review stringlengths 5 16.6k | score stringclasses 5
values | area stringclasses 12
values | text stringlengths 31 5.65k |
|---|---|---|---|
While existing training-free proxies can usually achieve compelling search results, they typically can not work consistently better than #Params in practice (White et al., 2022). So, this paper aims to improve such a state of affairs. Specifically, this paper firstly theoretically proves that the convergence of DNNs is... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
While existing training-free proxies can usually achieve compelling search results, they typically can not work consistently better than #Params in practice (White et al., 2022). So, this paper aims to improve such a state of affairs. Specifically, this paper firstly theoretically proves that the convergence of... |
In this paper, the authors focus on studying the detection of images generated by diffusion models. Specifically, this paper evaluates the performance of current methods and then analyzes the difference between GAN-generated and DM-generated images.
Strengths:
1. This paper focus on an ignored problem, the detection o... | Recommendation: 5: marginally below 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 focus on studying the detection of images generated by diffusion models. Specifically, this paper evaluates the performance of current methods and then analyzes the difference between GAN-generated and DM-generated images.
Strengths:
1. This paper focus on an ignored problem, the det... |
This paper proposes a quantum NTK framework to improve the training convergence of QNNs for the GSP problem via a symmetric ansatz design with a small effective dimension. It also proposes a novel symmetric pruning algorithm to extract the symmetric ansatz from the overparameterized and assymetric ansatz. Empirical res... | Recommendation: 8: accept, good paper | Area: General Machine Learning | Review:
This paper proposes a quantum NTK framework to improve the training convergence of QNNs for the GSP problem via a symmetric ansatz design with a small effective dimension. It also proposes a novel symmetric pruning algorithm to extract the symmetric ansatz from the overparameterized and assymetric ansatz. Empir... |
The authors study the universal approximation problem of functions using neural networks. They showed that, the minimum width to approximate any functions in $L^p(\mathcal K, \mathbb{R}^{d_y})$ and $C(\mathcal K, \mathbb{R}^{d_y})$ is at least $w^*_{\min} = \max (d_x, d_y)$ for any activation functions. Furthermore, th... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The authors study the universal approximation problem of functions using neural networks. They showed that, the minimum width to approximate any functions in $L^p(\mathcal K, \mathbb{R}^{d_y})$ and $C(\mathcal K, \mathbb{R}^{d_y})$ is at least $w^*_{\min} = \max (d_x, d_y)$ for any activation functions. Further... |
The paper proposes a locality sensitive hashing scheme, applicable to approximate nearest neighbor search, using Polar codes as the sparse collection of hash values.
The paper does not prove any rigorous guarantees or bounds on the proposed scheme. It's validated solely empirically.
It's an interesting idea, but its va... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
The paper proposes a locality sensitive hashing scheme, applicable to approximate nearest neighbor search, using Polar codes as the sparse collection of hash values.
The paper does not prove any rigorous guarantees or bounds on the proposed scheme. It's validated solely empirically.
It's an interesting idea, bu... |
Inspired by Neural collapse, this paper proposed to fix a learnable classifier as a geometric structure for few-shot incremental learning.
The pre-assigned classifier could avoid optimization conflict among sessions in the training stage. They also provided a detailed theoretical analysis to show their method could hold ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
Inspired by Neural collapse, this paper proposed to fix a learnable classifier as a geometric structure for few-shot incremental learning.
The pre-assigned classifier could avoid optimization conflict among sessions in the training stage. They also provided a detailed theoretical analysis to show their method cou... |
This paper studies the conditions under which diffusion model can work well for purification of adversarially perturbed samples. A simple diffusion-based purification method named DensePure is proposed by using majority vote and achieves higher certified accuracy in comparison with other method on CIFAR-10 and ImageNe... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper studies the conditions under which diffusion model can work well for purification of adversarially perturbed samples. A simple diffusion-based purification method named DensePure is proposed by using majority vote and achieves higher certified accuracy in comparison with other method on CIFAR-10 and... |
This paper investigates the role of regularization in the explainability of GNNs. They find that regularization pursues a balance between feature attribution and selection as well as that optimal regularization is related to the sparsity of the explanations.
Strength
======
- The paper introduces GIBE, which is simila... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper investigates the role of regularization in the explainability of GNNs. They find that regularization pursues a balance between feature attribution and selection as well as that optimal regularization is related to the sparsity of the explanations.
Strength
======
- The paper introduces GIBE, which i... |
propose a Bayesian technique by enforcing a prior over parameters for self-supervised learning. The main idea is based on the BYOL to learn the representation and combine that with Cyclical SGHMC. Paper applies the method to two datasets for semi-supervised classification and one dataset for out-of-distribution detecti... | Recommendation: 3: reject, not good enough | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
propose a Bayesian technique by enforcing a prior over parameters for self-supervised learning. The main idea is based on the BYOL to learn the representation and combine that with Cyclical SGHMC. Paper applies the method to two datasets for semi-supervised classification and one dataset for out-of-distribution... |
This paper studies the explainability of graph neural networks (GNNs) using subgraph matching. The paper first identifies the limitation of most existing works, which is the usage of black-box explainer to generate explanatory subgraph. Based on that, the authors propose a matching-then-ranking framework named MatchExp... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper studies the explainability of graph neural networks (GNNs) using subgraph matching. The paper first identifies the limitation of most existing works, which is the usage of black-box explainer to generate explanatory subgraph. Based on that, the authors propose a matching-then-ranking framework named ... |
This paper provides a comprehensive analysis of Pre-trained Protein Language Models (PPLM) and specific Pre-trained Antibody Language Models on the predictions of different antibody tasks and introduces a new pre-trained method that better utilizes antibody-specific information to achieve a pre-trained antibody languag... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper provides a comprehensive analysis of Pre-trained Protein Language Models (PPLM) and specific Pre-trained Antibody Language Models on the predictions of different antibody tasks and introduces a new pre-trained method that better utilizes antibody-specific information to achieve a pre-trained antibody... |
This paper proposed a novel approach to generate synthetic tabular data using large language models.
Strengths:
The authors proposed an interesting way to use large language models. It's a novel use case.
Weaknesses:
1. The computation cost for the proposed approach seems profitably expensive. In table 6, both GRea... | Recommendation: 5: marginally below the acceptance threshold | Area: Generative models | Review:
This paper proposed a novel approach to generate synthetic tabular data using large language models.
