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
This paper puts forward a method of mutual information estimation within neural networks where the stochasticity of observations is ensured by Gaussian dropout. The authors then use MC methods to estimate entropies and conditional entropies. The proposed approach is compared to MI estimation based on binning and the au...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper puts forward a method of mutual information estimation within neural networks where the stochasticity of observations is ensured by Gaussian dropout. The authors then use MC methods to estimate entropies and conditional entropies. The proposed approach is compared to MI estimation based on binning an...
This paper theoretically studies the generalization of a multi-layer neural network containing a self-attention module (resembling a simplified vision transformer), trained with SGD on a dataset consists of a mixture of label-relevant and irrelevant tokens. This paper is the first to formally study the generalization o...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This paper theoretically studies the generalization of a multi-layer neural network containing a self-attention module (resembling a simplified vision transformer), trained with SGD on a dataset consists of a mixture of label-relevant and irrelevant tokens. This paper is the first to formally study the generali...
The authors propose a technique for improving distributionally robust techniques for individual fairness by generating antidote samples. These adversarial samples are constrained to be on the manifold of the data distribution. The antidote data generator is a generative adversarial network (GAN) model that takes sample...	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 authors propose a technique for improving distributionally robust techniques for individual fairness by generating antidote samples. These adversarial samples are constrained to be on the manifold of the data distribution. The antidote data generator is a generative adversarial network (GAN) model that take...
This paper focuses on the ability of large language models (LLMs) and humans to handle ambiguity in task prompts, of the type that are currently popular in the LLM literature. A small suite of six synthetic tasks are created. Each asks the model (or human) to identify whether a sentence contains a particular category ...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper focuses on the ability of large language models (LLMs) and humans to handle ambiguity in task prompts, of the type that are currently popular in the LLM literature. A small suite of six synthetic tasks are created. Each asks the model (or human) to identify whether a sentence contains a particular c...
This paper looks at a new method for adding gaussian noise to nodes during training such that the amount of noise added is a learned parameter. The paper, the does some theoretical analysis, to obtain some heuristics about the effect of the noise. This followed up by empirical work looking at the evolution of the norms...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper looks at a new method for adding gaussian noise to nodes during training such that the amount of noise added is a learned parameter. The paper, the does some theoretical analysis, to obtain some heuristics about the effect of the noise. This followed up by empirical work looking at the evolution of t...
This paper proposes a new secure aggregation protocol P2PRISM for decentralized peer-to-peer learning, against malicious adversaries, in various network topologies. Strength: This paper considers a fairly interesting setting where no centralized trusted server is presented to aggregate the model updates as in federate...	Recommendation: 3: reject, not good enough	Area: General Machine Learning	Review: This paper proposes a new secure aggregation protocol P2PRISM for decentralized peer-to-peer learning, against malicious adversaries, in various network topologies. Strength: This paper considers a fairly interesting setting where no centralized trusted server is presented to aggregate the model updates as in ...
The paper adopts a reinforcement learning framework to estimate the long-term treatment effects in nonstationary environments. The main contributions lies in the development of a practical algorithm to estimate causal effect under nonstationarity. The algorithm is justified via both theoretical results, synthetic envir...	Recommendation: 5: marginally below the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper adopts a reinforcement learning framework to estimate the long-term treatment effects in nonstationary environments. The main contributions lies in the development of a practical algorithm to estimate causal effect under nonstationarity. The algorithm is justified via both theoretical results, synthet...
The authors propose DiffSBDD, an approach to structure-based drug design (SBDD) that aims to generate molecules in 3D that are conditioned on a protein binding interface. The approach is E(3)-equivariant and leverages the denoising diffusion probabilistic model (DDPM) formulation inspired by non-equilibrium thermodynam...	Recommendation: 3: reject, not good enough	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The authors propose DiffSBDD, an approach to structure-based drug design (SBDD) that aims to generate molecules in 3D that are conditioned on a protein binding interface. The approach is E(3)-equivariant and leverages the denoising diffusion probabilistic model (DDPM) formulation inspired by non-equilibrium the...
