Subset (19)

iclr23 · 14k rows

Split (3)

train · 8.4k rows

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.

review stringlengths 5 16.6k	score stringclasses 5 values	area stringclasses 12 values	text stringlengths 31 5.65k
Under certain mixture of Gaussian assumption of the data distribution, the paper provides an information-theoretic justification of the VICReg method for self-supervised learning. Additionally, a generalization bound for the down-stream tasks is established. Strength The main strength of the paper is its justifi...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: Under certain mixture of Gaussian assumption of the data distribution, the paper provides an information-theoretic justification of the VICReg method for self-supervised learning. Additionally, a generalization bound for the down-stream tasks is established. Strength The main strength of the paper is its...
This paper tackles the problem of 3D-aware human generation from 2D images. To generate at a high resolution, EVA3D proposes to use compositional multiple NeRFs to do the generation. To overcome difficulties in training with imbalanced human dataset, the paper proposes several training strategies. Results on several h...	Recommendation: 6: marginally above the acceptance threshold	Area: Generative models	Review: This paper tackles the problem of 3D-aware human generation from 2D images. To generate at a high resolution, EVA3D proposes to use compositional multiple NeRFs to do the generation. To overcome difficulties in training with imbalanced human dataset, the paper proposes several training strategies. Results on s...
This paper is proposing a modified neural ordinary differential equation (NODE) that augments the function that describes the vector field with an additional term that introduces oscillations into the dynamics. The authors show that this modification can overcome limitations of standard NODEs by (i) proofing that the n...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper is proposing a modified neural ordinary differential equation (NODE) that augments the function that describes the vector field with an additional term that introduces oscillations into the dynamics. The authors show that this modification can overcome limitations of standard NODEs by (i) proofing th...
Main contributions: 1) Identify implicature understanding as a useful benchmark of LLM's communication ability. 2) Design a protocol for comparing implicature understanding of LLMs versus that of humans. 3) Roll-out the protocol on a host of existing LLMs. Providing a deeper analysis of (a) how different properties of ...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: Main contributions: 1) Identify implicature understanding as a useful benchmark of LLM's communication ability. 2) Design a protocol for comparing implicature understanding of LLMs versus that of humans. 3) Roll-out the protocol on a host of existing LLMs. Providing a deeper analysis of (a) how different proper...
The paper considers the problem of channel equalization for wired-line communications systems, and proposes a machine learning based equalizer inspired by forward-backward algorithm. The proposed approach itself might have some novelty, however, I cannot recommend the acceptance of the paper, mainly due to the lack of...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper considers the problem of channel equalization for wired-line communications systems, and proposes a machine learning based equalizer inspired by forward-backward algorithm. The proposed approach itself might have some novelty, however, I cannot recommend the acceptance of the paper, mainly due to the...
This paper deals with the problem of prompt-style finetuning when gradient information is not available. The solution is using adaboost to quickly learn prompt-style learners. Experimental results on standard benchmark showed competitive results over previous black-box methods. - Strength 1. Clearly written, well moti...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper deals with the problem of prompt-style finetuning when gradient information is not available. The solution is using adaboost to quickly learn prompt-style learners. Experimental results on standard benchmark showed competitive results over previous black-box methods. - Strength 1. Clearly written, w...
This paper focused on adapting large-scaled pre-trained vision-language model (i.e., CLIP) to downstream datasets under the few-shot setting. Compared to the recent related work CoOp, this work (1) learns multiple prompts, (2) uses both local features and global features, and (3) measures the similarity of prompts and ...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper focused on adapting large-scaled pre-trained vision-language model (i.e., CLIP) to downstream datasets under the few-shot setting. Compared to the recent related work CoOp, this work (1) learns multiple prompts, (2) uses both local features and global features, and (3) measures the similarity of prom...