Strengths:
The authors proposed an interesting way to use large language models. It's a novel use case.
Weaknesses:
1. The computation cost for the proposed approach seems profitably expensive. In table 6, b... |
The research is on developing an adaptive optimization algorithm called AdaDQH, which proposes a new technique to approximate the Hessian matrix in the second-order gradient descent scheme in a computationally efficient manner. The authors demonstrate the optimizer’s performance on datasets from Natural Language Proces... | Recommendation: 8: accept, good paper | Area: Optimization (eg, convex and non-convex optimization) | Review:
The research is on developing an adaptive optimization algorithm called AdaDQH, which proposes a new technique to approximate the Hessian matrix in the second-order gradient descent scheme in a computationally efficient manner. The authors demonstrate the optimizer’s performance on datasets from Natural Languag... |
Authors propose AdaStride, a differentiable layer that learns an input-dependent and irregularly sampled downsampling.
Strengths:
* Learning adaptive downsampling is a fundamental deep learning problem with very large potential impact across all types of tasks, models and modalities.
* This approach displays two char... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
Authors propose AdaStride, a differentiable layer that learns an input-dependent and irregularly sampled downsampling.
Strengths:
* Learning adaptive downsampling is a fundamental deep learning problem with very large potential impact across all types of tasks, models and modalities.
* This approach displays ... |
This paper address the problem of building an end-to-end trained dialogue system in which responses are coherent both in terms of semantics and also in terms of personality and sentiment. The approach adopted is to use an encoder-decoder architecture in which the intermediate encoding is manipulated by a mapping funct... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper address the problem of building an end-to-end trained dialogue system in which responses are coherent both in terms of semantics and also in terms of personality and sentiment. The approach adopted is to use an encoder-decoder architecture in which the intermediate encoding is manipulated by a mappi... |
The paper proposes a framework for incorporating background knowledge into reinforcement learning, Knowledge-Grounded RL. The approach uses an embedding-based attention-mechanism model that learns to attend to external and internal knowledge based on observations. The proposed framework is agnostic to the policy traini... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper proposes a framework for incorporating background knowledge into reinforcement learning, Knowledge-Grounded RL. The approach uses an embedding-based attention-mechanism model that learns to attend to external and internal knowledge based on observations. The proposed framework is agnostic to the polic... |
The authors proposed a novel model-based meta-learning method, specifically applied to the problem of transfer learning, where support and query data come from two separate (but related) tasks. This formulation allows flexibility in terms of models that can be applied, in that the model used for support examples need n... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The authors proposed a novel model-based meta-learning method, specifically applied to the problem of transfer learning, where support and query data come from two separate (but related) tasks. This formulation allows flexibility in terms of models that can be applied, in that the model used for support example... |
The authors propose the attribution method ``Stick-breaking Path Integration'' which averages the attribution maps obtained by integrating over several different paths using the Integrated Gradient attribution method. They show superior performance qualitatively and quantitatively (pixel flipping, ROAR) compared to oth... | 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 the attribution method ``Stick-breaking Path Integration'' which averages the attribution maps obtained by integrating over several different paths using the Integrated Gradient attribution method. They show superior performance qualitatively and quantitatively (pixel flipping, ROAR) compare... |
The paper uses eye tracking studies on developers and correlates the patterns with the attention from code-related deep learning models (CodeGen and Code-J). CodeGen has higher agreement with developers (median = 0.23). The paper also reports interaction patterns. While attention-agnostic schemes (such as position) do ... | 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 uses eye tracking studies on developers and correlates the patterns with the attention from code-related deep learning models (CodeGen and Code-J). CodeGen has higher agreement with developers (median = 0.23). The paper also reports interaction patterns. While attention-agnostic schemes (such as posit... |
The authors proposed a new protein language model ProtFIM, and a new protein sequence design framework via ProtFIM.
By comparing the performance with previous models via their new evaluation scheme ( SEIFER), ProtFIM achieved similar performance but with less parameters.
Strengths:
1. The article is well written and ... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The authors proposed a new protein language model ProtFIM, and a new protein sequence design framework via ProtFIM.
By comparing the performance with previous models via their new evaluation scheme ( SEIFER), ProtFIM achieved similar performance but with less parameters.
Strengths:
1. The article is well writ... |
The paper introduces a novel GPU-optimized verifier (FGV) for geometric transformations [1]. This verifier enables training using a sampling-based [2] robust classification loss leading to robust networks that are efficiently verifiable with the FGV. For small datasets (MNIST / CIFAR10), the proposed algorithm is order... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper introduces a novel GPU-optimized verifier (FGV) for geometric transformations [1]. This verifier enables training using a sampling-based [2] robust classification loss leading to robust networks that are efficiently verifiable with the FGV. For small datasets (MNIST / CIFAR10), the proposed algorithm ... |
This paper proposes to transfer teacher knowledge in response, feature, and relation together.
Very bad paper, no novelty.
| Recommendation: 1: strong reject | Area: Deep Learning and representational learning | Review:
This paper proposes to transfer teacher knowledge in response, feature, and relation together.
Very bad paper, no novelty.
Recommendation:
1 |
This paper introduces a framework to improve and certify robustness against certain geometric perturbations. The main contribution lies in the parallelization of geometric transformations and their inverses. These improvements enable efficient training as proposed in the paper, as well as efficient LiRPA verification. ... | 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 introduces a framework to improve and certify robustness against certain geometric perturbations. The main contribution lies in the parallelization of geometric transformations and their inverses. These improvements enable efficient training as proposed in the paper, as well as efficient LiRPA verifi... |
This paper proposes an algorithm Decentralized Optimistic hypeRpolicy mIrror deScent (DORIS) for decentralized policy learning in Markov games with function approximations. The setting is "policy revealing", where the opponents (could be adversarial) would reveal its previous policies to the agent for making decisions.... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes an algorithm Decentralized Optimistic hypeRpolicy mIrror deScent (DORIS) for decentralized policy learning in Markov games with function approximations. The setting is "policy revealing", where the opponents (could be adversarial) would reveal its previous policies to the agent for making de... |
This paper motivates the need for Pareto optimal solutions from the fact that most domain generalization methods can be seen as regularized ERM problems that struggle to outperform ERM.
