimax problem, with perturbations i... |
The authors conduct a study of convolutional auto-encoders to assess the importance of different dimensions of the autoencoder bottleneck on model quality for a range of tasks. They additionally assess whether or not these models learn identity functions when the number of pixels in the input matches the number of feat... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
The authors conduct a study of convolutional auto-encoders to assess the importance of different dimensions of the autoencoder bottleneck on model quality for a range of tasks. They additionally assess whether or not these models learn identity functions when the number of pixels in the input matches the number... |
The authors propose to use second order method only on the loss function -- not throughout the whole network in the training process.
They are also putting a L2 term in the formulation.
I don't see a strength in the formulation. Personally I have tried this method without introducing the L2 term and my finding is not q... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The authors propose to use second order method only on the loss function -- not throughout the whole network in the training process.
They are also putting a L2 term in the formulation.
I don't see a strength in the formulation. Personally I have tried this method without introducing the L2 term and my finding ... |
This paper is about analyzing the behavior of stochastic gradient descent (SGD) using Kolmogorov complexity.
The authors are using an “entropy compression argument” to prove the termination (and convergence) of SGD under weaker assumptions as well as provide a lower bound on the amount of randomness needed for SGD to... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper is about analyzing the behavior of stochastic gradient descent (SGD) using Kolmogorov complexity.
The authors are using an “entropy compression argument” to prove the termination (and convergence) of SGD under weaker assumptions as well as provide a lower bound on the amount of randomness needed fo... |
This work proposes DeepHV, a deep neural network based model, to approximate hypervolume for multi-objective optimization. The idea of model-based hypervolume approximation has been proposed in the previous work (HVNet). This work's contribution is to design a more powerful equivariant layer to further exploit hypervol... | Recommendation: 6: marginally above the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
This work proposes DeepHV, a deep neural network based model, to approximate hypervolume for multi-objective optimization. The idea of model-based hypervolume approximation has been proposed in the previous work (HVNet). This work's contribution is to design a more powerful equivariant layer to further exploit ... |
Current deep and physics-informed machine learning models struggle to correctly forecast dynamical systems in out-of-distribution (OOD) settings. This paper proposes a new approach consisting of a meta-leaning strategy combined with causal structural discovery. The method is evaluated on standard ODE benchmarks, showin... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
Current deep and physics-informed machine learning models struggle to correctly forecast dynamical systems in out-of-distribution (OOD) settings. This paper proposes a new approach consisting of a meta-leaning strategy combined with causal structural discovery. The method is evaluated on standard ODE benchmarks... |
This paper considers learning a quantisation of colour space, and the parallels to how humans learn colour names.
A novel deep network architecture for colour quantisation is proposed, which is learned by optimising accuracy on a task of interest (image classification or object detection) when using images with quanti... | Recommendation: 3: reject, not good enough | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper considers learning a quantisation of colour space, and the parallels to how humans learn colour names.
A novel deep network architecture for colour quantisation is proposed, which is learned by optimising accuracy on a task of interest (image classification or object detection) when using images wit... |
This paper tries to detect adversarial examples from a perspective that models residual networks as discrete dynamical systems. The detector studies trajectories of samples in space, through time, to distinguish between clean and adversarial examples. Based on this rationale, the authors also apply transport regulariza... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper tries to detect adversarial examples from a perspective that models residual networks as discrete dynamical systems. The detector studies trajectories of samples in space, through time, to distinguish between clean and adversarial examples. Based on this rationale, the authors also apply transport re... |
As implied by the author, multi-agent systems can be studied in the sense of social behavior. Where each agent needs to be an expert in a particular task, they should share a common understanding of the whole space with other agents. To this end, a wide variety of methods are proposed to fulfill both of these criteria.... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
As implied by the author, multi-agent systems can be studied in the sense of social behavior. Where each agent needs to be an expert in a particular task, they should share a common understanding of the whole space with other agents. To this end, a wide variety of methods are proposed to fulfill both of these c... |
This paper presents context autoencoder (CAE), which follows BEiT and uses an additional latent contextual regressor to make predictions for the masked patches with cross attention over the visible patches. By doing that we can encourage the masked patches' representation predicted from visible patches are aligned with... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper presents context autoencoder (CAE), which follows BEiT and uses an additional latent contextual regressor to make predictions for the masked patches with cross attention over the visible patches. By doing that we can encourage the masked patches' representation predicted from visible patches are alig... |
This paper proposes offline to online optimistic finetuning (O3F), an algorithm for offline to online RL. The algorithm maintains the same offline learning objective during online learning, but uses a different action distribution that selects actions with potentially higher Q values. This is achieved by simple adding ... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes offline to online optimistic finetuning (O3F), an algorithm for offline to online RL. The algorithm maintains the same offline learning objective during online learning, but uses a different action distribution that selects actions with potentially higher Q values. This is achieved by simple... |
