review stringlengths 5 16.6k | score stringclasses 5
values | area stringclasses 12
values | text stringlengths 31 5.65k |
|---|---|---|---|
This paper proposed a StarGraph method to learn entity representations for large-scale knowledge graphs. The StarGraph first generates an incomplete two-hop neighborhood subgraph for each target node by picking part of the anchors and nodes within the 2-hop neighborhood. Then a transformer network is applied to obtain ... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper proposed a StarGraph method to learn entity representations for large-scale knowledge graphs. The StarGraph first generates an incomplete two-hop neighborhood subgraph for each target node by picking part of the anchors and nodes within the 2-hop neighborhood. Then a transformer network is applied to... |
This paper proposes to embed two sets of data into a common low dimensional space and at the same time estimate the correspondences between the data points of two sets. The proposed method combines Multidimensional Scaling (MDS) and Wasserstein Procrustes analysis into a joint optimization problem, which is solved usi... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
This paper proposes to embed two sets of data into a common low dimensional space and at the same time estimate the correspondences between the data points of two sets. The proposed method combines Multidimensional Scaling (MDS) and Wasserstein Procrustes analysis into a joint optimization problem, which is so... |
This paper proposes a new pipeline for zero-shot captioning. It first establishes a text decoder to inverse the text embedding from the CLIP to sentences. The authors further develop an embedding projection technique to project the image embedding to a weighted sum of the memorized text embeddings, which can significan... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a new pipeline for zero-shot captioning. It first establishes a text decoder to inverse the text embedding from the CLIP to sentences. The authors further develop an embedding projection technique to project the image embedding to a weighted sum of the memorized text embeddings, which can si... |
The authors introduce a real robot RL benchmarking platform for dexterous manipulation and, as a main contribution of this paper, collect two offline RL datasets and benchmark current popular Offline RL algorithms on these two datasets. This provides a point of reference to track future progress in Offline RL algorithm... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The authors introduce a real robot RL benchmarking platform for dexterous manipulation and, as a main contribution of this paper, collect two offline RL datasets and benchmark current popular Offline RL algorithms on these two datasets. This provides a point of reference to track future progress in Offline RL a... |
This paper proposes a new time-varying integration duration for HMC using the roots of Chebyshev polynomials.
In a very restrictive setting, they paper proves an upper bound on the number of iterations required to reach Wasserstrain-2 distance less than an a specified error threshold. This setting is where exact HMC m... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper proposes a new time-varying integration duration for HMC using the roots of Chebyshev polynomials.
In a very restrictive setting, they paper proves an upper bound on the number of iterations required to reach Wasserstrain-2 distance less than an a specified error threshold. This setting is where exa... |
This paper presents a loss function for classification that allows to mitigate the issue of spurious correlations in data where there is a majority group in which correlations are present, and a minority group without spurious correlations.
The proposed method consists of training two networks: 1) a ERM network that... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper presents a loss function for classification that allows to mitigate the issue of spurious correlations in data where there is a majority group in which correlations are present, and a minority group without spurious correlations.
The proposed method consists of training two networks: 1) a ERM netw... |
The paper presents a method for hierarchical representation learning of spatiotemporal features in long-term video prediction. The proposed method is called: Dynamic Latent Hierarchy (DLH). The method distinguishes between features that are changing and those that are not changing in the video sequence. DLH is able to ... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
The paper presents a method for hierarchical representation learning of spatiotemporal features in long-term video prediction. The proposed method is called: Dynamic Latent Hierarchy (DLH). The method distinguishes between features that are changing and those that are not changing in the video sequence. DLH is ... |
The paper aims at building "novel algorithm that designs exploration incentives via learnable representations of the dynamics model by embedding the neural dynamics into a kernel space induced by the system noise."
Strength:
The aim of the paper seems ambitious and interesting
Weaknesses:
The paper is not clear
The pa... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper aims at building "novel algorithm that designs exploration incentives via learnable representations of the dynamics model by embedding the neural dynamics into a kernel space induced by the system noise."
Strength:
The aim of the paper seems ambitious and interesting
Weaknesses:
The paper is not clea... |
In this paper, the authors propose a method to train efficient models with limited labeled data. By using task similarity metrics, they conduct weighted knowledge distillation with pretrained models from different sources. It achieves better results than transferring from ImageNet pretrained model or FixMatch.
Strength... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
In this paper, the authors propose a method to train efficient models with limited labeled data. By using task similarity metrics, they conduct weighted knowledge distillation with pretrained models from different sources. It achieves better results than transferring from ImageNet pretrained model or FixMatch.
... |
Motivated by quadrature rules, this paper studies the convergence of shallow ReLU networks trained on a quadratic loss defined by a symmetric positive definite matrix $P$. They study the convergence rate along different components of the matrix determined by the NTK and the matrix $P$ for general target functions. Wh... | Recommendation: 8: accept, good paper | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
Motivated by quadrature rules, this paper studies the convergence of shallow ReLU networks trained on a quadratic loss defined by a symmetric positive definite matrix $P$. They study the convergence rate along different components of the matrix determined by the NTK and the matrix $P$ for general target functi... |
The authors proposed a robust GAN-inversion method to cope with restoring images with unknown corruption and detecting unknown defects. They addressed the problem by involving the mask in the optimization process, and encouraging it to be sparse. Besides, to further close the GAN gap, they proposed to finetune the GAN ... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
The authors proposed a robust GAN-inversion method to cope with restoring images with unknown corruption and detecting unknown defects. They addressed the problem by involving the mask in the optimization process, and encouraging it to be sparse. Besides, to further close the GAN gap, they proposed to finetune ... |
The authors construct an architecture based on sparse-distributed memory (SDM) and multi-layer perceptrons and show that this architecture is naturally robust to catastrophic forgetting. This architecture uses sparse activations of neurons (top-K) with a couple of tricks to improve trainability. One one continual learn... | Recommendation: 8: accept, good paper | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
The authors construct an architecture based on sparse-distributed memory (SDM) and multi-layer perceptrons and show that this architecture is naturally robust to catastrophic forgetting. This architecture uses sparse activations of neurons (top-K) with a couple of tricks to improve trainability. One one continu... |
This paper studies the pruning problem of graph neural networks (together with the adjacency matrix). It claims that the GNN performance drops sharply when the graph sparsity is beyond a certain extent in UGS. To address this issue, this work adopts two approaches: (1) adding a new auxiliary loss based on WD to better ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper studies the pruning problem of graph neural networks (together with the adjacency matrix). It claims that the GNN performance drops sharply when the graph sparsity is beyond a certain extent in UGS. To address this issue, this work adopts two approaches: (1) adding a new auxiliary loss based on WD to... |
Mostly survey, not clear what the contribution is. The submission considers option RL. The major challenges the submission tackles is (1) the poor sample efficiency and (2) hard optimization for such a problem. The authors argue that the cause of the two challenges is the semi-MDP framework of existing option RL method... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
Mostly survey, not clear what the contribution is. The submission considers option RL. The major challenges the submission tackles is (1) the poor sample efficiency and (2) hard optimization for such a problem. The authors argue that the cause of the two challenges is the semi-MDP framework of existing option R... |
This paper introduces differential private diffusion models (SPDMs) and studies DP-SGD for DPDMs training. This paper proposes a modification of the diffusion model training objective called noise multiplicity to boost performance. The DPDMs experiments show improvement in image generation tasks.