The goal of this study is to attack the semi-supervised learning problem for graph-structured data. The paper focus on bridging the gap between GNNs and the limiting GPs by deriving the covariance kernel that incorporates the graph inductive bias as GNNs do. More specifically, the paper has shown that the GP can be de...	Recommendation: 5: marginally below the acceptance threshold	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: The goal of this study is to attack the semi-supervised learning problem for graph-structured data. The paper focus on bridging the gap between GNNs and the limiting GPs by deriving the covariance kernel that incorporates the graph inductive bias as GNNs do. More specifically, the paper has shown that the GP c...
The paper proposed a method to detect malicious nodes by converting the numerical features to rank features and then apply malicious nodes detection algorithm. Strength: 1. The idea of converting numerical features to rank features is inspiring. Will combine them both further improve the detection accuracy? Weakness: ...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: The paper proposed a method to detect malicious nodes by converting the numerical features to rank features and then apply malicious nodes detection algorithm. Strength: 1. The idea of converting numerical features to rank features is inspiring. Will combine them both further improve the detection accuracy? We...
In the proposed work, the authors present E3bind, an fully-equivariant and differentiable approach for protein-ligand docking. Similar works have recently become available in the literature, the claimed advantage of the presented one being that it iteratively docks the ligand into a (possibly unknown) protein binding p...	Recommendation: 5: marginally below the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: In the proposed work, the authors present E3bind, an fully-equivariant and differentiable approach for protein-ligand docking. Similar works have recently become available in the literature, the claimed advantage of the presented one being that it iteratively docks the ligand into a (possibly unknown) protein b...
The paper presents a variance reduction technique that improves the Straight-Through version of the Gumbel-Softmax estimator by taking the zero-temperature limit. The paper further shows that the proposed estimator could be decomposed into a sum of straight-through estimator and DARN. The superior performance of the pr...	Recommendation: 8: accept, good paper	Area: Generative models	Review: The paper presents a variance reduction technique that improves the Straight-Through version of the Gumbel-Softmax estimator by taking the zero-temperature limit. The paper further shows that the proposed estimator could be decomposed into a sum of straight-through estimator and DARN. The superior performance o...
The authors proposed a method that maps a single 2D image of a scene to a persistent 3D scene representation, enabling novel view synthesis and disentangled representation of the movable and immovable components of the scene. The enabler is a new representation called conditional neural groudplans, motivated by the Bir...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: The authors proposed a method that maps a single 2D image of a scene to a persistent 3D scene representation, enabling novel view synthesis and disentangled representation of the movable and immovable components of the scene. The enabler is a new representation called conditional neural groudplans, motivated by...
The paper is considering the problem of learning robotics manipulation in a multi-object setting. The specific problem considered is the successive repositioning of two cubes by a manipulator in the presence of a variable number of distractor cubes. The authors argue that previous approaches, based on GNNs do not lea...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper is considering the problem of learning robotics manipulation in a multi-object setting. The specific problem considered is the successive repositioning of two cubes by a manipulator in the presence of a variable number of distractor cubes. The authors argue that previous approaches, based on GNNs do...
This paper proposes a method for creating a concept bottleneck model from any given model. The authors claim that the proposed approach does not sacrifice the performance of the model, and retains the interpretability benefits of concept bottleneck models along with easy model editing. Strengths: It is an extremely we...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper proposes a method for creating a concept bottleneck model from any given model. The authors claim that the proposed approach does not sacrifice the performance of the model, and retains the interpretability benefits of concept bottleneck models along with easy model editing. Strengths: It is an extr...
The paper proposes to improve the stability, accuracy, and efficiency of implicit graph neural networks (IGNN) by new parameterizations (i.e., the Cayley transform-based orthogonal parameterization and monotone parameterization) and advanced solvers (i.e., operator splitting, Anderson acceleration). Theoretical justifi...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposes to improve the stability, accuracy, and efficiency of implicit graph neural networks (IGNN) by new parameterizations (i.e., the Cayley transform-based orthogonal parameterization and monotone parameterization) and advanced solvers (i.e., operator splitting, Anderson acceleration). Theoretical...
This paper presents a new method of hierarchical entity typing that represents mentions and types using a hyper-rectangle representation, models the mention-type relation using the intersection of these representations, and classifies per-mention types according to the mass of this intersection. The mass representation...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper presents a new method of hierarchical entity typing that represents mentions and types using a hyper-rectangle representation, models the mention-type relation using the intersection of these representations, and classifies per-mention types according to the mass of this intersection. The mass repres...