The paper introduces CANIFE, a method for auditing the empirical privacy guarantees of federated learning algorithms. In short, CANIFE does the following: It considers a canary client with a single data point that contributes in k "frozen" rounds, and doesn't contribute in k other "frozen" rounds (frozen means the glo... | Recommendation: 5: marginally below the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper introduces CANIFE, a method for auditing the empirical privacy guarantees of federated learning algorithms. In short, CANIFE does the following: It considers a canary client with a single data point that contributes in k "frozen" rounds, and doesn't contribute in k other "frozen" rounds (frozen means... |
This paper presents a simple method of prompt learning for Vision-Langauge model (in particular the CLIP). It is an extension based on existing work CoOp, where it leverages learnable embeddings as textual input and performs classification on this basis. Empirically, DoFo shows great improvement, where it reaches 73.2%... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper presents a simple method of prompt learning for Vision-Langauge model (in particular the CLIP). It is an extension based on existing work CoOp, where it leverages learnable embeddings as textual input and performs classification on this basis. Empirically, DoFo shows great improvement, where it reach... |
This paper focuses on Federated learning (FL) with heterogeneous data distributions. In this case, a globally shared model may not achieve a good local generalization performance. So, the author put forward a method that each node can still exploit global data through FL but, at the same time identify and collaborate o... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
This paper focuses on Federated learning (FL) with heterogeneous data distributions. In this case, a globally shared model may not achieve a good local generalization performance. So, the author put forward a method that each node can still exploit global data through FL but, at the same time identify and colla... |
The paper provides an empirical study about a setting where a portion of the offline RL dataset doesn't include actions. To exploit action-missing data, this work proposes to learn an inverse dynamics model on data with actions to generate proxy actions from state transitions. A set of empirical studies on d4rl gym-loc... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper provides an empirical study about a setting where a portion of the offline RL dataset doesn't include actions. To exploit action-missing data, this work proposes to learn an inverse dynamics model on data with actions to generate proxy actions from state transitions. A set of empirical studies on d4rl... |
This paper investigates the effect of pre-training dataset on transfer learning, when pretraining by text caption contrastive matching using CLIP. Several different pre-training datasets are compared, using six transfer task datasets (three in later experiments), and in both few-shot and full-data scenarios. The expe... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper investigates the effect of pre-training dataset on transfer learning, when pretraining by text caption contrastive matching using CLIP. Several different pre-training datasets are compared, using six transfer task datasets (three in later experiments), and in both few-shot and full-data scenarios. ... |
A generative model generates molecules conditioned on a protein pocket, and optionally also conditioned on a reference molecule or some desired substructures. Atom types, bonds, and atom positions are generated one at a time using an autoregressive flow model.
The key novelty here seems to be the idea of conditioning t... | Recommendation: 3: reject, not good enough | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
A generative model generates molecules conditioned on a protein pocket, and optionally also conditioned on a reference molecule or some desired substructures. Atom types, bonds, and atom positions are generated one at a time using an autoregressive flow model.
The key novelty here seems to be the idea of condit... |
This paper performs a mixture of things to try to visualize ViTs as opposed to CNNs. First, they slightly extended the methodology of (Simonyan et al. 2014) by adding some data augmentations, and ran the optimization to find images that maximally activate a certain filter. This is met largely with adversarial examples,... | 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 performs a mixture of things to try to visualize ViTs as opposed to CNNs. First, they slightly extended the methodology of (Simonyan et al. 2014) by adding some data augmentations, and ran the optimization to find images that maximally activate a certain filter. This is met largely with adversarial e... |
The paper presents a privacy-preserving federated learning protocol. Unlike prior works, it avoids running the training in a privacy-preserving manner but instead uses local training on sample-answer pairs that have been obtained using private inference.
This is an interesting idea, but not enough attention is given t... | Recommendation: 3: reject, not good enough | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper presents a privacy-preserving federated learning protocol. Unlike prior works, it avoids running the training in a privacy-preserving manner but instead uses local training on sample-answer pairs that have been obtained using private inference.
This is an interesting idea, but not enough attention is... |
Discovering latent relationships within training samples is an important problem that has been studied in depth before. In this work, the authors propose solving the partial label learning (PPL) problem with multi-agent reinforcement learning (RL). Their solution uses an attention-based graph neural network (GNN) to le... | Recommendation: 3: reject, not good enough | Area: Unsupervised and Self-supervised learning | Review:
Discovering latent relationships within training samples is an important problem that has been studied in depth before. In this work, the authors propose solving the partial label learning (PPL) problem with multi-agent reinforcement learning (RL). Their solution uses an attention-based graph neural network (GN... |
The paper studies the few-shot KG completion problem and proposes a hierarchical relation learning method for this task. The authors propose three levels of relational information to learn and refine the meta-representation of few-shot relations. Specifically, a contrastive learning-based context-level relation learnin... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper studies the few-shot KG completion problem and proposes a hierarchical relation learning method for this task. The authors propose three levels of relational information to learn and refine the meta-representation of few-shot relations. Specifically, a contrastive learning-based context-level relation... |