The authors propose a gradient-weighted visualized explanation technique for object detectors called ODAM. The technique produces instance-specific heat maps indicating the regions that have an impact on the object prediction. It combines attributions from the class and object localization prediction to produce a singl...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The authors propose a gradient-weighted visualized explanation technique for object detectors called ODAM. The technique produces instance-specific heat maps indicating the regions that have an impact on the object prediction. It combines attributions from the class and object localization prediction to produce...
This paper introduces a latent variable model for both generating and clustering neural activity. In simulated datasets, they study the performance of the model and its ability to recover underlying ensembles in a synthetic and real-world toroid task. Overall, the paper provides an interesting model and a strong set of...	Recommendation: 6: marginally above the acceptance threshold	Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)	Review: This paper introduces a latent variable model for both generating and clustering neural activity. In simulated datasets, they study the performance of the model and its ability to recover underlying ensembles in a synthetic and real-world toroid task. Overall, the paper provides an interesting model and a stron...
This paper evaluates the performance of deepfake detectors on images generated by five diffusion models (DM) and five GANs. It also analyzes the spectral representation of DM-generated images. (+) The paper is an easy-to-read paper evaluating how well images generated by DM can be detected The paper evaluates the perfo...	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 evaluates the performance of deepfake detectors on images generated by five diffusion models (DM) and five GANs. It also analyzes the spectral representation of DM-generated images. (+) The paper is an easy-to-read paper evaluating how well images generated by DM can be detected The paper evaluates t...
The paper proposes an alternative objective for Reinforcement Learning called ReMax. The authors claim that an agent optimizing this objective naturally develops an exploratory behavior, differently from standard RL, where exploration typically should be enforced, for instance adding an entropy bonus. The main idea con...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper proposes an alternative objective for Reinforcement Learning called ReMax. The authors claim that an agent optimizing this objective naturally develops an exploratory behavior, differently from standard RL, where exploration typically should be enforced, for instance adding an entropy bonus. The main ...
The paper introduces a library (open-sourced) called RL4LMs, which is for optimizing generation models using RL. The library is compatible with HuggingFace. Second, the paper comes up with a GRUE (general reinforced-language understanding evaluation) benchmark, which consists of six language generation tasks (under the...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper introduces a library (open-sourced) called RL4LMs, which is for optimizing generation models using RL. The library is compatible with HuggingFace. Second, the paper comes up with a GRUE (general reinforced-language understanding evaluation) benchmark, which consists of six language generation tasks (u...
This paper investigates the problem of building robust models by only enhancing a part of the whole model (subnetwork). The paper is mostly theoretical and propose a new concept of semi-robust. The authors also provide empirical evidences to support their claim. # Strength By only enhancing a part of the whole network,...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: This paper investigates the problem of building robust models by only enhancing a part of the whole model (subnetwork). The paper is mostly theoretical and propose a new concept of semi-robust. The authors also provide empirical evidences to support their claim. # Strength By only enhancing a part of the whole ...
The paper proposes a framework for online knowledge distillation from an intentionally diversified ensemble of teachers. A common hypothesis for the success of distillation from an ensemble is that diverse teacher models increase the efficacy of the ensemble. The paper provides a formal framework to enhance teacher div...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposes a framework for online knowledge distillation from an intentionally diversified ensemble of teachers. A common hypothesis for the success of distillation from an ensemble is that diverse teacher models increase the efficacy of the ensemble. The paper provides a formal framework to enhance tea...
The paper proposes a new framework for measuring trustworthiness of a model including new proposed metrics that can evaluate various models comparingly. The framework is applicable to text and image domains and is validated through experiments on datasets from each domain accordingly. The results showed that a model wi...	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 proposes a new framework for measuring trustworthiness of a model including new proposed metrics that can evaluate various models comparingly. The framework is applicable to text and image domains and is validated through experiments on datasets from each domain accordingly. The results showed that a ...