This paper proposes a method to build neural networks where an output is provably monotonic with respect to certain inputs. This is intended for applications where domain knowledge exists for such monotonic relations. For such applications, the benefits are better robustness and better interpretability. This paper cont...	Recommendation: 3: reject, not good enough	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper proposes a method to build neural networks where an output is provably monotonic with respect to certain inputs. This is intended for applications where domain knowledge exists for such monotonic relations. For such applications, the benefits are better robustness and better interpretability. This pa...
In this work, the authors propose a new way to learn positional encoding in Graph Transformers by extracting relationships between distant nodes in the graph. To evaluate the proposed approach, the authors design a simple graph task, Grid Histogram Counting, as well as use several benchmark datasets. Pros: The experime...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: In this work, the authors propose a new way to learn positional encoding in Graph Transformers by extracting relationships between distant nodes in the graph. To evaluate the proposed approach, the authors design a simple graph task, Grid Histogram Counting, as well as use several benchmark datasets. Pros: The ...
This paper proposed a new way of in-context prompting of language model, by combining chain-of-thought reasoning and action decision generation to guid the model generation. The model performance are evaluated on several language and text-based games. Strength: The proposed approach is well justified. The performanc...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposed a new way of in-context prompting of language model, by combining chain-of-thought reasoning and action decision generation to guid the model generation. The model performance are evaluated on several language and text-based games. Strength: The proposed approach is well justified. The pe...
This paper explores encoding high-resolution weather data using implicit neural representation (INR). In INR, a neural network takes as input a coordinate (here, transformed latitute, longitude, time and pressure) and output the compressed values (weather data here) at that coordinates. The approach is compared to seve...	Recommendation: 8: accept, good paper	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper explores encoding high-resolution weather data using implicit neural representation (INR). In INR, a neural network takes as input a coordinate (here, transformed latitute, longitude, time and pressure) and output the compressed values (weather data here) at that coordinates. The approach is compared...
The problem of solving computer vision tasks with neural networks on large images is considered. Authors propose a method that splits an image into patches and then applying greedy iterative approach (IPS iterative patch selection) to select the representative ones. Selection is performed by creating a buffer and itera...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The problem of solving computer vision tasks with neural networks on large images is considered. Authors propose a method that splits an image into patches and then applying greedy iterative approach (IPS iterative patch selection) to select the representative ones. Selection is performed by creating a buffer a...
The paper proposed a new approach composed of multiple components for systematic generalization on language tasks. It achieves good results on multiple datasets including SCAN, PCFG, and a task for arithmetic reasoning. Strengths: * Domain: the paper explores an important and growing domain of systematic gene...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: The paper proposed a new approach composed of multiple components for systematic generalization on language tasks. It achieves good results on multiple datasets including SCAN, PCFG, and a task for arithmetic reasoning. Strengths: * Domain: the paper explores an important and growing domain of systema...
The main idea in this paper is to do adversarial imitation learning (AIL) with a discriminator trained via the infoNCE loss, rather than the binary cross-entropy loss. The paper argues that this choice of loss function will result in "smoother" and more "semantically meaningful" reward signals for training the policy. ...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The main idea in this paper is to do adversarial imitation learning (AIL) with a discriminator trained via the infoNCE loss, rather than the binary cross-entropy loss. The paper argues that this choice of loss function will result in "smoother" and more "semantically meaningful" reward signals for training the ...
Spatial Resolved Temporal Networks (SpaRTeN) are described in this paper as a novel, composite deep learning model for online, unsupervised representation learning in a spatially constrained latent space. In the proposed approach, two distinct components are simultaneously learned: a recurrent neural network ensemble t...	Recommendation: 5: marginally below the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: Spatial Resolved Temporal Networks (SpaRTeN) are described in this paper as a novel, composite deep learning model for online, unsupervised representation learning in a spatially constrained latent space. In the proposed approach, two distinct components are simultaneously learned: a recurrent neural network en...