The paper suggests that this is due to two key problems. One is the approximation of the original OOD objective (often encoded as a re... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper motivates the need for Pareto optimal solutions from the fact that most domain generalization methods can be seen as regularized ERM problems that struggle to outperform ERM.
The paper suggests that this is due to two key problems. One is the approximation of the original OOD objective (often encoded... |
This paper proposes a method for controllable generation of images conditioned on image collages as input. The proposed approach, called “mixing and matching scenes” (M&Ms), is an adversarially trained generative image model conditioned on appearance features and spatial positions of collage elements. M&M combines the... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
This paper proposes a method for controllable generation of images conditioned on image collages as input. The proposed approach, called “mixing and matching scenes” (M&Ms), is an adversarially trained generative image model conditioned on appearance features and spatial positions of collage elements. M&M comb... |
The paper consider a class of predictive models that are quadratic in parameters (equation 1 and 19), and prove progressive sharpening (section 2.1) and edge of stability (ESO, section 2.2) in two dimension. The paper also claims that the two properties could be general property in high dimensional non-linear models.... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The paper consider a class of predictive models that are quadratic in parameters (equation 1 and 19), and prove progressive sharpening (section 2.1) and edge of stability (ESO, section 2.2) in two dimension. The paper also claims that the two properties could be general property in high dimensional non-linear... |
The manuscript observes the problem of fixed contributions of ground truth knowledge and teacher knowledge at different blocks of the student networks during knowledge distillation training. The author proposes a bi-level optimization scheme to balance the knowledge on the lower level and update the network based on op... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The manuscript observes the problem of fixed contributions of ground truth knowledge and teacher knowledge at different blocks of the student networks during knowledge distillation training. The author proposes a bi-level optimization scheme to balance the knowledge on the lower level and update the network bas... |
The authors point out drawbacks of current benchmarks that claim to evaluate memory in RL.
They propose a novel benchmark suite for evaluating the memory component of agents trained in partially observable environments.
The benchmark consists of three novel environments that vary in their difficulty and can be scaled t... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The authors point out drawbacks of current benchmarks that claim to evaluate memory in RL.
They propose a novel benchmark suite for evaluating the memory component of agents trained in partially observable environments.
The benchmark consists of three novel environments that vary in their difficulty and can be ... |
The paper proposes to pre-compute contextual token embeddings and use them in the decoder of generative retrievers (as a way to enable them to use external knowledge). The model performs favorably on the KILT retrieval tasks when trained on a subset of KILT.
Strengths:
- It is a reasonable approach to injecting exter... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper proposes to pre-compute contextual token embeddings and use them in the decoder of generative retrievers (as a way to enable them to use external knowledge). The model performs favorably on the KILT retrieval tasks when trained on a subset of KILT.
Strengths:
- It is a reasonable approach to injecti... |
The paper presents a method for differentiable volume rendering using a set of Gaussian ellipsoid primitives. The key idea is to use a large set of Gaussian ellipsoids to parameterize the density and color within a volume and then to render using ray marching to solve the volume rendering equation. Conveniently, integr... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper presents a method for differentiable volume rendering using a set of Gaussian ellipsoid primitives. The key idea is to use a large set of Gaussian ellipsoids to parameterize the density and color within a volume and then to render using ray marching to solve the volume rendering equation. Conveniently... |
This paper proposes a method for evaluating the autonomy perception module from the perspective of a planner. This is an important topic and starts to receive increasing attention in recent years, yet still underexplored. Specifically, the paper proposes to formulate the expected utility maximization objective as an Hi... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a method for evaluating the autonomy perception module from the perspective of a planner. This is an important topic and starts to receive increasing attention in recent years, yet still underexplored. Specifically, the paper proposes to formulate the expected utility maximization objective ... |
A new approach is introduced for constructing benchmarks for the application of consistent rules in language models. This approach is used to create a new dataset that is then used to evaluate T5 models on 3 dimensions: applying learned rules, scaling to larger sets of rules and compositional generalization.
**Streng... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
A new approach is introduced for constructing benchmarks for the application of consistent rules in language models. This approach is used to create a new dataset that is then used to evaluate T5 models on 3 dimensions: applying learned rules, scaling to larger sets of rules and compositional generalization.
... |
The paper proposes a gradient based rule for selecting intervention targets that is used with the ENCO experimental causal discovery algorithm. Empirical results are favorable for the proposed method compared to other procedures for ranking intervention targets.
The paper is generally well written and the empirical res... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The paper proposes a gradient based rule for selecting intervention targets that is used with the ENCO experimental causal discovery algorithm. Empirical results are favorable for the proposed method compared to other procedures for ranking intervention targets.
The paper is generally well written and the empir... |
This paper proposes a network family named Multi-head Recurrent Layer Attention (MRLA). It has two module variants: MRLA and MRLA-light. Both of the variants take a self-attention format. A major difference is that MRLA looks back at its previous module’s keys and values, while the light variant only uses the input fea... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a network family named Multi-head Recurrent Layer Attention (MRLA). It has two module variants: MRLA and MRLA-light. Both of the variants take a self-attention format. A major difference is that MRLA looks back at its previous module’s keys and values, while the light variant only uses the i... |
The authors propose a domain-invariant adaptation approach for listwise learning to rank (l2r). The authors extend a bound from the literature to l2r setting. Then a typical approach for domain adaptation is used with adversarial learning. The authors experiment in the textual domain with one source and three target do... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The authors propose a domain-invariant adaptation approach for listwise learning to rank (l2r). The authors extend a bound from the literature to l2r setting. Then a typical approach for domain adaptation is used with adversarial learning. The authors experiment in the textual domain with one source and three t... |
The paper proposes a semi-Markov conditional random field model that integrates a filtering step to eliminate irrelevant segments when performing named entity recognition in text. According to the authors this helps reducing the complexity compared to a semi-Markov CRF and dramatically reduces the search space.
Stren... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper proposes a semi-Markov conditional random field model that integrates a filtering step to eliminate irrelevant segments when performing named entity recognition in text. According to the authors this helps reducing the complexity compared to a semi-Markov CRF and dramatically reduces the search space.... |
This paper applies two approaches for improving the speed of ASR models: learning to identify salient tokens and skip those that are identified not salient, and weight pruning. The results show some improvements on the SUPERB set and LibriSpeech-100h. The authors also benchmark the inference speed on Pixel 3 phone and ... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper applies two approaches for improving the speed of ASR models: learning to identify salient tokens and skip those that are identified not salient, and weight pruning. The results show some improvements on the SUPERB set and LibriSpeech-100h. The authors also benchmark the inference speed on Pixel 3 ph... |
This paper proposes a causal framework to formalize SSDA, and theoretically explain what characteristics should a robust domain adaptation model have. They also discuss the maximal training data utilization and present a collaboratively debiasing learning framework to make use of the training data.