Strengths:
1. Studyi... | Recommendation: 3: reject, not good enough | Area: Generative models | Review:
This paper introduces differential private diffusion models (SPDMs) and studies DP-SGD for DPDMs training. This paper proposes a modification of the diffusion model training objective called noise multiplicity to boost performance. The DPDMs experiments show improvement in image generation tasks.
Strengths:
1... |
This paper proposes a measure, namely Image Downscaling Assessment by Rate-Distortion (IDA-RD), to quantitatively evaluate image downscaling algorithms.
The authors use traditional downscaling algorithms such as bicubic, bilinear, and nearest neighbor interpolation, as well as state-of-the-art downscaling algorithms to... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a measure, namely Image Downscaling Assessment by Rate-Distortion (IDA-RD), to quantitatively evaluate image downscaling algorithms.
The authors use traditional downscaling algorithms such as bicubic, bilinear, and nearest neighbor interpolation, as well as state-of-the-art downscaling algor... |
This paper analyzes the effect of various activation functions on server accuracy in federated learning with heterogeneous clients. It shows that Tanh-based activation functions outperform ReLU-based activation functions in most cases, providing the guidelines for selecting activation functions in various federated lea... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
This paper analyzes the effect of various activation functions on server accuracy in federated learning with heterogeneous clients. It shows that Tanh-based activation functions outperform ReLU-based activation functions in most cases, providing the guidelines for selecting activation functions in various feder... |
The paper proposes HYPER, a mechanism for transfer learning among retrieval tasks in different domains and datasets. For each query, it first dynamically generates a hyper-prompt to synthesize the query by attending to a given number of shared prompts, then encodes the synthesized prompt into prefixes, and employs the ... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper proposes HYPER, a mechanism for transfer learning among retrieval tasks in different domains and datasets. For each query, it first dynamically generates a hyper-prompt to synthesize the query by attending to a given number of shared prompts, then encodes the synthesized prompt into prefixes, and empl... |
This paper extends an SDP approach for Gaussian mixture models clustering by introducing cluster labels as model parameters.
Strength
======
- The paper is well written, except for the references/citation style. It's really hard to find the papers that you cited in the main manuscript in the References section as cita... | Recommendation: 3: reject, not good enough | Area: Unsupervised and Self-supervised learning | Review:
This paper extends an SDP approach for Gaussian mixture models clustering by introducing cluster labels as model parameters.
Strength
======
- The paper is well written, except for the references/citation style. It's really hard to find the papers that you cited in the main manuscript in the References section... |
The paper propose STEDIE, a new model that disentangles object representations based on interactions, into interaction-relevant relational features and interaction-irrelevant global features without direct supervision. Across different experiments, the authors showcases the model's effectiveness.
I want to preface this... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
The paper propose STEDIE, a new model that disentangles object representations based on interactions, into interaction-relevant relational features and interaction-irrelevant global features without direct supervision. Across different experiments, the authors showcases the model's effectiveness.
I want to pref... |
The paper introduces Denoising Diffusion Sampler. The task at hand is to sample from an un-normalized target density with an unknown normalization constant; and also to estimate said constant. Inspired by Denoising Diffusion generative models, the paper proposes a reverse-time SDE of a process that diffuses the target ... | Recommendation: 6: marginally above the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The paper introduces Denoising Diffusion Sampler. The task at hand is to sample from an un-normalized target density with an unknown normalization constant; and also to estimate said constant. Inspired by Denoising Diffusion generative models, the paper proposes a reverse-time SDE of a process that diffuses the... |
This work presents an approach for learning better policies for imitation learning (both offline/online) in the presence of action noise; notably, this work makes a distinction between state-independent action noise (e.g., an annotator picking a random action for a state) vs. state-dependent action noise (e.g., a compl... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This work presents an approach for learning better policies for imitation learning (both offline/online) in the presence of action noise; notably, this work makes a distinction between state-independent action noise (e.g., an annotator picking a random action for a state) vs. state-dependent action noise (e.g.,... |
The authors present a method to synthesize photo-realistic images of objects from a given category while the approach allows for disentangled representation in terms of the camera pose, crude object shape, and foreground/background appearance. The approach introduces the concept of template object coordinates (TOCS) wh... | Recommendation: 8: accept, good paper | Area: Generative models | Review:
The authors present a method to synthesize photo-realistic images of objects from a given category while the approach allows for disentangled representation in terms of the camera pose, crude object shape, and foreground/background appearance. The approach introduces the concept of template object coordinates (... |
The paper proposes to combine graph neural networks and neural ordinary differential equations (ODEs) to make the learned model generalize to particle systems with arbitrary sizes. It further studies how different inductive biases that could be embedded into the framework affect the performance. On two synthetic datase... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper proposes to combine graph neural networks and neural ordinary differential equations (ODEs) to make the learned model generalize to particle systems with arbitrary sizes. It further studies how different inductive biases that could be embedded into the framework affect the performance. On two syntheti... |
This paper proposes a transfer-based adversarial attack for black-box video models. The surrogate model is built via prompt tuning from a pre-trained image model with additional temporal prompt tokens. It is validated that a surrogate model with the learned temporal prompts (or "dynamic cues") achieves better adversari... | 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 proposes a transfer-based adversarial attack for black-box video models. The surrogate model is built via prompt tuning from a pre-trained image model with additional temporal prompt tokens. It is validated that a surrogate model with the learned temporal prompts (or "dynamic cues") achieves better a... |
Authors believe there are problems with naive softmax function used in ML training. The design of softmax will lead to 2 potential problems: overfitting and mismatch between softmax loss and prediction goal. They proposed AS-SOFTMAX to address so. Authors addressed how these 2 problems could affect the final performanc... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
Authors believe there are problems with naive softmax function used in ML training. The design of softmax will lead to 2 potential problems: overfitting and mismatch between softmax loss and prediction goal. They proposed AS-SOFTMAX to address so. Authors addressed how these 2 problems could affect the final pe... |
This paper introduced kernel weak quadratic costs to solve the problem that NOT with the weak quadratic cost might learn fake plans which are not optimal.