This paper tackles the problem of identifying runtime errors in a program in a static setting. Their proposed model is a modification of an Instruction Pointer Attention GNN (IPA-GNN) proposed in Bieber et al. (2020) [1], which learns to execute the program one instruction at a time in a "continuous" manner using embed... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper tackles the problem of identifying runtime errors in a program in a static setting. Their proposed model is a modification of an Instruction Pointer Attention GNN (IPA-GNN) proposed in Bieber et al. (2020) [1], which learns to execute the program one instruction at a time in a "continuous" manner usi... |
This paper considers a problem of person re-identification in a more general setting, i.e., domain generalizable person ReID without demographics. To address this, it introduces distributionally robust optimization (DRO) to learn robust person ReID models that perform well on all possible data distributions within the ... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper considers a problem of person re-identification in a more general setting, i.e., domain generalizable person ReID without demographics. To address this, it introduces distributionally robust optimization (DRO) to learn robust person ReID models that perform well on all possible data distributions wit... |
This paper proposes a MLP based spiking neural network, termed spiking MLP-Mixer, which contains the Spiking Token Block, Spiking Channel Block, and Speaking MLP. The authors state that they achieve good performance on the Image-1k dataset, Cifar10, Cifar100. This work suggests the importance of integrating global and ... | Recommendation: 3: reject, not good enough | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
This paper proposes a MLP based spiking neural network, termed spiking MLP-Mixer, which contains the Spiking Token Block, Spiking Channel Block, and Speaking MLP. The authors state that they achieve good performance on the Image-1k dataset, Cifar10, Cifar100. This work suggests the importance of integrating glo... |
This paper proposes a new evaluation setting for Continual Learning in the setting of large number of tasks with recurring tasks. The authors propose SCoLe, an experimental framework for this setting and report their findings throughout the manuscript. The key claimed insight is the observation that Continual Learning ... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper proposes a new evaluation setting for Continual Learning in the setting of large number of tasks with recurring tasks. The authors propose SCoLe, an experimental framework for this setting and report their findings throughout the manuscript. The key claimed insight is the observation that Continual L... |
This paper introduces a method for learning diffusion models over neural fields. More specifically, the authors perform diffuson directly on coordinate (such as pixel locations) and feature (such as RGB values) pairs, allowing the model to handle a wide variety of data that can be expressed as a neural field. The autho... | Recommendation: 8: accept, good paper | Area: Generative models | Review:
This paper introduces a method for learning diffusion models over neural fields. More specifically, the authors perform diffuson directly on coordinate (such as pixel locations) and feature (such as RGB values) pairs, allowing the model to handle a wide variety of data that can be expressed as a neural field. T... |
The paper investigates the near anomalies detection and points a significant decrease in performance of state of the art approaches on near anomalies cases. A new solution is proposed: 1) using SDE models trained to generate data from the training distribution. 2) generate a dataset from the trained SDE. 3) train a bin... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
The paper investigates the near anomalies detection and points a significant decrease in performance of state of the art approaches on near anomalies cases. A new solution is proposed: 1) using SDE models trained to generate data from the training distribution. 2) generate a dataset from the trained SDE. 3) tra... |
This paper aims to address two issues regarding the Clip-like model: 1) degraded accuracy and robustness when inferring by retrieving textual class names (the zero-shot protocol); 2) breaking the well-established vision-language alignment (linear probing). To combine the best of both worlds, this paper proposes Decompo... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper aims to address two issues regarding the Clip-like model: 1) degraded accuracy and robustness when inferring by retrieving textual class names (the zero-shot protocol); 2) breaking the well-established vision-language alignment (linear probing). To combine the best of both worlds, this paper proposes... |
This paper investigates the perceptual straightness of a large set of modern vision models. Authors start from the observation that representation straightness is a known property of biological vision. They then define a metric to capture the average output curvature on a video and evaluate a wide range of models (resn... | Recommendation: 6: marginally above the acceptance threshold | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
This paper investigates the perceptual straightness of a large set of modern vision models. Authors start from the observation that representation straightness is a known property of biological vision. They then define a metric to capture the average output curvature on a video and evaluate a wide range of mode... |
This paper proposes an end-to-end method so-called ``Jointist'' to simultaneously address 3 audio tasks consisting respectively
in separating the source signals, transcribing the music and recognizing the instruments.
The proposed method is a deep neural network which combines together 3 existing neural network archit... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes an end-to-end method so-called ``Jointist'' to simultaneously address 3 audio tasks consisting respectively
in separating the source signals, transcribing the music and recognizing the instruments.
The proposed method is a deep neural network which combines together 3 existing neural networ... |
This paper focuses on the task of noisy label learning (NLL), which proposes set-level self-supervised learning (SLSSL) to model noisy label data.
SLSSL performs self-supervised learning at mini-batch levels with observed noisy labels.
The proposed method relabel the samples from label i using label j, which is call... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper focuses on the task of noisy label learning (NLL), which proposes set-level self-supervised learning (SLSSL) to model noisy label data.
SLSSL performs self-supervised learning at mini-batch levels with observed noisy labels.