This work examines whether the secret information that may be contained in the training data set can be extracted from a model which satisfies RDP. This problem is positioned as a problem of intermediate difficulty between the reconstruction and membership attacks. The authors define leakage as the probability that an ...	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 examines whether the secret information that may be contained in the training data set can be extracted from a model which satisfies RDP. This problem is positioned as a problem of intermediate difficulty between the reconstruction and membership attacks. The authors define leakage as the probability ...
This paper presents a novel problem setting named adversary-aware partial label learning. The authors propose an Adversary-Aware loss function and the immature teacher within momentum (ITWM) disambiguation algorithm to tackle this problem. (+) The proposed new setting is potentially interesting. (+) The experimental r...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper presents a novel problem setting named adversary-aware partial label learning. The authors propose an Adversary-Aware loss function and the immature teacher within momentum (ITWM) disambiguation algorithm to tackle this problem. (+) The proposed new setting is potentially interesting. (+) The experi...
This paper formulates skip connection based models a learnable Markov chain. The paper then introduces the content of an efficient Markov chain which maps data from the input to the target domain efficeintly. A simple regularisation scheme is introduce to enable the learning of efficient Markov chains. Evaluation is pe...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper formulates skip connection based models a learnable Markov chain. The paper then introduces the content of an efficient Markov chain which maps data from the input to the target domain efficeintly. A simple regularisation scheme is introduce to enable the learning of efficient Markov chains. Evaluati...
The paper proposes PETALS, a system that enables inference or finetune processing of large language models (LLM) without the cost of HPC hardware, but rather by sharing a large pool of consumer-grade hardware. PETALS uses pipeline parallelism to partition LLM layers over the geographically distributed devices and propo...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper proposes PETALS, a system that enables inference or finetune processing of large language models (LLM) without the cost of HPC hardware, but rather by sharing a large pool of consumer-grade hardware. PETALS uses pipeline parallelism to partition LLM layers over the geographically distributed devices a...
The paper tackles target-aware small molecule generation. Specifically, small synthetic ligand molecules are generated using a deep generative model, such that the molecules best fit into the binding site of a larger protein. To this end, the paper jointly models the ligand and the protein using a generative diffusion ...	Recommendation: 5: marginally below the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: The paper tackles target-aware small molecule generation. Specifically, small synthetic ligand molecules are generated using a deep generative model, such that the molecules best fit into the binding site of a larger protein. To this end, the paper jointly models the ligand and the protein using a generative di...
The paper proposes an approach for boosting the effectiveness of kernelized Ridge regression by first learning a set of random features through a deep-learning inspired preprocessing step, then applying kernelized Ridge regression to the learned features. - Strength: While much work has been done on constructing a...	Recommendation: 3: reject, not good enough	Area: Unsupervised and Self-supervised learning	Review: The paper proposes an approach for boosting the effectiveness of kernelized Ridge regression by first learning a set of random features through a deep-learning inspired preprocessing step, then applying kernelized Ridge regression to the learned features. - Strength: While much work has been done on constr...
The manuscript proposes to use homotopy optimization between a stabilized version of a neural ODE and the original neural ODE to learn dynamical systems from time series data. The proposed method is tested on three synthetic problems, where it is shown to outperform a naive gradient descent training of the neural ODE. ...	Recommendation: 3: reject, not good enough	Area: Optimization (eg, convex and non-convex optimization)	Review: The manuscript proposes to use homotopy optimization between a stabilized version of a neural ODE and the original neural ODE to learn dynamical systems from time series data. The proposed method is tested on three synthetic problems, where it is shown to outperform a naive gradient descent training of the neur...
This paper explores neural image compression and focuses on improving existing methods in several ways: 1. A single model that allows variable bit rate encodings, variable encode compute, and faster run time. Although previous research has explored these items in isolation, I do not know of a model that incorporates a...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper explores neural image compression and focuses on improving existing methods in several ways: 1. A single model that allows variable bit rate encodings, variable encode compute, and faster run time. Although previous research has explored these items in isolation, I do not know of a model that incorp...