And it conducts the theoretical and empirical analysis of the saddle point optimization problem of NOT algorithm for the weak quadratic cost.
Finally, they analyze... | Recommendation: 5: marginally below the acceptance threshold | Area: Generative models | Review:
This paper introduced kernel weak quadratic costs to solve the problem that NOT with the weak quadratic cost might learn fake plans which are not optimal.
And it conducts the theoretical and empirical analysis of the saddle point optimization problem of NOT algorithm for the weak quadratic cost.
Finally, they... |
This paper considers a new class of f-divergences that incorporate Lipschitz continuous functions in the variational representation of the f-divergence. These new class of divergences interpolate between the 1-Wasserstein metric and f-divergence when the Lipschitz parameter of the function class varies from 0 to infini... | Recommendation: 5: marginally below the acceptance threshold | Area: Generative models | Review:
This paper considers a new class of f-divergences that incorporate Lipschitz continuous functions in the variational representation of the f-divergence. These new class of divergences interpolate between the 1-Wasserstein metric and f-divergence when the Lipschitz parameter of the function class varies from 0 t... |
The paper describes an evaluation framework based on representation similarity to study and explain the way VAEs learn and diagnose some of the main problems with this type of model. One aspect that sets this work apart is that it looks at representations at every layer, not just the inputs and latent space.
Strengths:... | Recommendation: 8: accept, good paper | Area: Generative models | Review:
The paper describes an evaluation framework based on representation similarity to study and explain the way VAEs learn and diagnose some of the main problems with this type of model. One aspect that sets this work apart is that it looks at representations at every layer, not just the inputs and latent space.
St... |
Although there have been studies that extended MARL into offline learning concepts and multi-tasking learning, no studies have attempted both simultaneously. Thus, the current research is meaningful in that it tries to tackle both offline and multi-task learning simultaneously in MARL. The principal methodology that th... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
Although there have been studies that extended MARL into offline learning concepts and multi-tasking learning, no studies have attempted both simultaneously. Thus, the current research is meaningful in that it tries to tackle both offline and multi-task learning simultaneously in MARL. The principal methodology... |
This paper studied to improve the generalization of open-world class-agnostic object detection with geometric cues. In practice, depth and normal maps predicted from pretrained Omnidata model were used to train an object proposal network for pseudo-labeling unannotated novel objects. The resulting Geometry-guided Open-... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper studied to improve the generalization of open-world class-agnostic object detection with geometric cues. In practice, depth and normal maps predicted from pretrained Omnidata model were used to train an object proposal network for pseudo-labeling unannotated novel objects. The resulting Geometry-guid... |
The paper introduces a new framework for pre-trained representations of molecules. It first discusses issues related to masked language modeling approaches for molecular graphs (AttrMask), e.g., negative transfer. To overcome these issues, the paper proposes using a VQ-VAE-based tokenizer to leverage context-dependent ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper introduces a new framework for pre-trained representations of molecules. It first discusses issues related to masked language modeling approaches for molecular graphs (AttrMask), e.g., negative transfer. To overcome these issues, the paper proposes using a VQ-VAE-based tokenizer to leverage context-de... |
In this paper, the authors propose an approach for active intervention targeting that can be applied to existing gradient-based causal discovery methods by defining a gradient-based score to guide intervention design. The authors have conducted extensive experiments and detailed analysis, but the method overall lacks s... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
In this paper, the authors propose an approach for active intervention targeting that can be applied to existing gradient-based causal discovery methods by defining a gradient-based score to guide intervention design. The authors have conducted extensive experiments and detailed analysis, but the method overall... |
This paper studies the supervised learning problem under missing data, and proposed a solution without the need for imputation based by utilizing an attention based latent space model for dealing with missingness in tabular datasets. The authors shows that certain regularization occurs for the latent space representati... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper studies the supervised learning problem under missing data, and proposed a solution without the need for imputation based by utilizing an attention based latent space model for dealing with missingness in tabular datasets. The authors shows that certain regularization occurs for the latent space repr... |
The paper presents a method to disentangle the semantic space in a GAN model by using Frechet means. The pipeline has two steps. First, the semantic subspace is constructed by the Frechet mean in the Grassmannian manifold of the intrinsic tangent spaces. Then, the Frechet basis of the semantic subspace is constructed b... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
The paper presents a method to disentangle the semantic space in a GAN model by using Frechet means. The pipeline has two steps. First, the semantic subspace is constructed by the Frechet mean in the Grassmannian manifold of the intrinsic tangent spaces. Then, the Frechet basis of the semantic subspace is const... |
This paper argued that Softmax cross entropy loss function may lead to overfitting for large-output space problems. To resolve such issue, the author propose Adaptive Sparse Ssoftmax (AS-Softmax) that masked out a non-target logit when the target logit exceed the non-target logit by a specific margin. The author also e... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper argued that Softmax cross entropy loss function may lead to overfitting for large-output space problems. To resolve such issue, the author propose Adaptive Sparse Ssoftmax (AS-Softmax) that masked out a non-target logit when the target logit exceed the non-target logit by a specific margin. The autho... |
This paper presents a new method for trajectory forecasting. The proposed method consists of two ideas. (1) It introduces a new constraint such that (1a) predicted trajectories from consecutive inputs should not be different (temporal consistency) and (1b) the predicted trajectories should also be consistent regarding ... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper presents a new method for trajectory forecasting. The proposed method consists of two ideas. (1) It introduces a new constraint such that (1a) predicted trajectories from consecutive inputs should not be different (temporal consistency) and (1b) the predicted trajectories should also be consistent re... |
The paper scales masked autoencoders (MAE) to image-text multimodal data, it investigates the use of modality-agnostic unified encoders, which learns generalizable representations across a number of downstream tasks.