The proposed method relabel the samples from label i using label j, which... |
The paper presents a novel approach to MIR tasks which uses a modular setup combining instrument recognition, transcription, and source separation models into a joint pipeline. The authors perform experiments to evaluate the proposed approach, but the results are mixed (with the only improvement being in a listening st... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper presents a novel approach to MIR tasks which uses a modular setup combining instrument recognition, transcription, and source separation models into a joint pipeline. The authors perform experiments to evaluate the proposed approach, but the results are mixed (with the only improvement being in a list... |
This paper studies the rank behavior of deep neural networks with three different types of rank definitions: the maximum of network Jacobi rank, bottlenet rank, and representation cost. The authors rigorously demonstrate the intrinsic connections among those three concepts, revealing a fascinating property of general d... | Recommendation: 8: accept, good paper | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper studies the rank behavior of deep neural networks with three different types of rank definitions: the maximum of network Jacobi rank, bottlenet rank, and representation cost. The authors rigorously demonstrate the intrinsic connections among those three concepts, revealing a fascinating property of g... |
This paper studies out-of-distribution detection, aiming at making classification models excel at discerning ID and OOD data. It is an important problem for safety-critical applications, and has attracted increasing attention recently. The authors claim that they adopt an original approach based on the functional view ... | Recommendation: 6: marginally above the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper studies out-of-distribution detection, aiming at making classification models excel at discerning ID and OOD data. It is an important problem for safety-critical applications, and has attracted increasing attention recently. The authors claim that they adopt an original approach based on the function... |
This paper studies the Selective classification problem when data contains many noisy samples which should be filtered out. The authors present a novel loss function to train the selector g (used to classify whether data samples are informative or uninformative) given predictor f. They also prove that the optimal selec... | Recommendation: 5: marginally below the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper studies the Selective classification problem when data contains many noisy samples which should be filtered out. The authors present a novel loss function to train the selector g (used to classify whether data samples are informative or uninformative) given predictor f. They also prove that the optim... |
This paper proposes a simple modification of multi-view augmentation for use with representation learning, as well as an additional training framework to go with it. The main idea is to create a richer set of background with higher resolution than available with the multi-crop strategy. Because the augmented data is mo... | Recommendation: 8: accept, good paper | Area: Unsupervised and Self-supervised learning | Review:
This paper proposes a simple modification of multi-view augmentation for use with representation learning, as well as an additional training framework to go with it. The main idea is to create a richer set of background with higher resolution than available with the multi-crop strategy. Because the augmented da... |
The paper introduces a method to correct the outputs of language models using
another language model in order to satisfy semantic constraints. A function
$v(y)$ is needed to measure the quality of the output hypotheses $y$. The
corrector is trained to improve this quality while staying close to the original
hypothesis.... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper introduces a method to correct the outputs of language models using
another language model in order to satisfy semantic constraints. A function
$v(y)$ is needed to measure the quality of the output hypotheses $y$. The
corrector is trained to improve this quality while staying close to the original
hyp... |
The paper introduces a novel 3D-aware generative image model, StyleMorph, which can disentangle 3D shape, camera pose, appearance and background for high-quality image synthesis. By bridging 3D morphable models with GAN synthesis and a canonical coordinate system, dense correspondences among generated objects can be p... | Recommendation: 8: accept, good paper | Area: Generative models | Review:
The paper introduces a novel 3D-aware generative image model, StyleMorph, which can disentangle 3D shape, camera pose, appearance and background for high-quality image synthesis. By bridging 3D morphable models with GAN synthesis and a canonical coordinate system, dense correspondences among generated objects ... |
This work considers the problem of sampling transition paths between two metastable states of a molecular system. This problem is difficult as the energy barrier between the states may be large, making it computationally expensive for traditional MD simulation. This work first relates this problem to literature on opti... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This work considers the problem of sampling transition paths between two metastable states of a molecular system. This problem is difficult as the energy barrier between the states may be large, making it computationally expensive for traditional MD simulation. This work first relates this problem to literature... |
This paper proposed a new model for causal proxy learning for average dose-response function estimation of multidimensional treatments. This is done by based on combining contrastive regularizer to learn the proxy representation, and ranked weighting method to de-bias the treatment assignment mechanism estimation. The... | Recommendation: 5: marginally below the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
This paper proposed a new model for causal proxy learning for average dose-response function estimation of multidimensional treatments. This is done by based on combining contrastive regularizer to learn the proxy representation, and ranked weighting method to de-bias the treatment assignment mechanism estimat... |
This paper investigates the synthesis problem in multi-document summarization. Then, this paper proposes a simple method: Given a set of documents, the proposed method generates diverse candidates using the diverse beam search and selects one candidate that aligns with the expected aggregate property of inputs.
Strengt... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper investigates the synthesis problem in multi-document summarization. Then, this paper proposes a simple method: Given a set of documents, the proposed method generates diverse candidates using the diverse beam search and selects one candidate that aligns with the expected aggregate property of inputs.... |
The manuscript discusses the need to consider a broader class of invariances in protein modelling than the rotation and translational invariances typically considered. The paper introduces the concept of conditional invariance, defined by transformations that modify the sidechain degrees of freedom of a protein, while ... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The manuscript discusses the need to consider a broader class of invariances in protein modelling than the rotation and translational invariances typically considered. The paper introduces the concept of conditional invariance, defined by transformations that modify the sidechain degrees of freedom of a protein... |
The authors propose an Implicit Reinforcement Learning via Supervised Learning (RvS) methods by leveraging the implicit model---the composition of $\arg\min$ with a general function approximator $f_\theta$ to represent the policy ($\hat{a} = \arg\min_a f_\theta (s, a)$) [1]---instead of the traditional explicit model (... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The authors propose an Implicit Reinforcement Learning via Supervised Learning (RvS) methods by leveraging the implicit model---the composition of $\arg\min$ with a general function approximator $f_\theta$ to represent the policy ($\hat{a} = \arg\min_a f_\theta (s, a)$) [1]---instead of the traditional explicit... |
The authors present SAGE or SEMANTIC-AWARE GLOBAL EXPLANATIONS model specifically for handling named entity recognition problems. They present a method to produce highly interpretable global rules to explain NLP classifiers.
The authors have provided motivation to the problem, which is certainly important. They presen... | Recommendation: 3: reject, not good enough | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The authors present SAGE or SEMANTIC-AWARE GLOBAL EXPLANATIONS model specifically for handling named entity recognition problems. They present a method to produce highly interpretable global rules to explain NLP classifiers.
