Subset (19)

iclr23 · 14k rows

Split (3)

train · 8.4k rows

Subsets and Splits

No community queries yet

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

review stringlengths 5 16.6k	score stringclasses 5 values	area stringclasses 12 values	text stringlengths 31 5.65k
This paper studied the effect of backdoor attacks in the multi-agent attacker scenario, where each attacker constructs backdoor attacks on it's individual dataset. There is a single defender learner who receives poisoned datasets from all attackers, and learn a model on that dataset. The paper discovered a backfire ef...	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 studied the effect of backdoor attacks in the multi-agent attacker scenario, where each attacker constructs backdoor attacks on it's individual dataset. There is a single defender learner who receives poisoned datasets from all attackers, and learn a model on that dataset. The paper discovered a bac...
The authors study the effectiveness of an iterative stochastic minimization algorithm for SAM (sharpness aware minimization): at each iteration the batch is further reduced in size using a Top-K rule based on the value of the loss. The goal of this choice is to reduce the computational expense, taking the cost closer t...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The authors study the effectiveness of an iterative stochastic minimization algorithm for SAM (sharpness aware minimization): at each iteration the batch is further reduced in size using a Top-K rule based on the value of the loss. The goal of this choice is to reduce the computational expense, taking the cost ...
This paper presents a unified framework for vision, language and multi-modal tasks. Instead of directly predicting labels with different formats, a unified task representation is designed to enable joint training of multiple tasks. The method exhibits state-of-the-art performance on the GRIT benchmarks and comparable r...	Recommendation: 8: accept, good paper	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper presents a unified framework for vision, language and multi-modal tasks. Instead of directly predicting labels with different formats, a unified task representation is designed to enable joint training of multiple tasks. The method exhibits state-of-the-art performance on the GRIT benchmarks and comp...
Similar to MIMO in wireless communication, multi-input multi-output training improves network performance by optimizing multiple subnetworks simultaneously. Previous MIMO network architecture uses CNN. This paper proposes MixViT, the first MIMO framework for vision transformers which takes advantage of ViTs’ inherent m...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: Similar to MIMO in wireless communication, multi-input multi-output training improves network performance by optimizing multiple subnetworks simultaneously. Previous MIMO network architecture uses CNN. This paper proposes MixViT, the first MIMO framework for vision transformers which takes advantage of ViTs’ in...
This paper introduces blind training to preserve the data and model privacy via shuffled Transformers. An intriguing finding is proposed that, inputs and the model weights of the Transformer encoder blocks, the backbone of Transformer, can be shuffled without degrading the model performance. strength: This work focuse...	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 introduces blind training to preserve the data and model privacy via shuffled Transformers. An intriguing finding is proposed that, inputs and the model weights of the Transformer encoder blocks, the backbone of Transformer, can be shuffled without degrading the model performance. strength: This wor...
In causal discovery, one starts with input variables and then learns a causal graph describing the relationships between the input variables. In recent years, a complementary question, which asks "how to learn causal variables from low-level raw data (e.g., images)?" has gained attention and this task is often termed "...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: In causal discovery, one starts with input variables and then learns a causal graph describing the relationships between the input variables. In recent years, a complementary question, which asks "how to learn causal variables from low-level raw data (e.g., images)?" has gained attention and this task is often ...
This paper contains two main contributions. Firstly, the authors have collected a new dataset for fitness activity recognition in a controlled environment (the setting of the videos is controlled by the authors). The dataset contains 5511 videos with 40 fine-grained activity labels. They then additionally collected a m...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper contains two main contributions. Firstly, the authors have collected a new dataset for fitness activity recognition in a controlled environment (the setting of the videos is controlled by the authors). The dataset contains 5511 videos with 40 fine-grained activity labels. They then additionally colle...
The authors propose a GAN based approach to generate attacks and also claim to enhance the detection of illicit activity in various domains. They use three type of loss functions to train a generator to generate the attacks. The domain they focus on is money laundering and recommendation systems. The authors claim ther...	Recommendation: 3: reject, not good enough	Area: Generative models	Review: The authors propose a GAN based approach to generate attacks and also claim to enhance the detection of illicit activity in various domains. They use three type of loss functions to train a generator to generate the attacks. The domain they focus on is money laundering and recommendation systems. The authors cl...