Strengths:
(a) The paper is overall well-written, the problem well formulated and novel to me. Generali... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper scales masked autoencoders (MAE) to image-text multimodal data, it investigates the use of modality-agnostic unified encoders, which learns generalizable representations across a number of downstream tasks.
Strengths:
(a) The paper is overall well-written, the problem well formulated and novel to me. ... |
In the paper, the authors consider the federated learning problem. The authors provide an algorithm that compresses the gradient before sending it from client to server (to combat communication cost), accounting for error of compression from previous iteration. Two versions of the algorithm is presented: Fed-EF-SGD and... | Recommendation: 6: marginally above the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
In the paper, the authors consider the federated learning problem. The authors provide an algorithm that compresses the gradient before sending it from client to server (to combat communication cost), accounting for error of compression from previous iteration. Two versions of the algorithm is presented: Fed-EF... |
The paper proposes a framework for stochastic optimizers named Admeta. Based on SGD and R-Adam, the authors provide two implementations, AdmetaS and AdmetaR.
Specifically, the author contributes a new gradient decent algorithm framework in the style of SGD and Adam optimizer. First, the author identifies the problem ... | Recommendation: 6: marginally above the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
The paper proposes a framework for stochastic optimizers named Admeta. Based on SGD and R-Adam, the authors provide two implementations, AdmetaS and AdmetaR.
Specifically, the author contributes a new gradient decent algorithm framework in the style of SGD and Adam optimizer. First, the author identifies the ... |
The paper proposed to loss term for deep regression models to increase the entropy of the learned representation and with that increase the performance of the predictive model. An additional term is also proposed to decrease the entropy of the class conditioned representation.
Strengths
-Well written.
-Technically soun... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The paper proposed to loss term for deep regression models to increase the entropy of the learned representation and with that increase the performance of the predictive model. An additional term is also proposed to decrease the entropy of the class conditioned representation.
Strengths
-Well written.
-Technica... |
A natural way to train qualitatively different policies for cooperative tasks is to penalize a new joint policy for performing well when individual components of this policy are paired with existing policies. While this approach can identify multiple "incompatible" solutions to a cooperative task, this incompatibility... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
A natural way to train qualitatively different policies for cooperative tasks is to penalize a new joint policy for performing well when individual components of this policy are paired with existing policies. While this approach can identify multiple "incompatible" solutions to a cooperative task, this incompa... |
This work proposes two approach, optimism-based algorithm and policy-based algorithm for infinite-horizon average reward constrained linear MDP under different assumptions and achieve better convergence upper bounds.
Strength.
This paper is sufficiently complete. It proposes new algorithms, builds a theoretically-imp... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This work proposes two approach, optimism-based algorithm and policy-based algorithm for infinite-horizon average reward constrained linear MDP under different assumptions and achieve better convergence upper bounds.
Strength.
This paper is sufficiently complete. It proposes new algorithms, builds a theoretic... |
This paper presents a method for long-term (up to 300 frames) video prediction. The technical method includes
1) encoding latent code from consecutive frames,
2) a temporal transformer with autoregressive sampling,
3) a MaskGit for decoding the dynamic prior, and
4) a CNN decoder to decode the frame latent code.
To ... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
This paper presents a method for long-term (up to 300 frames) video prediction. The technical method includes
1) encoding latent code from consecutive frames,
2) a temporal transformer with autoregressive sampling,
3) a MaskGit for decoding the dynamic prior, and
4) a CNN decoder to decode the frame latent c... |
This paper investigates a common structural characteristic between winning tickets. The authors argue that the found winning tickets indeed share a similar structure, the sign of the connections. To analyze this feature quantitatively, they devise a new metric (sign-aware structural distance) that measures the degree o... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper investigates a common structural characteristic between winning tickets. The authors argue that the found winning tickets indeed share a similar structure, the sign of the connections. To analyze this feature quantitatively, they devise a new metric (sign-aware structural distance) that measures the ... |
This work develops a new method to explain predictions from black-box ML models. The method learns two separate modules to perform the following: 1) output an attribution score for each feature and each class (similar to methods like LIME), and 2) output selection probabilities for each feature (similar to methods like... | Recommendation: 3: reject, not good enough | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This work develops a new method to explain predictions from black-box ML models. The method learns two separate modules to perform the following: 1) output an attribution score for each feature and each class (similar to methods like LIME), and 2) output selection probabilities for each feature (similar to meth... |
The paper proposes an approach to learn counterfactual functions based on the properties that the axiomatic definition of counterfactuals require (composition, reversibility, effectiveness). Evaluating the degree to which this properties are satisfied allows to assess the distance between an ideal counterfactual and it... | Recommendation: 8: accept, good paper | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The paper proposes an approach to learn counterfactual functions based on the properties that the axiomatic definition of counterfactuals require (composition, reversibility, effectiveness). Evaluating the degree to which this properties are satisfied allows to assess the distance between an ideal counterfactua... |
This paper works on an interesting topic of generating images based on tabular data. It provides a benchmark dataset and proposes some approaches with varying levels of complexity and tradeoffs.
The main idea of the paper is interesting. But the motivation what we could benefit from the generated images is not very cl... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper works on an interesting topic of generating images based on tabular data. It provides a benchmark dataset and proposes some approaches with varying levels of complexity and tradeoffs.