The authors have provided motivation to the problem, which is certainly important. The... |
The authors analyzed the problem of identifying differing prediction mechanisms amongst different sub-populations that are likely to differ. They provided theoretical outlining of the mechanism and proposed a novel algorithm to explore the differences. The provided empirical analysis of the explore differences and prov... | 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 analyzed the problem of identifying differing prediction mechanisms amongst different sub-populations that are likely to differ. They provided theoretical outlining of the mechanism and proposed a novel algorithm to explore the differences. The provided empirical analysis of the explore differences ... |
In this work the authors give a framework for evaluating blindspot detection methods (BDMs). The authors define BDMs as methods taking labeled examples and a classifier, and outputting "semantically meaningful" groups of points. The authors introduce a new BDM, and a toy evaluation setting in which they perform image ... | Recommendation: 3: reject, not good enough | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
In this work the authors give a framework for evaluating blindspot detection methods (BDMs). The authors define BDMs as methods taking labeled examples and a classifier, and outputting "semantically meaningful" groups of points. The authors introduce a new BDM, and a toy evaluation setting in which they perfor... |
The paper presents a human full-body avatar model that combines human NeRF and image-based rendering methods. The proposed model achieves significant improvement for unseen pose and unseen person synthesis. Extensive quantitative and qualitative evaluations have been conducted to measure the model's effectiveness.
Pape... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper presents a human full-body avatar model that combines human NeRF and image-based rendering methods. The proposed model achieves significant improvement for unseen pose and unseen person synthesis. Extensive quantitative and qualitative evaluations have been conducted to measure the model's effectivene... |
The paper introduces an evolutionary approach to train value functions in actor-critic methods. The main motivation for this approach is to provide a better value approximation, a higher correlation between the estimated gradient and the true gradient, and a lower variance. The authors propose to periodically perturb h... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper introduces an evolutionary approach to train value functions in actor-critic methods. The main motivation for this approach is to provide a better value approximation, a higher correlation between the estimated gradient and the true gradient, and a lower variance. The authors propose to periodically p... |
This paper proposes a complex design to classify Images: it involves w2v, bert, user defined and bayesian features, that are sent to DNN,,
Results look good., but on the author'ś data.
Strengths
it merges important ideas
the results are good.
Weaknesses¨:
Why? Often decisions are not clearly explained, eg Bert +/... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
This paper proposes a complex design to classify Images: it involves w2v, bert, user defined and bayesian features, that are sent to DNN,,
Results look good., but on the author'ś data.
Strengths
it merges important ideas
the results are good.
Weaknesses¨:
Why? Often decisions are not clearly explained, eg... |
The authors present a GNN architecture combining Structured State Spaces model and graph structure learning for spatiotemporal modeling of multivariate signals.
The proposed model has two major advantages: (1) it leverages S4 to capture long-range temporal dependencies in signals and (2) it is able to dynamically lea... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The authors present a GNN architecture combining Structured State Spaces model and graph structure learning for spatiotemporal modeling of multivariate signals.
The proposed model has two major advantages: (1) it leverages S4 to capture long-range temporal dependencies in signals and (2) it is able to dynamic... |
The paper considers the problem of choosing the clipping threshold in DP-SGD routines and proposes two related modifications to the original fixed clipping technique of Abadi et al. They then show that the DP-SGD method with the proposed clipping has desirable theoretical properties in terms of convergence to stationar... | 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 considers the problem of choosing the clipping threshold in DP-SGD routines and proposes two related modifications to the original fixed clipping technique of Abadi et al. They then show that the DP-SGD method with the proposed clipping has desirable theoretical properties in terms of convergence to s... |
The authors propose a method to infer the arithmetic expressions performed by a language model while solving a mathematical reasoning question with a chain-of-thought style output. The inferred expressions are then calculated outside of the LM. This aims to improve the performance of LLMs on mathematical reasoning task... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The authors propose a method to infer the arithmetic expressions performed by a language model while solving a mathematical reasoning question with a chain-of-thought style output. The inferred expressions are then calculated outside of the LM. This aims to improve the performance of LLMs on mathematical reason... |
The paper proposes a hazard gradient penalty algorithm for regularising neural ode-based survival models. Additionally, the paper formulates a theoretical connection of the proposed hazard gradient penalty to KL divergence minimization of KL ($p(t|x) || p(t| x^{\prime})$), where $x^{\prime}$ are neighbouring data poi... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper proposes a hazard gradient penalty algorithm for regularising neural ode-based survival models. Additionally, the paper formulates a theoretical connection of the proposed hazard gradient penalty to KL divergence minimization of KL ($p(t|x) || p(t| x^{\prime})$), where $x^{\prime}$ are neighbouring ... |
This paper proposes an efficient one-shot neural architecture search (NAS) method that progressively freezes the architectures. The authors show that the first few blocks become similar among the candidate solutions during the evolutionary architecture search and then develop an acceleration method of architecture sear... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
This paper proposes an efficient one-shot neural architecture search (NAS) method that progressively freezes the architectures. The authors show that the first few blocks become similar among the candidate solutions during the evolutionary architecture search and then develop an acceleration method of architect... |
This paper proposes a method for automatically constructing chain-of-thought prompts for multistep reasoning in large language models by using zero-shot chain-of-thought prompting (i.e. prepending "let's think step by step" to the generation) to generate rationales for selected prompt questions, and various heuristics ... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a method for automatically constructing chain-of-thought prompts for multistep reasoning in large language models by using zero-shot chain-of-thought prompting (i.e. prepending "let's think step by step" to the generation) to generate rationales for selected prompt questions, and various heu... |
This work proposes to use forward processes other than the Gaussian noise. This work learns using the x0 prediction formulation. This author proposes TACoS to build the reverse process. Experiments provide various types of forward processes.