This paper analyzes the "reliability" of large language models (LLM) with prompt engineering or in-context learning. Authors consider four "facets" of reliability: (1) robustness to domain shift, (2) neutrality w.r.t. social biases (3) uncertainty calibration and (4) ability to update the learned knowledge. For each f...	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 analyzes the "reliability" of large language models (LLM) with prompt engineering or in-context learning. Authors consider four "facets" of reliability: (1) robustness to domain shift, (2) neutrality w.r.t. social biases (3) uncertainty calibration and (4) ability to update the learned knowledge. Fo...
This paper focuses on the problem of reliable explanations. The authors find that existing methods can be inconsistent or unstable. Therefore, there is an impending need to quantify the uncertainty of such explanation methods in order to understand when explanations are trustworthy. The authors introduce a novel uncert...	Recommendation: 5: marginally below the acceptance threshold	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: This paper focuses on the problem of reliable explanations. The authors find that existing methods can be inconsistent or unstable. Therefore, there is an impending need to quantify the uncertainty of such explanation methods in order to understand when explanations are trustworthy. The authors introduce a nove...
This work addresses VLM few-shot learning with prompt tuning. It addressed the overfitting of the learned prompts by regularizing the direction of gradients. The direction is enforced not in the opposite direction of the gradient computed with the human-selected prompt. The paper provides a theoretical justification to...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This work addresses VLM few-shot learning with prompt tuning. It addressed the overfitting of the learned prompts by regularizing the direction of gradients. The direction is enforced not in the opposite direction of the gradient computed with the human-selected prompt. The paper provides a theoretical justific...
The paper introduces a model-based RL approach that learns a transformer-based world model and additional techniques to further improve the sample-efficiency. Specifically, VAE is trained to learn visual information from raw pixels, and autoregressive dynamics model based on Transformer-XL architecture is trained upon ...	Recommendation: 6: marginally above the acceptance threshold	Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)	Review: The paper introduces a model-based RL approach that learns a transformer-based world model and additional techniques to further improve the sample-efficiency. Specifically, VAE is trained to learn visual information from raw pixels, and autoregressive dynamics model based on Transformer-XL architecture is train...
To tackle the distribution shift and missing data problem of time series, this paper proposes a latent space spectral decomposition method for simultaneous time series forecasting and imputation. The latent vector is optimized individually for unseen data and thus can generalize well to unseen data distribution. Pros:...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: To tackle the distribution shift and missing data problem of time series, this paper proposes a latent space spectral decomposition method for simultaneous time series forecasting and imputation. The latent vector is optimized individually for unseen data and thus can generalize well to unseen data distribution...
This paper proposes a variant of masked autoencoders (MAE), which is specifically designed for certified robustness tasks. Specifically, they add Gaussian noise in the pretaining process and use the consistency regularization method for finetuning. With a much smaller computational complexity, they obtain better result...	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 variant of masked autoencoders (MAE), which is specifically designed for certified robustness tasks. Specifically, they add Gaussian noise in the pretaining process and use the consistency regularization method for finetuning. With a much smaller computational complexity, they obtain bette...
The authors propose a simple modification of scheduled sampling for transformers that allows them to get better performance on a variety of tasks, including NMT and text generation. Importantly, they analyze why exactly their method does better and produce pretty conclusive quantitative results of this. This paper does...	Recommendation: 3: reject, not good enough	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The authors propose a simple modification of scheduled sampling for transformers that allows them to get better performance on a variety of tasks, including NMT and text generation. Importantly, they analyze why exactly their method does better and produce pretty conclusive quantitative results of this. This pa...
The paper proposes an unsupervised network (graph) embedding technique. Rather than using unstructured vectors to represent nodes, the embedding framework uses sums of Kronecker products of learnable vectors to represent the nodes. This is inspired by entanglement in quantum physics where similar representations are us...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The paper proposes an unsupervised network (graph) embedding technique. Rather than using unstructured vectors to represent nodes, the embedding framework uses sums of Kronecker products of learnable vectors to represent the nodes. This is inspired by entanglement in quantum physics where similar representation...
The paper proposes a curriculum learning (CL) method to train BERT, which is treated as blackbox (no need to modify the original implementation). The CL method consists of several stages, each of which is to replace rare words / phrases with syntactical tags (e.g. 1000 with CD). By doing that, BERT will only need to le...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: The paper proposes a curriculum learning (CL) method to train BERT, which is treated as blackbox (no need to modify the original implementation). The CL method consists of several stages, each of which is to replace rare words / phrases with syntactical tags (e.g. 1000 with CD). By doing that, BERT will only ne...