The main idea of the paper is interesting. But the motivation what we could benefit from the generated images is not... |
This paper proposes a user interface for analyzing differences between ML models (ostensibly public cloud models, but it could be anything). While the idea of segmenting model input for the purpose of model understanding is interesting, this paper is basically describing a fairly primitive GUI for the (already publish... | Recommendation: 3: reject, not good enough | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper proposes a user interface for analyzing differences between ML models (ostensibly public cloud models, but it could be anything). While the idea of segmenting model input for the purpose of model understanding is interesting, this paper is basically describing a fairly primitive GUI for the (already... |
The paper constructs a novel type of architecture for monotonic neural networks by adding a single residual connection to an expressive (weight-constrained) Lipschitz network. The resulting network architecture is simple in implementation and theory foundations, robust, and highly expressive. The algorithm is compared ... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper constructs a novel type of architecture for monotonic neural networks by adding a single residual connection to an expressive (weight-constrained) Lipschitz network. The resulting network architecture is simple in implementation and theory foundations, robust, and highly expressive. The algorithm is c... |
This paper proposes a method for compositional visual reasoning in videos, based on recent advances in self-supervised pretraining. The method first trains a spacial-temporal transformer that reconstructs the video frames under the masked autoencoder paradigm, then performs reasoning via transfer learning. The authors ... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper proposes a method for compositional visual reasoning in videos, based on recent advances in self-supervised pretraining. The method first trains a spacial-temporal transformer that reconstructs the video frames under the masked autoencoder paradigm, then performs reasoning via transfer learning. The ... |
The paper treats instance segmentation as a graph learning problem, via casting a point cloud as a graph, followed by a feature extraction pipeline of applying the multi-head attention (MHA) module [1] and triangular multiplicative updates [2]. The final instance segmentation predictions are obtained by finding the con... | Recommendation: 3: reject, not good enough | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper treats instance segmentation as a graph learning problem, via casting a point cloud as a graph, followed by a feature extraction pipeline of applying the multi-head attention (MHA) module [1] and triangular multiplicative updates [2]. The final instance segmentation predictions are obtained by finding... |
In this paper, the authors propose generative adversarial neural tangent kernel (GA-NTK) and its variants.
They model the discriminator as a Gaussian process whose mean (and covariance) is governed by NTK.
Also, they can be performed via a single-level training process as a whole, because the training dynamics of their... | Recommendation: 8: accept, good paper | Area: Generative models | Review:
In this paper, the authors propose generative adversarial neural tangent kernel (GA-NTK) and its variants.
They model the discriminator as a Gaussian process whose mean (and covariance) is governed by NTK.
Also, they can be performed via a single-level training process as a whole, because the training dynamics ... |
Summary
This paper introduces Graph Neural Bandits (GNB) in the context of personalized recommendation. Specifically, the authors leverage graph neural networks to learn users’ “coarse-grained” collaborative instead of “fine-grained” for bandits optimization. All in all, the proposed method shows good performance acro... | Recommendation: 6: marginally above the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
Summary
This paper introduces Graph Neural Bandits (GNB) in the context of personalized recommendation. Specifically, the authors leverage graph neural networks to learn users’ “coarse-grained” collaborative instead of “fine-grained” for bandits optimization. All in all, the proposed method shows good performa... |
The paper introduces a test-time adaptation method to improve the adversarial robustness of a neural network by proposing (a) a test-time objective based on existing self-supervised training methods (e.g., RotNet [Gidaris et al., 2018]) and (b) a pre-training objective based on incorporating gradient-based meta-learnin... | Recommendation: 5: marginally below the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper introduces a test-time adaptation method to improve the adversarial robustness of a neural network by proposing (a) a test-time objective based on existing self-supervised training methods (e.g., RotNet [Gidaris et al., 2018]) and (b) a pre-training objective based on incorporating gradient-based meta... |
This paper provides an analysis of the effect of training convolutional neural networks with a particular image size in the generalisation to different image sizes at inference time, as a consequence of downscaling operations within the architecture and specifically the padding necessary for enabling flexible input siz... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper provides an analysis of the effect of training convolutional neural networks with a particular image size in the generalisation to different image sizes at inference time, as a consequence of downscaling operations within the architecture and specifically the padding necessary for enabling flexible i... |
The aurhors provide GRAPHSPLINENETS, a novel deep learning approach to speed up simulation of physical systems with spatio-temporal continuous outputs by exploiting the synergy between graph neural networks (GNN) and orthogonal spline collocation (OSC).
Two differentiable OSC (time-oriented OSC and spatial-oriented OS... | Recommendation: 3: reject, not good enough | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The aurhors provide GRAPHSPLINENETS, a novel deep learning approach to speed up simulation of physical systems with spatio-temporal continuous outputs by exploiting the synergy between graph neural networks (GNN) and orthogonal spline collocation (OSC).
Two differentiable OSC (time-oriented OSC and spatial-ori... |
**Summary**
This paper considers continuous self-training for ASR. This paper build on the observation that a previous method slimIPL performance degrades as the number of pretraining step M increase. The authors hypothesize that pretraining would cause overfitting when the supervised data is limited. To address this m... | Recommendation: 8: accept, good paper | Area: Unsupervised and Self-supervised learning | Review:
**Summary**
This paper considers continuous self-training for ASR. This paper build on the observation that a previous method slimIPL performance degrades as the number of pretraining step M increase. The authors hypothesize that pretraining would cause overfitting when the supervised data is limited. To addres... |
The paper presents a unified story for shape derivatives of neural implicit surfaces. The method enables projection onto deformations that preserve volume or are approximately rigid, and it suggests a path toward enforcing other kinds of constraints. Results are limited and only qualitative.
## Strengths
- The paper is... | Recommendation: 5: marginally below the acceptance threshold | Area: Generative models | Review:
The paper presents a unified story for shape derivatives of neural implicit surfaces. The method enables projection onto deformations that preserve volume or are approximately rigid, and it suggests a path toward enforcing other kinds of constraints. Results are limited and only qualitative.