Strength:
This paper demonstrates that it is possible to use forward process ... | Recommendation: 5: marginally below the acceptance threshold | Area: Generative models | Review:
This work proposes to use forward processes other than the Gaussian noise. This work learns using the x0 prediction formulation. This author proposes TACoS to build the reverse process. Experiments provide various types of forward processes.
Strength:
This paper demonstrates that it is possible to use forward ... |
This paper proposes a new way for molecule docking, which is to use diffusion model to form the docking pose prediction as a generation problem. In this work, the prediction is mainly based on the 3D translation group, rotation group and torsion angle change. The authors first critique the regression-based ligand propo... | 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 new way for molecule docking, which is to use diffusion model to form the docking pose prediction as a generation problem. In this work, the prediction is mainly based on the 3D translation group, rotation group and torsion angle change. The authors first critique the regression-based liga... |
The authors proposed an algorithm for interpreting deep models, which is an extension of Fel et al., 2021. The main tool it relies on is called Sobol first-order sensitivity. The experiments are demonstrated on different CNN variants, including VGG, ResNet and DenseNet.
- The main criticism is on the novelty. Most of ... | Recommendation: 3: reject, not good enough | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The authors proposed an algorithm for interpreting deep models, which is an extension of Fel et al., 2021. The main tool it relies on is called Sobol first-order sensitivity. The experiments are demonstrated on different CNN variants, including VGG, ResNet and DenseNet.
- The main criticism is on the novelty. ... |
This paper is about adversarial training for robust accuracy against adversarial attacks. The authors introduce a new procedure to improve the generalization performance of adversarially-trained networks by using self-supervised test-time fine-tuning. In addition, to determine a good starting point for test-time adaptat... | Recommendation: 6: marginally above the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper is about adversarial training for robust accuracy against adversarial attacks. The authors introduce a new procedure to improve the generalization performance of adversarially-trained networks by using self-supervised test-time fine-tuning. In addition, to determine a good starting point for test-time... |
The paper addresses overestimation issues in multi-agent RL, and shows that the issues not only come from target Q function for each individual agent, but also has an accumulated overestimation effect in the online global Q-network. To address his problem, the paper proposes Dual Ensembled MUltiagent Q-learning with hy... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper addresses overestimation issues in multi-agent RL, and shows that the issues not only come from target Q function for each individual agent, but also has an accumulated overestimation effect in the online global Q-network. To address his problem, the paper proposes Dual Ensembled MUltiagent Q-learning... |
The paper studies the effectiveness of self-pretraining, meaning applying pretraining *objectives* like masked language modeling to *datasets* specific for a single task. Remarkably, self-pretraining often achieves performance comparable to the conventional approach of pre-training on generalist web-scale text. This ca... | Recommendation: 8: accept, good paper | Area: Unsupervised and Self-supervised learning | Review:
The paper studies the effectiveness of self-pretraining, meaning applying pretraining *objectives* like masked language modeling to *datasets* specific for a single task. Remarkably, self-pretraining often achieves performance comparable to the conventional approach of pre-training on generalist web-scale text.... |
In this paper, the authors introduce a new version of the Concept Bottleneck Model (Koh et al., 2020) that dims to be more accurate and interpretable w.r.t. other Concept-based models, especially in regimes where concept supervision is limited.
The proposed method is inspired by a theorem depicting the trade-off betwe... | 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 introduce a new version of the Concept Bottleneck Model (Koh et al., 2020) that dims to be more accurate and interpretable w.r.t. other Concept-based models, especially in regimes where concept supervision is limited.
The proposed method is inspired by a theorem depicting the trade-o... |
In this paper, to solve the expensive computing resources problem of current dynamic mixup method, it proposed a method trying to transfer the decoupling mechanism of dynamic methods from the data level to the objective function level and propose the general decoupled mixup loss. The experimental results on supervise... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
In this paper, to solve the expensive computing resources problem of current dynamic mixup method, it proposed a method trying to transfer the decoupling mechanism of dynamic methods from the data level to the objective function level and propose the general decoupled mixup loss. The experimental results on s... |
This paper focuses on the problem of difficulty amplification: trained models exhibit consistent differences in performance on groups that are "easy" and "difficult" even when there are no obvious spurious correlations or dataset imbalance. In particular, they show that the simplicity bias is a major reason for diffic... | 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 problem of difficulty amplification: trained models exhibit consistent differences in performance on groups that are "easy" and "difficult" even when there are no obvious spurious correlations or dataset imbalance. In particular, they show that the simplicity bias is a major reason fo... |
This work tries to improve certified training based on box propagation (DiffAI) / interval bound propagation (IBP). The proposed method tries to find a point with a high loss within the perturbation region by adversarial attack. Then box propagation is performed around the high-loss input point with a small box only, i... | Recommendation: 6: marginally above the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This work tries to improve certified training based on box propagation (DiffAI) / interval bound propagation (IBP). The proposed method tries to find a point with a high loss within the perturbation region by adversarial attack. Then box propagation is performed around the high-loss input point with a small box... |
In this paper, the authors proposed ahead-of-time (AoT) P-Tuning, a new prompt-tuning method that adds an input-dependent offset to each layer's activation (based on the vocabulary index). The proposed method does not increase the sequence length, leading to now significant computation increase. It achieves competitive... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