The paper leverage the variant of the OPE estimator to estimate the average treatment effects in non-stationary dynamics. The problem is interesting and challenging, however, the method proposed in the paper might be limited for practical use. Strength: The paper proposes a treatment effect estimator in non-stationar...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper leverage the variant of the OPE estimator to estimate the average treatment effects in non-stationary dynamics. The problem is interesting and challenging, however, the method proposed in the paper might be limited for practical use. Strength: The paper proposes a treatment effect estimator in non-s...
This paper under review uses heat diffusion process to understand some multi-scale behavior of the learned model around a test point. Summary scales which characterizes the model on different scales can be drawn. Strength: This paper is clearly written and well-organized. Weakness: The paper introduces many math nota...	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 under review uses heat diffusion process to understand some multi-scale behavior of the learned model around a test point. Summary scales which characterizes the model on different scales can be drawn. Strength: This paper is clearly written and well-organized. Weakness: The paper introduces many m...
This present a novel neural-networks-based algorithm to compute optimal transport (OT) plans and maps for general cost functionals. The algorithm generalizes prior OT methods for weak and strong cost functionals. And it constructs a functional to map data distributions with preserving the class-wise structure of data...	Recommendation: 3: reject, not good enough	Area: Generative models	Review: This present a novel neural-networks-based algorithm to compute optimal transport (OT) plans and maps for general cost functionals. The algorithm generalizes prior OT methods for weak and strong cost functionals. And it constructs a functional to map data distributions with preserving the class-wise structure...
The paper introduces MiDAS, a novel method to address the problem of out-of-domain texts for fake news classification. MiDAS applies an adaptive model selector that utilizes the Lipschitz smoothness idea to estimate the model's relevancy. The author shows supremacy over several baselines and previously introduced metho...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper introduces MiDAS, a novel method to address the problem of out-of-domain texts for fake news classification. MiDAS applies an adaptive model selector that utilizes the Lipschitz smoothness idea to estimate the model's relevancy. The author shows supremacy over several baselines and previously introduc...
This paper proposes a discrete diffusion model for graph generation by combining discrete diffusion [Austin 2021] and graph transformer [Dwivedi and Bresson 2021]. Its main novelties are: - The diffusion model is discrete: sampling of node and edge features is performed after adding noise. - The denoising network is ...	Recommendation: 8: accept, good paper	Area: Generative models	Review: This paper proposes a discrete diffusion model for graph generation by combining discrete diffusion [Austin 2021] and graph transformer [Dwivedi and Bresson 2021]. Its main novelties are: - The diffusion model is discrete: sampling of node and edge features is performed after adding noise. - The denoising net...
The paper develops a method for using multiple environments during training to obtain a model that performs well on an unseen test environment. The method has both theoretical and experimental support. Strengths: The method developed in the paper is based on an interesting way to model different environments, and the ...	Recommendation: 6: marginally above the acceptance threshold	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: The paper develops a method for using multiple environments during training to obtain a model that performs well on an unseen test environment. The method has both theoretical and experimental support. Strengths: The method developed in the paper is based on an interesting way to model different environments, ...
This work finds it unclear to determine which surrogate gradient to apply in different tasks or networks in SNN. Seeking to solve this problem, the authors have done some experiments with different surrogate gradients across tasks and networks, and come to the conclusion that the derivative of fast sigmoid outperforms ...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This work finds it unclear to determine which surrogate gradient to apply in different tasks or networks in SNN. Seeking to solve this problem, the authors have done some experiments with different surrogate gradients across tasks and networks, and come to the conclusion that the derivative of fast sigmoid outp...
The paper proposes a text-driven image editing technique to refine images generated from large-scale text-to-image diffusion models. The paper relies on a crucial observation: Just modifying a text prompt when calling the generator results in structurally and compositionally very different images, even if the text prom...	Recommendation: 8: accept, good paper	Area: Generative models	Review: The paper proposes a text-driven image editing technique to refine images generated from large-scale text-to-image diffusion models. The paper relies on a crucial observation: Just modifying a text prompt when calling the generator results in structurally and compositionally very different images, even if the t...