## Strengths
- The ... |
The paper presents a method to learn 3D human from 2D image collections. By leveraging a 3D-aware generative model, the authors propose to integrate a compositional representation and a prior by SMPL, and an improved training strategy to enable digital human generation. The authors show that it is possible to sample hi... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
The paper presents a method to learn 3D human from 2D image collections. By leveraging a 3D-aware generative model, the authors propose to integrate a compositional representation and a prior by SMPL, and an improved training strategy to enable digital human generation. The authors show that it is possible to s... |
The authors suggest a new method for interpreting decisions by deep learning models using a variance-based sensitivity analysis method. By such analysis of the input features the authors claim that they can analyse not only the contribution of individual features but their relation as well, for example, such image patc... | Recommendation: 3: reject, not good enough | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The authors suggest a new method for interpreting decisions by deep learning models using a variance-based sensitivity analysis method. By such analysis of the input features the authors claim that they can analyse not only the contribution of individual features but their relation as well, for example, such im... |
This paper theoretically analyse how self-supervised pretraining can achieve better performance than full supervised training in a synthetic setting, with fixed data, a fixed architecture, and a fixed self-supervised training scheme. More precisely, it looks at how, when a precise data generating process is assumed and... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper theoretically analyse how self-supervised pretraining can achieve better performance than full supervised training in a synthetic setting, with fixed data, a fixed architecture, and a fixed self-supervised training scheme. More precisely, it looks at how, when a precise data generating process is ass... |
The manuscript introduced a bio-inspired representation learning that guarantees disentanglement in linear network, under some assumptions. The authors showed two neurobiological constraints i.e., nonnegativity and minimum energy efficiency (with respect to activity and weights), lead to linear factorization, under iid... | Recommendation: 8: accept, good paper | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
The manuscript introduced a bio-inspired representation learning that guarantees disentanglement in linear network, under some assumptions. The authors showed two neurobiological constraints i.e., nonnegativity and minimum energy efficiency (with respect to activity and weights), lead to linear factorization, u... |
The paper combines standard VAE with filter algorithms (for example extended Kalman filter). The VAE is used to encode and decode from observation space to the latent space (in the paper often referred to as pseudo-observation space). While the filter algorithm is used to estimate the marginal log likelihood over the l... | Recommendation: 3: reject, not good enough | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The paper combines standard VAE with filter algorithms (for example extended Kalman filter). The VAE is used to encode and decode from observation space to the latent space (in the paper often referred to as pseudo-observation space). While the filter algorithm is used to estimate the marginal log likelihood ov... |
Federated learning based on client-side uploading of spiking neural networks (SNNs) is considered. Each client's SNN model parameters are obtained by converting locally updated regular NNs (called ANNs here). The server obtains and downloads the aggregated global ANN parameters from the SNN parameters uploaded from the... | Recommendation: 6: marginally above the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
Federated learning based on client-side uploading of spiking neural networks (SNNs) is considered. Each client's SNN model parameters are obtained by converting locally updated regular NNs (called ANNs here). The server obtains and downloads the aggregated global ANN parameters from the SNN parameters uploaded ... |
In their manuscript entitled, "Deconstructing distributions: a pointwise framework of learning", the authors present an approach to identifying a certain class of outliers in training sets: those points for which improvements to the global model accuracy correspond to degradation in individual accuracy.
Perhaps the g... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
In their manuscript entitled, "Deconstructing distributions: a pointwise framework of learning", the authors present an approach to identifying a certain class of outliers in training sets: those points for which improvements to the global model accuracy correspond to degradation in individual accuracy.
Perha... |
This paper studies train-free metrics for NAS.
The performance of existing train-free metrics is often inconsistent across different search spaces, less consistent than even simply counting parameters.
This paper presents a new train-free metric based on a theoretical analysis on the correlation of gradient statistic... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper studies train-free metrics for NAS.
The performance of existing train-free metrics is often inconsistent across different search spaces, less consistent than even simply counting parameters.
This paper presents a new train-free metric based on a theoretical analysis on the correlation of gradient s... |
The authors revisit gradient-based OOD detection from the perspective of backpropagation and extend the decomposition G(x)=UV of GradNorm to more loss functions. Here G is a gradient-based detection score, V is the feature norm and U represents output information. According to this decomposition, the authors suggest ex... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
The authors revisit gradient-based OOD detection from the perspective of backpropagation and extend the decomposition G(x)=UV of GradNorm to more loss functions. Here G is a gradient-based detection score, V is the feature norm and U represents output information. According to this decomposition, the authors su... |
The paper proposes a method based on a binary neural network to learn classification rules from sequential data. The model contains a stackedOR layer for categorical variables, a logical AND layer and an OR layer. The training strategy includes dynamically-induced sparsity.
Strengths. The paper introduces a method t... | Recommendation: 6: marginally above the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper proposes a method based on a binary neural network to learn classification rules from sequential data. The model contains a stackedOR layer for categorical variables, a logical AND layer and an OR layer. The training strategy includes dynamically-induced sparsity.
Strengths. The paper introduces a ... |
This paper makes two contributions to reduce overfitting of video recognition models: Ghost Motion (GM) data augmentation which randomly shifts one RGB channel forward or backward in time, and logit smoothing (temperature scaling) to reduce overconfidence on background frames. The authors hypothesize that GM works by e... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper makes two contributions to reduce overfitting of video recognition models: Ghost Motion (GM) data augmentation which randomly shifts one RGB channel forward or backward in time, and logit smoothing (temperature scaling) to reduce overconfidence on background frames. The authors hypothesize that GM wo... |
Driven by the success of data augmentation in improving the transferability of adversarial samples, this paper conducts an empirical investigation about the best combination of different data augmentations towards boosting the adversarial transferability. Experiments across ten data augmentations lead to a new composit... | Recommendation: 3: reject, not good enough | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
Driven by the success of data augmentation in improving the transferability of adversarial samples, this paper conducts an empirical investigation about the best combination of different data augmentations towards boosting the adversarial transferability. Experiments across ten data augmentations lead to a new ... |
The paper proposes a learning based method for subgraph matching, using a graph neural network-based encoder-decoder architecture, and a reinforcement learning strategy. The GNN consists of two main modules: intra-graph message passing for modeling graph structures, and inter-graph message passing for query-graph match... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper proposes a learning based method for subgraph matching, using a graph neural network-based encoder-decoder architecture, and a reinforcement learning strategy. The GNN consists of two main modules: intra-graph message passing for modeling graph structures, and inter-graph message passing for query-gra... |
The paper is a combination of Latent Outlier Exposure (Qiu 2022) and an active learning strategy from the same group. The idea is to exploit knowledge about the outlier rate to control active learning sample selection during training. The assumption is that finding anomalies in the (unlabeled) training set will improve... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
The paper is a combination of Latent Outlier Exposure (Qiu 2022) and an active learning strategy from the same group. The idea is to exploit knowledge about the outlier rate to control active learning sample selection during training. The assumption is that finding anomalies in the (unlabeled) training set will... |
This paper aims to address the multi-domain long-tailed recognition problem.
(1) An augmentation strategy is proposed for this goal:
I. For each time, uniformly sample an image across all classes and uniformly sample an image across all domains.
II. Each image representation is disentangled by class-... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper aims to address the multi-domain long-tailed recognition problem.
(1) An augmentation strategy is proposed for this goal:
I. For each time, uniformly sample an image across all classes and uniformly sample an image across all domains.
II. Each image representation is disentangled b... |
This paper combines recent multi-agent RL approaches for finding stackelberg equilibria into one framework, with conditions on convergence. Also, ideas from meta-RL are used to improve the follower.