The paper proposed to utilize intermediate checkpoints from a single training process to protect data from unauthorized use. It also proposed a novel feature alignment (FA) technique that improves the accuracy of its proposed self-ensemble protection (SEP) method. FA uses an existing theory called neural collapse to al...	Recommendation: 6: marginally above the acceptance threshold	Area: Deep Learning and representational learning	Review: The paper proposed to utilize intermediate checkpoints from a single training process to protect data from unauthorized use. It also proposed a novel feature alignment (FA) technique that improves the accuracy of its proposed self-ensemble protection (SEP) method. FA uses an existing theory called neural collap...
Deep learning models often exhibit consistent error patterns, where these errors often correspond to hard subpopulations in the data they are deployed on. This paper introduces a framework for automatic distillation and surfacing of a model’s error patterns. In detail, this framework uses SVM to predict the error of th...	Recommendation: 8: accept, good paper	Area: Deep Learning and representational learning	Review: Deep learning models often exhibit consistent error patterns, where these errors often correspond to hard subpopulations in the data they are deployed on. This paper introduces a framework for automatic distillation and surfacing of a model’s error patterns. In detail, this framework uses SVM to predict the err...
Paper tackles the task of Class Agnostic Counting (CAC) in a few-shot setting, which involves predicting the overall count for the object of interest in an image, given few exemplars of object of interest from the same image. Previous works on CAC such as Ranjan et al, FamNet, rely on correlation between the exemplar f...	Recommendation: 5: marginally below the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: Paper tackles the task of Class Agnostic Counting (CAC) in a few-shot setting, which involves predicting the overall count for the object of interest in an image, given few exemplars of object of interest from the same image. Previous works on CAC such as Ranjan et al, FamNet, rely on correlation between the ex...
This submission proposes a new imputation algorithm called Markov-Blanket Miss-Forest (MBMF), which involves two phases: Markov Blanket-based Feature Selection (MBFS), and MissForest (MF) imputation (Stekhoven & Buhlmann, 2012). For each partially observed variable V_i, MBFS finds its Markov Blanket within some m-gra...	Recommendation: 8: accept, good paper	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: This submission proposes a new imputation algorithm called Markov-Blanket Miss-Forest (MBMF), which involves two phases: Markov Blanket-based Feature Selection (MBFS), and MissForest (MF) imputation (Stekhoven & Buhlmann, 2012). For each partially observed variable V_i, MBFS finds its Markov Blanket within so...
This paper provides an interesting approach for generating link embeddings in temporal graphs, which is converting the graph into a line graph and connecting edges (now nodes), based on the temporal difference they have in the initial graph. They show that with this representation, the adjacency matrix representation i...	Recommendation: 5: marginally below the acceptance threshold	Area: General Machine Learning	Review: This paper provides an interesting approach for generating link embeddings in temporal graphs, which is converting the graph into a line graph and connecting edges (now nodes), based on the temporal difference they have in the initial graph. They show that with this representation, the adjacency matrix represen...
This paper proposes a type of neural networks that uses lateral inhibition inspired layers. Lateral inhibition in this paper refers to the effect where the contrast of nearby neuron excitation in the lateral direction is increased for recognition. Inspired by this, the paper proposes to use a Gaussian low-pass filter w...	Recommendation: 1: strong reject	Area: Deep Learning and representational learning	Review: This paper proposes a type of neural networks that uses lateral inhibition inspired layers. Lateral inhibition in this paper refers to the effect where the contrast of nearby neuron excitation in the lateral direction is increased for recognition. Inspired by this, the paper proposes to use a Gaussian low-pass ...
This paper considers the semi-supervised community detection problem. In particular, they consider the degree-corrected block model (DCBM) where each node belongs to one of K communities and has a degree parameter. Then, the edges are generated between nodes according to the degree parameter of the two nodes and commun...	Recommendation: 8: accept, good paper	Area: General Machine Learning	Review: This paper considers the semi-supervised community detection problem. In particular, they consider the degree-corrected block model (DCBM) where each node belongs to one of K communities and has a degree parameter. Then, the edges are generated between nodes according to the degree parameter of the two nodes an...