In this paper, the authors proposed ahead-of-time (AoT) P-Tuning, a new prompt-tuning method that adds an input-dependent offset to each layer's activation (based on the vocabulary index). The proposed method does not increase the sequence length, leading to now significant computation increase. It achieves com... |
Data augmentations with common input transformations is a commonplace in ML training, and is known to improve performance. This work argues that example-agnostic common transformations are too restrictive and that example specific augmentations can allow much wider transformation. The proposed methods of learning insta... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
Data augmentations with common input transformations is a commonplace in ML training, and is known to improve performance. This work argues that example-agnostic common transformations are too restrictive and that example specific augmentations can allow much wider transformation. The proposed methods of learni... |
This paper studies quantitative separations between two families of antisymmetric functions: The Slater Ansatz which is a linear combination of Slater determinants and the more powerful Jastrow Ansatz where each Slater determinant is augmented by a symmetric prefactor (and hence remains antisymmetric). In this area of ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper studies quantitative separations between two families of antisymmetric functions: The Slater Ansatz which is a linear combination of Slater determinants and the more powerful Jastrow Ansatz where each Slater determinant is augmented by a symmetric prefactor (and hence remains antisymmetric). In this ... |
This paper considers the problem of domain generalization where the domain index is across time. The authors proposed a Bayesian framework to explicitly model the concept drift over time, i.e., predicting the updated model on the future domain based on historical domains. A recurrent neural network is employed to learn... | Recommendation: 8: accept, good paper | Area: General Machine Learning | Review:
This paper considers the problem of domain generalization where the domain index is across time. The authors proposed a Bayesian framework to explicitly model the concept drift over time, i.e., predicting the updated model on the future domain based on historical domains. A recurrent neural network is employed ... |
The author proposed a framework that treats model fitting with missing values as a latent space regularization problem using measure theoretical arguments. Specifically, the author translates the effect of missing values in the data space into the decreased mutual information between the latent variables and target var... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The author proposed a framework that treats model fitting with missing values as a latent space regularization problem using measure theoretical arguments. Specifically, the author translates the effect of missing values in the data space into the decreased mutual information between the latent variables and ta... |
A new model called NS-SSCL, neuro-symbolic self-supervised contrastive learning, is proposed for solving abstract reasoning problems in RAVEN and V-PROM, two abstract reasoning problems used for evaluating machine intelligence. The NS-SSCL method leverages SSCL to learn visual representation for attribute disentangleme... | Recommendation: 3: reject, not good enough | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
A new model called NS-SSCL, neuro-symbolic self-supervised contrastive learning, is proposed for solving abstract reasoning problems in RAVEN and V-PROM, two abstract reasoning problems used for evaluating machine intelligence. The NS-SSCL method leverages SSCL to learn visual representation for attribute disen... |
This paper proposes two measures that encapsulate the topological features of data: Persistent Intrinsic Dimension (PID) and Euclidicity. The authors claim that manifold learning with a fixed intrinsic dimension can be a restrictive assumption in general and it is worth thinking of allowing this dimension to vary for d... | Recommendation: 3: reject, not good enough | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper proposes two measures that encapsulate the topological features of data: Persistent Intrinsic Dimension (PID) and Euclidicity. The authors claim that manifold learning with a fixed intrinsic dimension can be a restrictive assumption in general and it is worth thinking of allowing this dimension to va... |
This paper studies graph-regularized MLPs performance limitations. They show that the node embeddings space from a conventional GR-MLP suffers from dimensional collapse (or spectral collapse). Their solution, ORTHO-REG, mitigates this issue by introducing a soft regularization term based on the correlation matrix of no... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper studies graph-regularized MLPs performance limitations. They show that the node embeddings space from a conventional GR-MLP suffers from dimensional collapse (or spectral collapse). Their solution, ORTHO-REG, mitigates this issue by introducing a soft regularization term based on the correlation matr... |
This introduces a hierarchical framework for multi-agent reinforcement learning. It formulates agent-task assignments as a linear programming problem. With the solution of LP, it generates a low-level policy to solve each sub-task in a cooperative manner. Some empirical experiments are performed to demonstrate the effe... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This introduces a hierarchical framework for multi-agent reinforcement learning. It formulates agent-task assignments as a linear programming problem. With the solution of LP, it generates a low-level policy to solve each sub-task in a cooperative manner. Some empirical experiments are performed to demonstrate ... |
This paper considers solving meta-learning (formulated as bilevel optimization with the inner level being tasks) using techniques from multi-objective optimization (MOO). Previous work along this line required computing gradients for each task at each iteration, which is expensive. This paper proposes an algorithm to f... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper considers solving meta-learning (formulated as bilevel optimization with the inner level being tasks) using techniques from multi-objective optimization (MOO). Previous work along this line required computing gradients for each task at each iteration, which is expensive. This paper proposes an algori... |
The paper studies approximation power of (residual) vanilla convolutional networks. The authors show that any translational equivalent/invariant mapping/functions can approximated by such networks, in the $L^p$ sense, if
- the number of channels is at least 2.
- the kernel size is at least 2 (in each coordinates)
- t... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The paper studies approximation power of (residual) vanilla convolutional networks. The authors show that any translational equivalent/invariant mapping/functions can approximated by such networks, in the $L^p$ sense, if
- the number of channels is at least 2.