Pros:
* This paper u... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a causal framework to formalize SSDA, and theoretically explain what characteristics should a robust domain adaptation model have. They also discuss the maximal training data utilization and present a collaboratively debiasing learning framework to make use of the training data.
Pros:
* This... |
The paper investigates the application of transformer models to repairing circuits given formal specifications. The transformer takes a faulty circuit represented in AIGER format and a specification in LTL, and outputs a circuit that hopefully satisfies the specification. The paper also comes with a detailed descriptio... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper investigates the application of transformer models to repairing circuits given formal specifications. The transformer takes a faulty circuit represented in AIGER format and a specification in LTL, and outputs a circuit that hopefully satisfies the specification. The paper also comes with a detailed de... |
This paper proposes a new framework for using pre-trained models for various recognition and reasoning tasks. The idea is to use a "generator" and an ensemble of "scorers" to iteratively improve the output: the generator proposes an output that the scorer scores. This general framework is demonstrated on multiple tasks... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper proposes a new framework for using pre-trained models for various recognition and reasoning tasks. The idea is to use a "generator" and an ensemble of "scorers" to iteratively improve the output: the generator proposes an output that the scorer scores. This general framework is demonstrated on multip... |
This paper tackles the problem of Out-of-distribution detection. A pre-trained object-detection network is firstly used to build a semantic graph. Then the graph embedding is further used to detect OOD samples. The experiment is conducted on the far-OOD and near-OOD setting.
Strength:
1. The idea of building a sema... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper tackles the problem of Out-of-distribution detection. A pre-trained object-detection network is firstly used to build a semantic graph. Then the graph embedding is further used to detect OOD samples. The experiment is conducted on the far-OOD and near-OOD setting.
Strength:
1. The idea of buildin... |
This paper proposed a new distance based on Lyapunov Spectrum (LS) for neural network hyperpruning. Here, the authors define ‘hyperpruning’ as finding a suitable pruning method with optimal hyperparameter configuration. The authors apply this method to RNN language model benchmarks and show good results on Penn Treeban... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposed a new distance based on Lyapunov Spectrum (LS) for neural network hyperpruning. Here, the authors define ‘hyperpruning’ as finding a suitable pruning method with optimal hyperparameter configuration. The authors apply this method to RNN language model benchmarks and show good results on Penn... |
The manuscript proposed a neural architecture search (NAS) algorithm, FedorAS, under cross-device federated learning setting. The proposed algorithm follows settings of one-shot NAS utilizing a super-net search space with weight sharing. It enables knowledge exchange between server and clients via sub-networks. And a n... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The manuscript proposed a neural architecture search (NAS) algorithm, FedorAS, under cross-device federated learning setting. The proposed algorithm follows settings of one-shot NAS utilizing a super-net search space with weight sharing. It enables knowledge exchange between server and clients via sub-networks.... |
In their work, the authors focus on congestion games and examine information theoretic conditions for identifying NE. Specifically, they identify the minimal assumptions required to efficiently determine if a policy is a Nash Equilibrium of a congestion game from a given dataset. The authors examine 3 different models,... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
In their work, the authors focus on congestion games and examine information theoretic conditions for identifying NE. Specifically, they identify the minimal assumptions required to efficiently determine if a policy is a Nash Equilibrium of a congestion game from a given dataset. The authors examine 3 different... |
This work synthesized classification datasets motivated by high-quality images generated by the pre-trained text-to-image model (GLIDE).
They used the dataset on three tasks: 1) zero-shot classification, 2) few-shot classification, and 3) pre-training for transfer learning.
They tuned the classifier of pre-trained CLIP... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This work synthesized classification datasets motivated by high-quality images generated by the pre-trained text-to-image model (GLIDE).
They used the dataset on three tasks: 1) zero-shot classification, 2) few-shot classification, and 3) pre-training for transfer learning.
They tuned the classifier of pre-trai... |
This paper proposes an acceleration method of guided diffusion sampling based on splitting numerical methods. Based on the finding that the high-order numerical methods are unsuitable for the conditional function, it develops a method based on Strang splitting and a combination of fourth and first-order numerical metho... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
This paper proposes an acceleration method of guided diffusion sampling based on splitting numerical methods. Based on the finding that the high-order numerical methods are unsuitable for the conditional function, it develops a method based on Strang splitting and a combination of fourth and first-order numeric... |
The authors propose an approach to speed up transformer based forecasting / enable it to scale better for longer context (history) windows feeding into the transformer forecasting model. Typically in space and time complexity transformer forecasting would scale quadratically with this context window (sequence) length.... | Recommendation: 1: strong reject | Area: Deep Learning and representational learning | Review:
The authors propose an approach to speed up transformer based forecasting / enable it to scale better for longer context (history) windows feeding into the transformer forecasting model. Typically in space and time complexity transformer forecasting would scale quadratically with this context window (sequence)... |
This paper points out that the predictive coding implementation from Rao and Ballard is essentially expectation maximization, and therefore can be sped up by using the incremental (aka partial) variant of EM instead of the full version. The paper also argues that this is more efficient than backpropagation, at least fo... | Recommendation: 5: marginally below the acceptance threshold | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
This paper points out that the predictive coding implementation from Rao and Ballard is essentially expectation maximization, and therefore can be sped up by using the incremental (aka partial) variant of EM instead of the full version. The paper also argues that this is more efficient than backpropagation, at ... |
This paper proposes a new approach for hyperparameter optimization (Deep Ranking Ensembles) in both the transfer and non-transfer learning setting.
The are 3 main components to this approach:
1. Use neural networks that learn to rank configurations as a surrogate model in the Bayesian optimization setting.
2. Create... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper proposes a new approach for hyperparameter optimization (Deep Ranking Ensembles) in both the transfer and non-transfer learning setting.