This paper proposes a novel deep learning framework named Cross-Protein Wasserstein Transformer (CPWT) to predict PPI sites through fine-grained cross-graph structural modeling, which promotes the PPI prediction from the perspective of sophisticated cross-protein structural modeling based on Wasserstein affinities. The...	Recommendation: 5: marginally below the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper proposes a novel deep learning framework named Cross-Protein Wasserstein Transformer (CPWT) to predict PPI sites through fine-grained cross-graph structural modeling, which promotes the PPI prediction from the perspective of sophisticated cross-protein structural modeling based on Wasserstein affinit...
This paper proposes a new method for dynamic large-scale multitask learning and conducts some experiments to study its effectiveness. Strength 1. This paper aims to provide a new and automatic method for dynamic large-scale multitask learning. Weaknesses 1. The main contribution of this paper is unclear. Although it c...	Recommendation: 1: strong reject	Area: Deep Learning and representational learning	Review: This paper proposes a new method for dynamic large-scale multitask learning and conducts some experiments to study its effectiveness. Strength 1. This paper aims to provide a new and automatic method for dynamic large-scale multitask learning. Weaknesses 1. The main contribution of this paper is unclear. Altho...
The paper aims to understand the training dynamics of the Attention mechanism through lexical prob $\beta$ and its learning proxy model. Strength: 1. The paper did some rigorous analysis of the attention training mechanism and drew a connection to the word alignment of the classical models. 2. There is a connection to...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: The paper aims to understand the training dynamics of the Attention mechanism through lexical prob $\beta$ and its learning proxy model. Strength: 1. The paper did some rigorous analysis of the attention training mechanism and drew a connection to the word alignment of the classical models. 2. There is a conne...
This paper proposes an explainable abstractive summarization framework by exploring the two research questions: RQ1 and RQ2. Experiment results on CNN/DailyMail dataset are given to demonstrate the effectiveness of the proposed SP framework. It is nice to propose a novel framework to explore the interpretability of abs...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes an explainable abstractive summarization framework by exploring the two research questions: RQ1 and RQ2. Experiment results on CNN/DailyMail dataset are given to demonstrate the effectiveness of the proposed SP framework. It is nice to propose a novel framework to explore the interpretabilit...
This paper presents a dataset containing synthetic models of power grids combined with the statistical results of dynamic simulations (quantifying single-node basin stability and survivability - abbreviated SNBS and SURV). The authors train a baseline GNN to show initial efficacy of the task of predicting SNBS and SURV...	Recommendation: 6: marginally above the acceptance threshold	Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )	Review: This paper presents a dataset containing synthetic models of power grids combined with the statistical results of dynamic simulations (quantifying single-node basin stability and survivability - abbreviated SNBS and SURV). The authors train a baseline GNN to show initial efficacy of the task of predicting SNBS ...
This paper presents an algorithm to accurately account for individual differential privacy guarantees in DPSGD. Strength: Compared with worst-case DP guarantees, the individual DP guarantee of a sample allows us to better understand the privacy guarantee of a specific data sample, and its impact over model parameters....	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 presents an algorithm to accurately account for individual differential privacy guarantees in DPSGD. Strength: Compared with worst-case DP guarantees, the individual DP guarantee of a sample allows us to better understand the privacy guarantee of a specific data sample, and its impact over model par...
While many techniques for model fairness have been proposed, the majority of them assume that the distributions of training and deployment data are identical, which is often not the case in practice. In particular, the bias between labels and sensitive groups changes, which may impair the performance of machine learnin...	Recommendation: 8: accept, good paper	Area: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)	Review: While many techniques for model fairness have been proposed, the majority of them assume that the distributions of training and deployment data are identical, which is often not the case in practice. In particular, the bias between labels and sensitive groups changes, which may impair the performance of machine...
The authors consider the problem of learning the intensity function of an inhomogeneous Poisson process (PP) over higher dimensional spaces: 1. The authors propose a certain family of generalized additive models (GAM) to model the log-intensity function of the PP and call their method the additive Poisson process (APP...	Recommendation: 3: reject, not good enough	Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)	Review: The authors consider the problem of learning the intensity function of an inhomogeneous Poisson process (PP) over higher dimensional spaces: 1. The authors propose a certain family of generalized additive models (GAM) to model the log-intensity function of the PP and call their method the additive Poisson proc...