This paper provides a nice overview of the existing literature for solving sequential Stackelberg games. I especially l... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper combines recent multi-agent RL approaches for finding stackelberg equilibria into one framework, with conditions on convergence. Also, ideas from meta-RL are used to improve the follower.
This paper provides a nice overview of the existing literature for solving sequential Stackelberg games. I espe... |
The authors consider the problem of strategic classification on graphs under a simplified linear graph model. The authors consider an iterative strategy for nodes changing features. An approximation learning algorithm is proposed to learn a robust classifier. The authors experiment on both synthetic 1d case and three r... | Recommendation: 3: reject, not good enough | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The authors consider the problem of strategic classification on graphs under a simplified linear graph model. The authors consider an iterative strategy for nodes changing features. An approximation learning algorithm is proposed to learn a robust classifier. The authors experiment on both synthetic 1d case and... |
The authors propose an augmented lagrangian approach to various offline decision making learning problems based on marginal importance sampling using function approximation. The authors show that their method exhibits improved qualities over behavior regularization, such as improved (sub)optimality. The present 4 varia... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The authors propose an augmented lagrangian approach to various offline decision making learning problems based on marginal importance sampling using function approximation. The authors show that their method exhibits improved qualities over behavior regularization, such as improved (sub)optimality. The present... |
This paper proposes a method for imitation learning (IL) when the expert and agent operate from different observation spaces. The paper proposes a new algorithm for dealing with this setting, which relies on two components: an importance weighting correction which modifies the standard GAIL objective, and a rejection m... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes a method for imitation learning (IL) when the expert and agent operate from different observation spaces. The paper proposes a new algorithm for dealing with this setting, which relies on two components: an importance weighting correction which modifies the standard GAIL objective, and a rej... |
The paper starts with pointing out that the prevalent perspective regarding modeling heterogeneous FL clients by looking at the label distribution is limiting, since, in practice, the data distribution may differ between clients, too. The authors propose to mitigate this concern using a variant of an iterative domain t... | Recommendation: 8: accept, good paper | Area: Generative models | Review:
The paper starts with pointing out that the prevalent perspective regarding modeling heterogeneous FL clients by looking at the label distribution is limiting, since, in practice, the data distribution may differ between clients, too. The authors propose to mitigate this concern using a variant of an iterative ... |
This paper studies the problem of investigating the depth of GNN. The authors propose a novel and interesting method to extend the depth of GNNs from a positive integer to a real value. It is claimed that negative depth enables high-pass frequency filtering functionality for graph heterophily while positive value enabl... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper studies the problem of investigating the depth of GNN. The authors propose a novel and interesting method to extend the depth of GNNs from a positive integer to a real value. It is claimed that negative depth enables high-pass frequency filtering functionality for graph heterophily while positive val... |
This work studies the problem of formulating several (non)linear two-layer neural networks (NNs) as equivalent convex optimization problems, where existing results of equivalent convex optimization problems are computationally expensive.
This work proposes to use Burer-Monteiro (BM) factorization to obtain equivalent ... | Recommendation: 6: marginally above the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
This work studies the problem of formulating several (non)linear two-layer neural networks (NNs) as equivalent convex optimization problems, where existing results of equivalent convex optimization problems are computationally expensive.
This work proposes to use Burer-Monteiro (BM) factorization to obtain equ... |
The authors propose to modify the LTH procedure by using a score that measures the layerwise importance in order to select the weights to be pruned. The authors propose a few possible scores and evaluate their performance on a set of experiments. The experiments show that in most of the cases the proposed method does n... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The authors propose to modify the LTH procedure by using a score that measures the layerwise importance in order to select the weights to be pruned. The authors propose a few possible scores and evaluate their performance on a set of experiments. The experiments show that in most of the cases the proposed metho... |
The paper presents a new approach called "Scrunch" to perform privacy-preserving training in MLaaS settings. The key idea is to split the model architecture into two parts (one for client called Encoder and another for the server called Classifier) and optimize both parts to maintain predictive performance while explic... | Recommendation: 3: reject, not good enough | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper presents a new approach called "Scrunch" to perform privacy-preserving training in MLaaS settings. The key idea is to split the model architecture into two parts (one for client called Encoder and another for the server called Classifier) and optimize both parts to maintain predictive performance whil... |
This paper proposes a more general setting for person re-identification, i.e., domain generalizable person re-identification without demographics (DGWD-ReID). Based on the distributionally robust optimization (DRO) method for addressing the underlying uncertainty of domain distribution, it proposes a variant of DRO, na... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a more general setting for person re-identification, i.e., domain generalizable person re-identification without demographics (DGWD-ReID). Based on the distributionally robust optimization (DRO) method for addressing the underlying uncertainty of domain distribution, it proposes a variant of... |
To address the information extraction (IE) setting for multiple tasks, the authors propose to formulate the IE problem as token-pair classification. The proposed formulation represents the spans and span-pair relations through token-pair interactions which is similar to table-filling approaches. The experiment results ... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
To address the information extraction (IE) setting for multiple tasks, the authors propose to formulate the IE problem as token-pair classification. The proposed formulation represents the spans and span-pair relations through token-pair interactions which is similar to table-filling approaches. The experiment ... |
The authors study the evolution under gradient flow of the training and test errors in the setting of Gaussian covariate models, which encompasses several other well studied settings. Using tools from random matrix theory, in particular the linear pencil, they calculate the exact asymptotic behavior of several quantiti... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The authors study the evolution under gradient flow of the training and test errors in the setting of Gaussian covariate models, which encompasses several other well studied settings. Using tools from random matrix theory, in particular the linear pencil, they calculate the exact asymptotic behavior of several ... |
In this paper, the authors proposed GPTQ, a practical method to perform low-bit weight quantization of large Generative Pre-trained (GPT) models in a post-training manner. GPTQ adapted from existing Hessian-based quantization method and performed modifications to improve efficiency by using a fixed per-row ordering, a ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
In this paper, the authors proposed GPTQ, a practical method to perform low-bit weight quantization of large Generative Pre-trained (GPT) models in a post-training manner. GPTQ adapted from existing Hessian-based quantization method and performed modifications to improve efficiency by using a fixed per-row orde... |
This paper proposes a fast approximation of the empirical neural tangent kernel. The authors theoretically prove the approximation accuracy of the proposed method, and use experiments to demonstrate its performance.