The are 3 main components to this approach:
1. Use neural networks that learn to rank configurations as a surrogate model in the Bayesian optimization setting.
2... |
The paper develops schemes to protect distributed gradient descent from stragglers building on previous work that induces redundancy in computation via error correcting codes. The authors specifically build on a technique called gradient coding that enables a master node to recover the sum of the gradients from $n$ dis... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
The paper develops schemes to protect distributed gradient descent from stragglers building on previous work that induces redundancy in computation via error correcting codes. The authors specifically build on a technique called gradient coding that enables a master node to recover the sum of the gradients from... |
The authors proposed a method, ARHGA, for inferring missing attributes and calibrating defective attributes in a heterogenous graph. The main idea of the paper is to subsample the graph and augment the attributes. The attributes were filled in using topological embeddings (based on random walk 2 vec on meta-paths) and ... | Recommendation: 1: strong reject | Area: Deep Learning and representational learning | Review:
The authors proposed a method, ARHGA, for inferring missing attributes and calibrating defective attributes in a heterogenous graph. The main idea of the paper is to subsample the graph and augment the attributes. The attributes were filled in using topological embeddings (based on random walk 2 vec on meta-pat... |
Consider the problem of synthetically generating data based on real data with multiple tables and links between tables, such as a movie recommendation dataset, and if possible, in a differentially private way. This paper suggests a new approach for doing so and applies it to real data.
Strengths
+ Solves a novel probl... | Recommendation: 8: accept, good paper | Area: Generative models | Review:
Consider the problem of synthetically generating data based on real data with multiple tables and links between tables, such as a movie recommendation dataset, and if possible, in a differentially private way. This paper suggests a new approach for doing so and applies it to real data.
Strengths
+ Solves a nov... |
This paper extends the stochastic latent actor-critic method to safe reinforcement learning by a safety critic. The experiments show the competitive performance of the proposed methods over existing approaches on benchmark pixel input tasks.
Strength:
The presentation of the paper is very clear.
The experimenta... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper extends the stochastic latent actor-critic method to safe reinforcement learning by a safety critic. The experiments show the competitive performance of the proposed methods over existing approaches on benchmark pixel input tasks.
Strength:
The presentation of the paper is very clear.
The exp... |
This paper presents a new concept, semirobusness, and a corresponding framework to study if the (semi-)robustness of a subnetwork can indicate the (semi-)robustness of the rest or the all network. The major contribution is a set of theorems to show that under assumptions related to the mutual information between repres... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper presents a new concept, semirobusness, and a corresponding framework to study if the (semi-)robustness of a subnetwork can indicate the (semi-)robustness of the rest or the all network. The major contribution is a set of theorems to show that under assumptions related to the mutual information betwee... |
Learning to Optimize (L2O) has emerged as a data-driven way to derive “learned optimizers” for specific problem classes. However, for very different unseen problems those learned optimizers can fail badly. This paper is the first to discuss L2O’s test-time self-adaptation to out-of-distribution tasks, providing both th... | Recommendation: 8: accept, good paper | Area: General Machine Learning | Review:
Learning to Optimize (L2O) has emerged as a data-driven way to derive “learned optimizers” for specific problem classes. However, for very different unseen problems those learned optimizers can fail badly. This paper is the first to discuss L2O’s test-time self-adaptation to out-of-distribution tasks, providing... |
This paper presents a model for image captioning, utilizing external knowledge with a graph-based model. The full model contains a transformer (a text tokenizer + a ViT based patch tokenizer), and a graph attention network to get the external knowledge from conceptnet 5.0. The paper claims to achieve state-of-the-art p... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper presents a model for image captioning, utilizing external knowledge with a graph-based model. The full model contains a transformer (a text tokenizer + a ViT based patch tokenizer), and a graph attention network to get the external knowledge from conceptnet 5.0. The paper claims to achieve state-of-t... |
This paper considered the problem of unsupervised test adaptation under covariate shift to achieve good fairness-accuracy trade-offs when a small amount of unlabeled data is available. The authors proposed a new weighted entropy based loss function to account for covariate shift, in combination with a representation ma... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper considered the problem of unsupervised test adaptation under covariate shift to achieve good fairness-accuracy trade-offs when a small amount of unlabeled data is available. The authors proposed a new weighted entropy based loss function to account for covariate shift, in combination with a represent... |
This paper studies multitask RL and proposes a new framework based on minimum description length (MDL) principle, called MDL-control (MDL-C). The framework contains the learning of the control policy and the default policy. The control policy is trained according to Regularized Polic Optimization regarding the default ... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper studies multitask RL and proposes a new framework based on minimum description length (MDL) principle, called MDL-control (MDL-C). The framework contains the learning of the control policy and the default policy. The control policy is trained according to Regularized Polic Optimization regarding the ... |
The authors study a relaxation of the attention mechanism. For this relaxation (which is still a non-convex problem) they propose a convexification. For single output models, they show how to solve the original problem using the convex equivalent problem.
Strength:
- the paper is in general easy to follow
Weaknesses:... | Recommendation: 3: reject, not good enough | Area: Optimization (eg, convex and non-convex optimization) | Review:
The authors study a relaxation of the attention mechanism. For this relaxation (which is still a non-convex problem) they propose a convexification. For single output models, they show how to solve the original problem using the convex equivalent problem.
Strength:
- the paper is in general easy to follow
Wea... |
This paper tackles the problem of indirect bias. The authors define how indirect bias in a corpus can be quantified, then introduce a method to reweight adjectives to correct their frequency in association with gendered nouns. These methods are then tested on a variety of embeddings, showing marginal improvements in so... | 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 tackles the problem of indirect bias. The authors define how indirect bias in a corpus can be quantified, then introduce a method to reweight adjectives to correct their frequency in association with gendered nouns. These methods are then tested on a variety of embeddings, showing marginal improvemen... |
To prompt token interaction over time, this paper proposes PatchBlender, a learnable matrix to mix tokens over time. Via inserting the PatchBlender in the middle layers of ViT, it improves the ability for temporal modeling. Experiments on MOVi-A and Something-Something v2 support the effectiveness of PatchBlender.
Stre... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
To prompt token interaction over time, this paper proposes PatchBlender, a learnable matrix to mix tokens over time. Via inserting the PatchBlender in the middle layers of ViT, it improves the ability for temporal modeling. Experiments on MOVi-A and Something-Something v2 support the effectiveness of PatchBlend... |
The paper proposed and analyzed the risk-aware reinforcement learning algorithm RA-UCB with coherent risk measures and non-linear function approximation.