The paper analyses the solution to the multi-agent adversarial RL (MaARL) problem via mean-field optimal control viewpoint. The authors characterizes the solution using the two-sided extremism principle (TSEP) and HJI equation then provides a connection between them. At the end, the authors provide a generalization bou...	Recommendation: 3: reject, not good enough	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper analyses the solution to the multi-agent adversarial RL (MaARL) problem via mean-field optimal control viewpoint. The authors characterizes the solution using the two-sided extremism principle (TSEP) and HJI equation then provides a connection between them. At the end, the authors provide a generaliza...
The authors propose a technique to prioritize which samples to select next in an active learning setting for node attribute completion in graphs. The proposed method uses metrics related to the "density" of a data point as well as its centrality in the graph. The proposed approach is used as input to SAT, a previously-...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The authors propose a technique to prioritize which samples to select next in an active learning setting for node attribute completion in graphs. The proposed method uses metrics related to the "density" of a data point as well as its centrality in the graph. The proposed approach is used as input to SAT, a pre...
This paper studies the convergence behavior of SGD. Specifically, it says that there is a way to project the learning process onto a set of orthogonal basis functions so that the leading coefficients in the basis functions exhibit 'smoothly' convergence behavior. While not explicitly stated in the paper, I think that i...	Recommendation: 6: marginally above the acceptance threshold	Area: Optimization (eg, convex and non-convex optimization)	Review: This paper studies the convergence behavior of SGD. Specifically, it says that there is a way to project the learning process onto a set of orthogonal basis functions so that the leading coefficients in the basis functions exhibit 'smoothly' convergence behavior. While not explicitly stated in the paper, I thin...
This work proposes a new method to produce robust models by data augmentations that corrupt semantic information. Semantic corruptions power nuisance-avoiding methods to build robust models against spurious correlations without requiring extra annotations or strong assumptions. This paper analyzes semantic corruptions ...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: This work proposes a new method to produce robust models by data augmentations that corrupt semantic information. Semantic corruptions power nuisance-avoiding methods to build robust models against spurious correlations without requiring extra annotations or strong assumptions. This paper analyzes semantic corr...
This work solves personalized federated learning (PFL) setting by developing feature alignment mechanism and classifier collaboration. Concretely, this feature alignment minimizes the distance between local feature representations and global centroids. And classifier collaboration utilizes linear combination of multipl...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: This work solves personalized federated learning (PFL) setting by developing feature alignment mechanism and classifier collaboration. Concretely, this feature alignment minimizes the distance between local feature representations and global centroids. And classifier collaboration utilizes linear combination of...
The work studies the robustness of dynamic neural networks (DyNNs) from four perspectives: 1) comparison of the black-box robustness between static and DyNNs; 2) The inference time of adversarial samples in DyNNs; 3) Dynamic mechanisms with the best robustness; 4) synthesize samples to reduce the effectiveness of DyNNs...	Recommendation: 3: reject, not good enough	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: The work studies the robustness of dynamic neural networks (DyNNs) from four perspectives: 1) comparison of the black-box robustness between static and DyNNs; 2) The inference time of adversarial samples in DyNNs; 3) Dynamic mechanisms with the best robustness; 4) synthesize samples to reduce the effectiveness ...
The paper presents an algorithm to learn shortest-path (SP) representation of nodes in a graph. Existing strategies to learn SP are based on random-walks that, based on the structure of the graph, have limitations in terms of performance as well as distance preservation. The authors propose a method called Between Cent...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: The paper presents an algorithm to learn shortest-path (SP) representation of nodes in a graph. Existing strategies to learn SP are based on random-walks that, based on the structure of the graph, have limitations in terms of performance as well as distance preservation. The authors propose a method called Betw...
The paper tackles the problem of estimating the effect of continuous and multi-dimensional treatments in the setting with unobserved confounders. The paper assumes access to high dimensional proxy variables for the unobserved confounders and proposes a Contrastive regulariser which learns a proxy representation for the...	Recommendation: 5: marginally below the acceptance threshold	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: The paper tackles the problem of estimating the effect of continuous and multi-dimensional treatments in the setting with unobserved confounders. The paper assumes access to high dimensional proxy variables for the unobserved confounders and proposes a Contrastive regulariser which learns a proxy representation...