- the kernel size is at least 2 (in each coordina... |
This paper suggests the use of post-training quantization to quantize neural image compression models, where the model parameters are compressed from Float32 format to Int8 precision. The weight, bias and activation parameters are processed with post-training optimization. Some theoretical justifications are provided t... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper suggests the use of post-training quantization to quantize neural image compression models, where the model parameters are compressed from Float32 format to Int8 precision. The weight, bias and activation parameters are processed with post-training optimization. Some theoretical justifications are pr... |
This paper considers the estimation of latent causal models, where low-level data like image pixels or high-dimensional vectors are observed, but not the underlying causal variables. Different from previous methods, the authors handle this problem in a Bayesian manner, by putting a prior on the latent causal structure ... | Recommendation: 3: reject, not good enough | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
This paper considers the estimation of latent causal models, where low-level data like image pixels or high-dimensional vectors are observed, but not the underlying causal variables. Different from previous methods, the authors handle this problem in a Bayesian manner, by putting a prior on the latent causal st... |
This paper tackles the issue of temporal drift. The proposed methodology consists of a recurrent neural network modelling the parameters of the classification/regression model, assuming constant topology of this later. This recurrent model aims to encapsulate the latent change in parameters through time. The parameters... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
This paper tackles the issue of temporal drift. The proposed methodology consists of a recurrent neural network modelling the parameters of the classification/regression model, assuming constant topology of this later. This recurrent model aims to encapsulate the latent change in parameters through time. The pa... |
The paper proposes a mathematical formalism for model fusion motivated by the TLP distance which is interesting. They illustrate the performance of their method by fusing various networks on both homogenous and heterogenous task settings, where they seem to perform slightly better than the OTFusion method upon which th... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The paper proposes a mathematical formalism for model fusion motivated by the TLP distance which is interesting. They illustrate the performance of their method by fusing various networks on both homogenous and heterogenous task settings, where they seem to perform slightly better than the OTFusion method upon ... |
This paper deals with generative modelling of videos, and specifically proposes a variety of improvements over StyleGAN-V. The improvements are intellectually appealing, and quite convincingly demonstrated. While Arxiv is starting to be full of video papers, many of which exceed the quality of this paper, they may not ... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
This paper deals with generative modelling of videos, and specifically proposes a variety of improvements over StyleGAN-V. The improvements are intellectually appealing, and quite convincingly demonstrated. While Arxiv is starting to be full of video papers, many of which exceed the quality of this paper, they ... |
In this paper, the authors introduce a graphical referential game to study the emergence of visual conventions between communicating agents. The framework is based on recent computational approaches exploring the emergence of language through language games. In particular, the authors exploit a “Listener / Speaker” arc... | Recommendation: 3: reject, not good enough | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
In this paper, the authors introduce a graphical referential game to study the emergence of visual conventions between communicating agents. The framework is based on recent computational approaches exploring the emergence of language through language games. In particular, the authors exploit a “Listener / Spea... |
As the title suggests, this paper explores domain generalization for small data through a probabilistic maximum mean discrepancy (MMD) approach. For this purpose a domain-invariant representation is being learned from multiple domains (sources) where the assumption is that these domains inherently have insufficient sam... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
As the title suggests, this paper explores domain generalization for small data through a probabilistic maximum mean discrepancy (MMD) approach. For this purpose a domain-invariant representation is being learned from multiple domains (sources) where the assumption is that these domains inherently have insuffic... |
This paper investigates potential privacy risk in federated learning. They found that existing privacy defenses in FL can be broken via a simple adaptive attack. In particular, the proposed learning-based approach (Learning to invert) aims to train an inversion model to reconstruct training samples from their gradient ... | 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 investigates potential privacy risk in federated learning. They found that existing privacy defenses in FL can be broken via a simple adaptive attack. In particular, the proposed learning-based approach (Learning to invert) aims to train an inversion model to reconstruct training samples from their g... |
This paper systematically looks for the best way to combine parameter-efficient fine-tuning methods. Authors consider 4 design decisions: (1) how to group layers together (that layers within one group are treated equally), (2) how to allocate the "budget" of trainable parameters between groups, (3) whether to tune each... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper systematically looks for the best way to combine parameter-efficient fine-tuning methods. Authors consider 4 design decisions: (1) how to group layers together (that layers within one group are treated equally), (2) how to allocate the "budget" of trainable parameters between groups, (3) whether to t... |
The authors propose a method which accelerates the architecture choice step of certain one-shot NAS methods. The method greedily freezes the first block choices during the architecture optimization phase with an evolutionary algorithm. This allows for caching the intermediate feature maps and therefore reduce the compu... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
The authors propose a method which accelerates the architecture choice step of certain one-shot NAS methods. The method greedily freezes the first block choices during the architecture optimization phase with an evolutionary algorithm. This allows for caching the intermediate feature maps and therefore reduce t... |
The paper proposes a framework for the adaptive discretization of continuous-action spaces in an online manner, by employing a two-phase learning structure: (1) latent action space learning; (2) RL on the learned action space.
**Strengths:**
- The target problem is very important: To have a unified approach for deali... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper proposes a framework for the adaptive discretization of continuous-action spaces in an online manner, by employing a two-phase learning structure: (1) latent action space learning; (2) RL on the learned action space.
**Strengths:**
- The target problem is very important: To have a unified approach f... |
This paper proposed to learn a multi-hop KBQA model which contains a retrieval and reasoning model that shares the same architecture. Unifying the retrieval and reasoning module let the model share more learned knowledge. Experiments show great performance on three benchmark multi-hop reasoning datasets.
The propose mo... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposed to learn a multi-hop KBQA model which contains a retrieval and reasoning model that shares the same architecture. Unifying the retrieval and reasoning module let the model share more learned knowledge. Experiments show great performance on three benchmark multi-hop reasoning datasets.
The pr... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.