**Strength:**
- The paper provides a risk-aware RL algorithm under coherent risk measures with a sub-linear performance guarantee.
- The authors overcome the challen... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The paper proposed and analyzed the risk-aware reinforcement learning algorithm RA-UCB with coherent risk measures and non-linear function approximation.
**Strength:**
- The paper provides a risk-aware RL algorithm under coherent risk measures with a sub-linear performance guarantee.
- The authors overcome the... |
This work proposes a novel universal molecular representation learning framework, called Uni-Mol, that incorporates 3D information. The network consisting of stack of SE(3)-equivariant transformer layers, is pretrained on two different tasks each involving molecular and protein pocket data respectively. This pretrainin... | Recommendation: 8: accept, good paper | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This work proposes a novel universal molecular representation learning framework, called Uni-Mol, that incorporates 3D information. The network consisting of stack of SE(3)-equivariant transformer layers, is pretrained on two different tasks each involving molecular and protein pocket data respectively. This pr... |
This paper considers the offline RL problem and proposed two techniques (trajectory weighting, and reward perturbation) to improve Decision Transformer (DT) and Reinforcement Learning via Supervised Learning (RvS).
Specifically, trajectory weighting transforms the trajectory distribution of the original offline dataset... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper considers the offline RL problem and proposed two techniques (trajectory weighting, and reward perturbation) to improve Decision Transformer (DT) and Reinforcement Learning via Supervised Learning (RvS).
Specifically, trajectory weighting transforms the trajectory distribution of the original offline... |
The paper studies reward-free learning for linear MDPs. In particular, the authors focus on reducing the number of deployment (or switching cost). The proposed algorithm achieves a sample complexity of $\tilde{O}(d^2H^5/\epsilon^2)$ and deployment complexity $O(H)$, where the order of $d,\epsilon$ are near-optimal in t... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The paper studies reward-free learning for linear MDPs. In particular, the authors focus on reducing the number of deployment (or switching cost). The proposed algorithm achieves a sample complexity of $\tilde{O}(d^2H^5/\epsilon^2)$ and deployment complexity $O(H)$, where the order of $d,\epsilon$ are near-opti... |
The paper focuses on the challenges of estimating post-nonlinear models (PNLs) for causal discovery in the bivariate case. Indeed, solving the practical issues of estimating PNLs is an important topic.
The problems with existing methods are that they can produce local minima which are trivial and meaningless solutions... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The paper focuses on the challenges of estimating post-nonlinear models (PNLs) for causal discovery in the bivariate case. Indeed, solving the practical issues of estimating PNLs is an important topic.
The problems with existing methods are that they can produce local minima which are trivial and meaningless s... |
This work develops a new hypothesis testing procedure for identifying harmful covariate shifts. To identify harmful covariate shifts, the authors propose to use disagreement cross entropy for multiple models and present Detectron (Disagreement) based on a permutation test.
- Strength
- The idea of using an ensemble ... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
This work develops a new hypothesis testing procedure for identifying harmful covariate shifts. To identify harmful covariate shifts, the authors propose to use disagreement cross entropy for multiple models and present Detectron (Disagreement) based on a permutation test.
- Strength
- The idea of using an e... |
The paper presents a diversity-aware query strategy for active anomaly detection. The proposed query strategy is based on DPP. A positive semi-definite pairwise similarity matrix between highest ranked anomalous instances is first constructed, and then a subset of instances is computed by maximizing its principal minor... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
The paper presents a diversity-aware query strategy for active anomaly detection. The proposed query strategy is based on DPP. A positive semi-definite pairwise similarity matrix between highest ranked anomalous instances is first constructed, and then a subset of instances is computed by maximizing its princip... |
### Summary
This paper addresses data and resource heterogeneity of each client in Federated Learning. This paper proposes a gating function $g_{\phi}(x, s_i)$ that allows easily adaptation to each client while taking sparsity into account like computation resources. As gating function generates masking parameters for... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
### Summary
This paper addresses data and resource heterogeneity of each client in Federated Learning. This paper proposes a gating function $g_{\phi}(x, s_i)$ that allows easily adaptation to each client while taking sparsity into account like computation resources. As gating function generates masking parame... |
This work studies a new learning system to model a non-equilibrium N-body system. In particular, the new method combines two learning components to jointly model the near-equilibrium (thermal) flow and far-from-equilibrium (non-thermal) particles. The experiments show that this hybrid model outperforms a model that onl... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This work studies a new learning system to model a non-equilibrium N-body system. In particular, the new method combines two learning components to jointly model the near-equilibrium (thermal) flow and far-from-equilibrium (non-thermal) particles. The experiments show that this hybrid model outperforms a model ... |
This paper proposes a new variational inference method for sampling from the distribution. The proposed method is a sort of extension of the SVGD method inspired by functional gradient boosting. That is, the drift term is modeled by the neural networks and is optimized to maximize the discrepancy under an appropriate r... | Recommendation: 8: accept, good paper | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
This paper proposes a new variational inference method for sampling from the distribution. The proposed method is a sort of extension of the SVGD method inspired by functional gradient boosting. That is, the drift term is modeled by the neural networks and is optimized to maximize the discrepancy under an appro... |
The authors propose to treat data augmentation in the context of a constrained optimization problem, where the constraints are defined using a set of desired invariances. The data augmentations are then chosen using an MCMC scheme based on the loss of the model (which is supposed to be invariant), while the strength of... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The authors propose to treat data augmentation in the context of a constrained optimization problem, where the constraints are defined using a set of desired invariances. The data augmentations are then chosen using an MCMC scheme based on the loss of the model (which is supposed to be invariant), while the str... |
This paper presents a learning-based method for computing geodesics on a triangulated mesh. The algorithm is similar to fast marching cubes with O(NlogN) complexity, except the distance updating step which is the one of the two main contributions of the paper.
The proposed distance update step encodes the 3-ring neighb... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper presents a learning-based method for computing geodesics on a triangulated mesh. The algorithm is similar to fast marching cubes with O(NlogN) complexity, except the distance updating step which is the one of the two main contributions of the paper.