This work tries to explain the asymptotic behaviors of two well-known approximate RL algorithms, Q-learning and SARSA, with $\epsilon$-greedy exploration. The theoretical analysis presented in this work uses Differential Inclusion (DI) theory to analyze value-based RL methods with discontinuous behavior policy changes....	Recommendation: 5: marginally below the acceptance threshold	Area: Theory (eg, control theory, learning theory, algorithmic game theory)	Review: This work tries to explain the asymptotic behaviors of two well-known approximate RL algorithms, Q-learning and SARSA, with $\epsilon$-greedy exploration. The theoretical analysis presented in this work uses Differential Inclusion (DI) theory to analyze value-based RL methods with discontinuous behavior policy ...
In this paper, the authors introduced a framework for explaining distribution shift using a transport map T between a source and target distribution. They constrained a relaxed form of optimal transport to theoretically define an interpretable mapping TIT and introduced two interpretable transport methods: k-sparse and...	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 introduced a framework for explaining distribution shift using a transport map T between a source and target distribution. They constrained a relaxed form of optimal transport to theoretically define an interpretable mapping TIT and introduced two interpretable transport methods: k-sp...
This paper observes that AT [1] focuses on the low-frequency area to achieving adversarial robustness. It empirically shows that white-box attack corrupts the high frequency domain. Authors propose Frequency Regularization (FR) which applies a regularization in the frequency domain with AT. Strength - This paper des...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper observes that AT [1] focuses on the low-frequency area to achieving adversarial robustness. It empirically shows that white-box attack corrupts the high frequency domain. Authors propose Frequency Regularization (FR) which applies a regularization in the frequency domain with AT. Strength - This p...
This paper attempts to propose an evaluation protocol for lightweight probing of unsupervised representations and investigates the correlation between RL performance and linear probing from a pretrained representation. Authors are testing this on a very specific class of self-predictive (recurrent) representation model...	Recommendation: 3: reject, not good enough	Area: Deep Learning and representational learning	Review: This paper attempts to propose an evaluation protocol for lightweight probing of unsupervised representations and investigates the correlation between RL performance and linear probing from a pretrained representation. Authors are testing this on a very specific class of self-predictive (recurrent) representati...
This paper focuses on the self-supervised learning. Based on masked image modeling, the authors introduce a new algorithm which uses corrupted as training sources instead of masked ones. They utilize a new trainable module including a pretrained transformer decoder to generate such corrupted images. The models are trai...	Recommendation: 6: marginally above the acceptance threshold	Area: Unsupervised and Self-supervised learning	Review: This paper focuses on the self-supervised learning. Based on masked image modeling, the authors introduce a new algorithm which uses corrupted as training sources instead of masked ones. They utilize a new trainable module including a pretrained transformer decoder to generate such corrupted images. The models ...
This paper proposes a new approach to multitask learning by retrieving relevant training examples from P3 and train the model on it. ### Strength 1. This paper proposes a simple yet effective idea on how to use existing resources (i.e., P3) to facilitate unseen tasks. 2. The paper is well-written and easy to follow. 3...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: This paper proposes a new approach to multitask learning by retrieving relevant training examples from P3 and train the model on it. ### Strength 1. This paper proposes a simple yet effective idea on how to use existing resources (i.e., P3) to facilitate unseen tasks. 2. The paper is well-written and easy to f...
Most of the paper is clearly written. A new dataset for incremental learning is provided. The fixed and dynamic modes of Global Prototype Encoding is proposed and outperform the state-of-arts. Strength: Most parts are explained clearly for the equations.Most of the procedures seems reasonable. Weakness: 1) How to le...	Recommendation: 6: marginally above the acceptance threshold	Area: Applications (eg, speech processing, computer vision, NLP)	Review: Most of the paper is clearly written. A new dataset for incremental learning is provided. The fixed and dynamic modes of Global Prototype Encoding is proposed and outperform the state-of-arts. Strength: Most parts are explained clearly for the equations.Most of the procedures seems reasonable. Weakness: 1) H...
In this paper, the authors focus on the multi-modal representation learning and cross-modal similarity estimation topic and propose the Chimera framework. The proposed framework uses the hyperbolic space to better represent complex multi-modal relationships. The topic model is also utilized. The proposed Chimera method...	Recommendation: 5: marginally below the acceptance threshold	Area: Deep Learning and representational learning	Review: In this paper, the authors focus on the multi-modal representation learning and cross-modal similarity estimation topic and propose the Chimera framework. The proposed framework uses the hyperbolic space to better represent complex multi-modal relationships. The topic model is also utilized. The proposed Chimer...