This paper proposes a framework inspired by Laurentzian audacity theory like learn directed graph representation. This framework is evaluted in different tasks and show how this framework can represent graphs with directed cycles. Strength - The experiments results show the superiority of the proposed spacetime graph....	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: This paper proposes a framework inspired by Laurentzian audacity theory like learn directed graph representation. This framework is evaluted in different tasks and show how this framework can represent graphs with directed cycles. Strength - The experiments results show the superiority of the proposed spacetim...
The paper investigates contrastive learning in the context of building sentence representations. It finds that the decoupled form that decomposes the contrastive loss into aligment and uniformity components can have a detrimental effect in this setting, due to different training dynamics compared to the InfoNCE loss. T...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper investigates contrastive learning in the context of building sentence representations. It finds that the decoupled form that decomposes the contrastive loss into aligment and uniformity components can have a detrimental effect in this setting, due to different training dynamics compared to the InfoNCE...
This paper presents a framework for protein thermostability prediction. A large-scale protein dataset with organism-level temperature annotations is curated, and one pretraining and one tuning module are proposed for prediction. The main contributions are: 1. A dataset of protein sequences and folded 3D structures data...	Recommendation: 5: marginally below the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper presents a framework for protein thermostability prediction. A large-scale protein dataset with organism-level temperature annotations is curated, and one pretraining and one tuning module are proposed for prediction. The main contributions are: 1. A dataset of protein sequences and folded 3D structu...
This paper seeks to approximate a given high-dimensional manifold with a product of some smaller, simpler manifolds. It learns this approximate mapping using neural network based optimization. The paper motivates this approach by discussing the concept of PGA (which is essentially Tangent PCA as stated here) and a spar...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: This paper seeks to approximate a given high-dimensional manifold with a product of some smaller, simpler manifolds. It learns this approximate mapping using neural network based optimization. The paper motivates this approach by discussing the concept of PGA (which is essentially Tangent PCA as stated here) an...
The paper proposes a scalable estimation method for non-parametric Markov network structures, using regularized score matching. They first introduce necessary and sufficient conditions of conditional independence between variables in general distributions for all data types (i.e., continuous, discrete, and mixed-type)...	Recommendation: 6: marginally above the acceptance threshold	Area: General Machine Learning	Review: The paper proposes a scalable estimation method for non-parametric Markov network structures, using regularized score matching. They first introduce necessary and sufficient conditions of conditional independence between variables in general distributions for all data types (i.e., continuous, discrete, and mix...
The paper presents a method for formulating particular multivariate functions based on indices as a particular tensor train decomposition. An overview of the construction is presented, where so-called derivative functions are used to define index dependences as well as the cores of the TT representation. Analysis of th...	Recommendation: 6: marginally above the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: The paper presents a method for formulating particular multivariate functions based on indices as a particular tensor train decomposition. An overview of the construction is presented, where so-called derivative functions are used to define index dependences as well as the cores of the TT representation. Analys...
This paper provides theoretical results on the formal verifiability of message passing neural networks (MPNNs), a popular class of graph neural network models, for graph and node classification tasks. In particular, it proves that, without assuming a bound on the degree of the graph, the formal verification of output ...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper provides theoretical results on the formal verifiability of message passing neural networks (MPNNs), a popular class of graph neural network models, for graph and node classification tasks. In particular, it proves that, without assuming a bound on the degree of the graph, the formal verification of...
This paper proposes a self-supervised or pretrained method for molecular representation learning. It can be summarized as follows, #### 1. A Transformer-based backbone. #### 2. Two self-supervised learning or pretraining tasks: 3D position recovery and masked atom prediction. #### 3. A few finetuning strategies are ...	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 self-supervised or pretrained method for molecular representation learning. It can be summarized as follows, #### 1. A Transformer-based backbone. #### 2. Two self-supervised learning or pretraining tasks: 3D position recovery and masked atom prediction. #### 3. A few finetuning strateg...