The proposed distance update step encodes the 3-rin... |
This paper proposes a method for offline reinforcement that relies on limited queries to the environment for dealing with out-of-distribution actions during the online learning phase. The authors use a state-action (learned) pseudometric to compare online and offline states and determine which actions are best to query... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes a method for offline reinforcement that relies on limited queries to the environment for dealing with out-of-distribution actions during the online learning phase. The authors use a state-action (learned) pseudometric to compare online and offline states and determine which actions are best ... |
This paper studies the encoding and decoding of faces in the (macaque) brain. To study decoding, they generate faces by sampling a random vector from the latent space of a GAN. They show the faces to a macaque monkey and record the neural response. A linear model is trained to map the neural response to the latent vect... | Recommendation: 6: marginally above the acceptance threshold | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
This paper studies the encoding and decoding of faces in the (macaque) brain. To study decoding, they generate faces by sampling a random vector from the latent space of a GAN. They show the faces to a macaque monkey and record the neural response. A linear model is trained to map the neural response to the lat... |
1, Introduce a task called multi-range relational modeling.
2, Their EurNet constructs the multi-relational graph, where each type of edge corresponds to short-, medium- or long-range spatial interactions.
3, The proposed GRMP separately performs (1) relational message aggregation on each individual feature channel and... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
1, Introduce a task called multi-range relational modeling.
2, Their EurNet constructs the multi-relational graph, where each type of edge corresponds to short-, medium- or long-range spatial interactions.
3, The proposed GRMP separately performs (1) relational message aggregation on each individual feature cha... |
I am not in good position to review
I am not in good position to review
I am not in good position to review | Recommendation: 5: marginally below the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
I am not in good position to review
I am not in good position to review
I am not in good position to review
Recommendation:
5 |
This paper introduces a layer-wise pruning method by minimizing the reconstruction errors of nonlinear outputs. Unlike the previous methods that compute the errors before the nonlinear function, the proposed method does it after the nonlinearity and focuses on reducing the errors in the non-saturating regions. The auth... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper introduces a layer-wise pruning method by minimizing the reconstruction errors of nonlinear outputs. Unlike the previous methods that compute the errors before the nonlinear function, the proposed method does it after the nonlinearity and focuses on reducing the errors in the non-saturating regions. ... |
This paper studies how skip connections affect CNNs in the sense of learning theory. The authors investigate this by approximating CNN with kernels based on NTK and the gaussian process. They derive the analytic forms of the kernel and analyze their eigenvalues.
Strengths
- The effect of skip connection is rigorously ... | Recommendation: 5: marginally below the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper studies how skip connections affect CNNs in the sense of learning theory. The authors investigate this by approximating CNN with kernels based on NTK and the gaussian process. They derive the analytic forms of the kernel and analyze their eigenvalues.
Strengths
- The effect of skip connection is rig... |
In this work, the authors propose an efficient single-loop iterative procedure to compute the Gromov Wasserstein (GW) distance between two sets. The GW distance can be formulated as a quadratic programming with a doubly-stochastic constraint. Many of the existing approaches for solving this optimization are double loop... | Recommendation: 8: accept, good paper | Area: Optimization (eg, convex and non-convex optimization) | Review:
In this work, the authors propose an efficient single-loop iterative procedure to compute the Gromov Wasserstein (GW) distance between two sets. The GW distance can be formulated as a quadratic programming with a doubly-stochastic constraint. Many of the existing approaches for solving this optimization are dou... |
In this paper, the authors observe the Boundary-caused Class Weights Confusion (BCWC) problem in semantic segmentation. To solve this problem, they propose E-CRF to fuse the CRF into CNN networks for more effective optimization. The E-CRF owns two advantages: use CRF to purify the feature representation of boundary pix... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
In this paper, the authors observe the Boundary-caused Class Weights Confusion (BCWC) problem in semantic segmentation. To solve this problem, they propose E-CRF to fuse the CRF into CNN networks for more effective optimization. The E-CRF owns two advantages: use CRF to purify the feature representation of boun... |
Use unsupervised skill learning with mutual information latent space skills, but derive the skill latent space from the latent space of a world model. The world model is represented with a recurrent neural network (GRU), where the internal state of the world model is the recurrent state of the GRU, and trained to match... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
Use unsupervised skill learning with mutual information latent space skills, but derive the skill latent space from the latent space of a world model. The world model is represented with a recurrent neural network (GRU), where the internal state of the world model is the recurrent state of the GRU, and trained ... |
This work proposes a novel identifiability analysis of a certain class of multimodal contrastive learning algorithms. The result is centered around a model with a latent variable partitioned into content (shared across modalities), style and and modality-specific information. The main theoretical result guarantees a fo... | Recommendation: 8: accept, good paper | Area: Unsupervised and Self-supervised learning | Review:
This work proposes a novel identifiability analysis of a certain class of multimodal contrastive learning algorithms. The result is centered around a model with a latent variable partitioned into content (shared across modalities), style and and modality-specific information. The main theoretical result guarant... |
In this paper, authors propose continuous-discrete convolution (CDConv) for joint modeling of geometric and sequential protein structures. This is motivated by the observation of regular sequence structure and irregular geometry structure of proteins. The proposed instantiation of CDConv is rotational invariant, which ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
In this paper, authors propose continuous-discrete convolution (CDConv) for joint modeling of geometric and sequential protein structures. This is motivated by the observation of regular sequence structure and irregular geometry structure of proteins. The proposed instantiation of CDConv is rotational invariant... |
This paper introduces an attack that corrupts the transformer models to extract private data from users' updates in federated learning. The threat model assumes that the server can send a malicious update to corrupt the user-side models and then can extract the original words and sentences entered from users' updates. ... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper introduces an attack that corrupts the transformer models to extract private data from users' updates in federated learning. The threat model assumes that the server can send a malicious update to corrupt the user-side models and then can extract the original words and sentences entered from users' u... |
The paper proposes and analyzes a computational model for the learning and transfer of schemas as a mean for abstraction. A computational model that decouples learning a schema from observation/emission is suggested, and its performance across a range of empirical simulations is assessed.
The paper clearly illustrates ... | Recommendation: 6: marginally above the acceptance threshold | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
The paper proposes and analyzes a computational model for the learning and transfer of schemas as a mean for abstraction. A computational model that decouples learning a schema from observation/emission is suggested, and its performance across a range of empirical simulations is assessed.
The paper clearly illu... |
This paper presents an add-on component (termed GrafT) that considers global dependencies and multi-scale information throughout the transformer network. The proposed method is evaluated on several fundamental visual recognition problems.