Strengths:
Although I did not thoroughly check all the proofs, the proposed methods and the theory look... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a fast approximation of the empirical neural tangent kernel. The authors theoretically prove the approximation accuracy of the proposed method, and use experiments to demonstrate its performance.
Strengths:
Although I did not thoroughly check all the proofs, the proposed methods and the the... |
This paper addresses an extended problem of open-set recognition (OSR), called Unified Open-set Recognition (UOSR). Unlike OSR, which only rejects testing samples from classes that the model has not seen during training, UOSR aims at rejecting both samples from unseen classes and also samples from seen classes but wron... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper addresses an extended problem of open-set recognition (OSR), called Unified Open-set Recognition (UOSR). Unlike OSR, which only rejects testing samples from classes that the model has not seen during training, UOSR aims at rejecting both samples from unseen classes and also samples from seen classes ... |
This paper focuses on class-incremental with repetition (CIR), which ranges from class- and domain-incremental learning. For building streams and managing repetition over time, the authors propose two CIR generators (i.e., slot-based and sampling-based), which are from two aspects to generate data streams. Moreover, th... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper focuses on class-incremental with repetition (CIR), which ranges from class- and domain-incremental learning. For building streams and managing repetition over time, the authors propose two CIR generators (i.e., slot-based and sampling-based), which are from two aspects to generate data streams. More... |
The authors propose GraphCG, a method for finding steerable factors in the latent space of already-learned deep generative model (DGM) for graphs. The method learns an energy-based model (EBM) in the latent space of a DGM with the goal of finding direction vectors $d_i$ that correspond to meaningful semantic modificati... | Recommendation: 5: marginally below the acceptance threshold | Area: Generative models | Review:
The authors propose GraphCG, a method for finding steerable factors in the latent space of already-learned deep generative model (DGM) for graphs. The method learns an energy-based model (EBM) in the latent space of a DGM with the goal of finding direction vectors $d_i$ that correspond to meaningful semantic mo... |
In order to fit molecular property prediction models that give good uncertainty estimates but also overcome issues associated with neural network methods that either overfit or underfit Gaussian Processes, the authors propose Adaptive Deep Kernel Fitting with Implicit Function Theorem (ADKF-IFT), a generalization of De... | Recommendation: 8: accept, good paper | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
In order to fit molecular property prediction models that give good uncertainty estimates but also overcome issues associated with neural network methods that either overfit or underfit Gaussian Processes, the authors propose Adaptive Deep Kernel Fitting with Implicit Function Theorem (ADKF-IFT), a generalizati... |
The paper proposes a fair learning objective for federated settings via Bounded Group Loss. The authors propose a scalable federated solver to find an approximate saddle point for the objective. Theoretically, they provide convergence and fairness guarantees for the method. Empirically, they show that their method can ... | Recommendation: 6: marginally above the acceptance threshold | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
The paper proposes a fair learning objective for federated settings via Bounded Group Loss. The authors propose a scalable federated solver to find an approximate saddle point for the objective. Theoretically, they provide convergence and fairness guarantees for the method. Empirically, they show that their met... |
The paper proposes a method to learn and generate "antidote data", which are comparable samples, to resist individual unfairness at minimal cost to model predictivity. The antidote data is generated using a GAN type generator. The paper discusses applications to various datasets and observes the effect of antidote data... | 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 method to learn and generate "antidote data", which are comparable samples, to resist individual unfairness at minimal cost to model predictivity. The antidote data is generated using a GAN type generator. The paper discusses applications to various datasets and observes the effect of antid... |
This paper focuses on the issue of requiring to access a complete and correct symbolic model when combining symbolic methods with reinforcement learning (RL). To address the issue, this work proposes a new method for learning optimistic symbolic approximations of the underlying model in an online fashion. Besides, theo... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper focuses on the issue of requiring to access a complete and correct symbolic model when combining symbolic methods with reinforcement learning (RL). To address the issue, this work proposes a new method for learning optimistic symbolic approximations of the underlying model in an online fashion. Besid... |
This paper presents a simple yet effective strategy to better balance different languages when pretraining a massively multilingual language model (such as mBERT, mT5, XLM-R, or XLM-E). The idea is quite straightforward but nicely motivated and convincingly executed: instead of the typically used temperature-based samp... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper presents a simple yet effective strategy to better balance different languages when pretraining a massively multilingual language model (such as mBERT, mT5, XLM-R, or XLM-E). The idea is quite straightforward but nicely motivated and convincingly executed: instead of the typically used temperature-ba... |
This paper studies zero-shot image classification with class hierarchy.
A Classification with Hierarchical Label sets (CHiLS) model is proposed to improved classical CLIP model.
This model leverages predefined subclasses, and perform CLIP on them first to obtain a set of class embeddings.
These subclass embeddings are ... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper studies zero-shot image classification with class hierarchy.
A Classification with Hierarchical Label sets (CHiLS) model is proposed to improved classical CLIP model.
This model leverages predefined subclasses, and perform CLIP on them first to obtain a set of class embeddings.
These subclass embeddi... |
This paper challenges the defense ability of adversarial training against clean-label availability poisoning attacks. It was believed that these attacks can hardly harm adversarially trained models. However, the proposed attack method in the paper substantially degrades the test accuracy of adversarially trained models... | Recommendation: 8: accept, good paper | Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics) | Review:
This paper challenges the defense ability of adversarial training against clean-label availability poisoning attacks. It was believed that these attacks can hardly harm adversarially trained models. However, the proposed attack method in the paper substantially degrades the test accuracy of adversarially traine... |
The paper studies the effect of label noise on the test error of overparameterized models. In particular, it shows that when the labels are sufficiently noisy, the test error increases again with model size (after the double descent peak). The authors show this theoretically in a simple linear random features model, an... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
The paper studies the effect of label noise on the test error of overparameterized models. In particular, it shows that when the labels are sufficiently noisy, the test error increases again with model size (after the double descent peak). The authors show this theoretically in a simple linear random features m... |
This paper points out that using softmax does not take the intra-class and inter-class feature variation into account, and this paper aims to interpret the high dimensional feature output that lies in a union of linear subspaces. In classification, each feature representation falls into one of the subspaces (K classes)... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper points out that using softmax does not take the intra-class and inter-class feature variation into account, and this paper aims to interpret the high dimensional feature output that lies in a union of linear subspaces. In classification, each feature representation falls into one of the subspaces (K ... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.