This papers studies the generalization properties of $L_q$-stable algorithms. This notion of stability is distribution-dependent, and thus less stringent analogue of uniform stability. The authors first derive a general inequality for the sum of functions of random variables with bounded difference (Theorem 1), and use...	Recommendation: 8: accept, good paper	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This papers studies the generalization properties of $L_q$-stable algorithms. This notion of stability is distribution-dependent, and thus less stringent analogue of uniform stability. The authors first derive a general inequality for the sum of functions of random variables with bounded difference (Theorem 1),...
This work develops a pipeline for unsupervised semantic segmentation. The distinguishing feature of this approach is its generalization from training on binary classification examples to multiple classes. This is achieved by clustering feature embeddings extracted with the help of unsupervised saliency detection method...	Recommendation: 8: accept, good paper	Area: Unsupervised and Self-supervised learning	Review: This work develops a pipeline for unsupervised semantic segmentation. The distinguishing feature of this approach is its generalization from training on binary classification examples to multiple classes. This is achieved by clustering feature embeddings extracted with the help of unsupervised saliency detectio...
The authors propose an approach for visual exploration using RGB cameras only. The approach is also able to build a topological map of the environment. To do the exploration, the authors use imitation learning to hypothesize the image feature in the next step and choose an action based on that feature (and current feat...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The authors propose an approach for visual exploration using RGB cameras only. The approach is also able to build a topological map of the environment. To do the exploration, the authors use imitation learning to hypothesize the image feature in the next step and choose an action based on that feature (and curr...
The authors study the problem of learning classes of deep, feedfoward neural networks with ReLU activations through with guarantees in terms of $L^\infty$ accuracy, through the angle of sample complexity requirements. More precisely, they consider a rather general model where a learner can adaptively query points of th...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The authors study the problem of learning classes of deep, feedfoward neural networks with ReLU activations through with guarantees in terms of $L^\infty$ accuracy, through the angle of sample complexity requirements. More precisely, they consider a rather general model where a learner can adaptively query poin...
FedLite is a framework for split FL that addresses on of the main problems this variant of FL has: high communication costs that result from sending the activations generated by the last layer (the split layer) that runs on the client side to the server AND downloading the gradients from the server which are needed to ...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: FedLite is a framework for split FL that addresses on of the main problems this variant of FL has: high communication costs that result from sending the activations generated by the last layer (the split layer) that runs on the client side to the server AND downloading the gradients from the server which are ne...
The paper uses the recursive vector quantization ANN method, which can also be seen as a hierarchical VQ approach. The proposed method provides an optimization-based approach to fine-tune this ANN method. The analysis was made on million-scale and billion-scale ANN datasets. Strengths: 1) The paper is novel and shows...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper uses the recursive vector quantization ANN method, which can also be seen as a hierarchical VQ approach. The proposed method provides an optimization-based approach to fine-tune this ANN method. The analysis was made on million-scale and billion-scale ANN datasets. Strengths: 1) The paper is novel a...
The manuscript proposes to to use transformer layers applied to 3D voxel tokens at different levels of detail in order to improve the reconstruction quality of monocular 3D reconstruction (given frame poses) over related work that typically relies on 3D CNNs for multi-view aggregation and fusion. The core contribution...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The manuscript proposes to to use transformer layers applied to 3D voxel tokens at different levels of detail in order to improve the reconstruction quality of monocular 3D reconstruction (given frame poses) over related work that typically relies on 3D CNNs for multi-view aggregation and fusion. The core cont...
This paper focusses on the setting of goal-conditioned RL where subgoals are provided by graph based planners. It introduces two algorithmic contributions: * a self imitation loss that is inspired by the "optimal substructure property" ie the fact that parts of an optimal path are themselves optimal. This idea is used ...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: This paper focusses on the setting of goal-conditioned RL where subgoals are provided by graph based planners. It introduces two algorithmic contributions: * a self imitation loss that is inspired by the "optimal substructure property" ie the fact that parts of an optimal path are themselves optimal. This idea ...