Strength:
The proposed method is evaluated on several fundamental vision tasks, ... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper presents an add-on component (termed GrafT) that considers global dependencies and multi-scale information throughout the transformer network. The proposed method is evaluated on several fundamental visual recognition problems.
Strength:
The proposed method is evaluated on several fundamental vision... |
The authors analyse the NNGP kernel and NTK of neural networks with sinusoidal activations. The NTK has an adjustable band-width parameter, which allows users to understand such kernel models in terms of a low-pass filter. A discussion on how the cutoff frequency is chosen is provided. The authors show that such kernel... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The authors analyse the NNGP kernel and NTK of neural networks with sinusoidal activations. The NTK has an adjustable band-width parameter, which allows users to understand such kernel models in terms of a low-pass filter. A discussion on how the cutoff frequency is chosen is provided. The authors show that suc... |
The paper discusses the use of bounded group loss (BGL) in the context of fair federated learning. The core objective is to learn a classifier that satisfies some predefined constraints on the expected loss function of each group on average across all clients. It is expected that the client-conditional data distributio... | 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 discusses the use of bounded group loss (BGL) in the context of fair federated learning. The core objective is to learn a classifier that satisfies some predefined constraints on the expected loss function of each group on average across all clients. It is expected that the client-conditional data dis... |
This paper investigates a practical setting in imitation learning, where a fraction of expert demonstrations are noisy demonstrations. The authors propose to select hard samples by measuring the uncertainty and update the model with the selected samples. The motivation comes from the neural network, which tends to fit ... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper investigates a practical setting in imitation learning, where a fraction of expert demonstrations are noisy demonstrations. The authors propose to select hard samples by measuring the uncertainty and update the model with the selected samples. The motivation comes from the neural network, which tends... |
This paper proposed a new method to learn optimization landscape features using latent space information with AE/VAE for downstream meta-learning tasks.
Strength:
1. The DoE2Vec can reconstruct a large set of functions and can be applied to downstream meta-learning tasks.
2. The combination of ELA and VAE-32 outperfor... | Recommendation: 5: marginally below the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper proposed a new method to learn optimization landscape features using latent space information with AE/VAE for downstream meta-learning tasks.
Strength:
1. The DoE2Vec can reconstruct a large set of functions and can be applied to downstream meta-learning tasks.
2. The combination of ELA and VAE-32 o... |
This paper study the generalisation capabilities of contrastive self-supervised learning models, showing that the data augmentation strategy is key to generalisation. They decompose the problem of learning visual representations with siamese networks into 3 crucial parts, alignment of positive samples, divergence of cl... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper study the generalisation capabilities of contrastive self-supervised learning models, showing that the data augmentation strategy is key to generalisation. They decompose the problem of learning visual representations with siamese networks into 3 crucial parts, alignment of positive samples, divergen... |
This paper proposes a method to train a conditional diffusion models by removing the necessary of annotation pairs. More detailed, they utilize the kNN retrieved image embedding as condition signal, and can further utilize caption embedding due to the almost-perfect joint distribution between text and images. The fram... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
This paper proposes a method to train a conditional diffusion models by removing the necessary of annotation pairs. More detailed, they utilize the kNN retrieved image embedding as condition signal, and can further utilize caption embedding due to the almost-perfect joint distribution between text and images. ... |
This paper presents an empirical study on how different robust pretraining protocols affect the robustness of different downstream tasks. In this regard, many different pretraining strategies are compared on a few small-scale pretraining-finetuning pairs (e.g., SVHN-> MNIST or CIFAR10->FMNIST) in terms of the $\ell_2$ ... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
This paper presents an empirical study on how different robust pretraining protocols affect the robustness of different downstream tasks. In this regard, many different pretraining strategies are compared on a few small-scale pretraining-finetuning pairs (e.g., SVHN-> MNIST or CIFAR10->FMNIST) in terms of the $... |
This paper proposes a new method to steal and defend transformer-based SSL encoders in both language and vision domains. The stealing can be completed using the returned representations with 40x fewer queries for the languages-related tasks. And the number of queries can be decreased further for vision encoders by util... | 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 proposes a new method to steal and defend transformer-based SSL encoders in both language and vision domains. The stealing can be completed using the returned representations with 40x fewer queries for the languages-related tasks. And the number of queries can be decreased further for vision encoders... |
The paper develops a pipeline for evaluating task-free continual algorithms using three common machine learning datasets. The core idea is to generate a stream of tasks from a dataset, e.g. mini-Imagenet, and then assign to each data instance a permutation index based on the Beta distribution to determine the time in t... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
The paper develops a pipeline for evaluating task-free continual algorithms using three common machine learning datasets. The core idea is to generate a stream of tasks from a dataset, e.g. mini-Imagenet, and then assign to each data instance a permutation index based on the Beta distribution to determine the t... |
This paper introduces variational classification (VC) by treating the inputs to the softmax layer of a classification model as latent variables with a mixture of Gaussians prior. To achieve this probabilistic interpretation of the softmax classifier, they derived an objective to be minimized, which is similar to the EL... | Recommendation: 5: marginally below the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
This paper introduces variational classification (VC) by treating the inputs to the softmax layer of a classification model as latent variables with a mixture of Gaussians prior. To achieve this probabilistic interpretation of the softmax classifier, they derived an objective to be minimized, which is similar t... |
This paper concerns language models in the pretrain-then-prompt-tune paradigm, and aims to introduce a prompting method to allow for solving multiple tasks. This can be described a continual learning setup, where the task ID is known to the model at both training and inference time. The method, called Progressive Pr... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper concerns language models in the pretrain-then-prompt-tune paradigm, and aims to introduce a prompting method to allow for solving multiple tasks. This can be described a continual learning setup, where the task ID is known to the model at both training and inference time. The method, called Progre... |
This paper analyzes the learning problem in three aspects. 1. From the perspective of information theory, when the loss is bounded and the entropy of the data distribution is small, one can learn only in the typical set; 2. Modeling RNN propagation as the iterative method that solves sparse sensing and RNN training is ... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper analyzes the learning problem in three aspects. 1. From the perspective of information theory, when the loss is bounded and the entropy of the data distribution is small, one can learn only in the typical set; 2. Modeling RNN propagation as the iterative method that solves sparse sensing and RNN trai... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.