review stringlengths 5 16.6k | score stringclasses 5
values | area stringclasses 12
values | text stringlengths 31 5.65k |
|---|---|---|---|
This paper explores how activate learning can help reduce annotation costs in the field of object detection of 3D point clouds. Specifically, to achieve efficient active learning with limited fixed annotation budgets, three selection criteria termed CRB are proposed to learn better sample acquisition of the 3D boxes an... | Recommendation: 8: accept, good paper | Area: General Machine Learning | Review:
This paper explores how activate learning can help reduce annotation costs in the field of object detection of 3D point clouds. Specifically, to achieve efficient active learning with limited fixed annotation budgets, three selection criteria termed CRB are proposed to learn better sample acquisition of the 3D ... |
This paper studies unsupervised word alignment using pre-trained multilingual LMs. The paper proposes an iterative method that repeats two steps: (1) generate word alignment using a LM, (2) using the word alignment, generate synthesis code-switched parallel text to fine-tune the LM. Empirically, this iterative method r... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper studies unsupervised word alignment using pre-trained multilingual LMs. The paper proposes an iterative method that repeats two steps: (1) generate word alignment using a LM, (2) using the word alignment, generate synthesis code-switched parallel text to fine-tune the LM. Empirically, this iterative ... |
The paper studies generalization bounds in federated learning with unparticipating clients. A two-level distribution framework is proposed, where the data distribution of each client is sampled according to a meta distribution, then the data samples at each client is sampled from the client's data distribution. Afterwa... | Recommendation: 6: marginally above the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
The paper studies generalization bounds in federated learning with unparticipating clients. A two-level distribution framework is proposed, where the data distribution of each client is sampled according to a meta distribution, then the data samples at each client is sampled from the client's data distribution.... |
This paper focuses on anomaly detection from time series. Specifically, this paper tackles this problem from a causal perspective. It first learns a causal structure from the data. The causal structure is assumed to be a directed acyclic graph (DAG) which captures stationary causal structure. The anomaly then can be de... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper focuses on anomaly detection from time series. Specifically, this paper tackles this problem from a causal perspective. It first learns a causal structure from the data. The causal structure is assumed to be a directed acyclic graph (DAG) which captures stationary causal structure. The anomaly then c... |
The paper presents a Robust GAN-inversion method for restoring gross corrupted images. They also extend RGI to Relaxed RGI for generator fine-tuning to mitigate the gap between the GAN learned manifold and the true image manifold. Their result achieves sota performance in two applications: mask-free semantic inpating a... | Recommendation: 6: marginally above the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
The paper presents a Robust GAN-inversion method for restoring gross corrupted images. They also extend RGI to Relaxed RGI for generator fine-tuning to mitigate the gap between the GAN learned manifold and the true image manifold. Their result achieves sota performance in two applications: mask-free semantic in... |
This paper proposes a new Graph Neural Network architecture enhancing the standard message passing architectures (MPNN). The main idea is to run various MPNNs on rooted subgraphs of the original graph. The authors show that their architecture I2-GNN is more powerful than previous MPNN. They show theoretical results whe... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a new Graph Neural Network architecture enhancing the standard message passing architectures (MPNN). The main idea is to run various MPNNs on rooted subgraphs of the original graph. The authors show that their architecture I2-GNN is more powerful than previous MPNN. They show theoretical res... |
The paper considers the semantic segmentation problem. The motivation of this paper is based on the relatively poor performance for the tailed classes and in order to overcome this issue, the author propose to use multiple experts to process pixels with different category frequency. In the experimental section, the pap... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper considers the semantic segmentation problem. The motivation of this paper is based on the relatively poor performance for the tailed classes and in order to overcome this issue, the author propose to use multiple experts to process pixels with different category frequency. In the experimental section,... |
This work proposes to use increasing stepsize in Xiao, 2022, as shown in Theorem 4.7, in natural policy gradient with log-linear policy parametrization, and the main result is that the convergence rate is linear plus approximation errors, improving the $O(1/\sqrt{T})$ rate plus approximation error terms in Agarwal et a... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This work proposes to use increasing stepsize in Xiao, 2022, as shown in Theorem 4.7, in natural policy gradient with log-linear policy parametrization, and the main result is that the convergence rate is linear plus approximation errors, improving the $O(1/\sqrt{T})$ rate plus approximation error terms in Agar... |
This paper aims to establish a sharp generalization error bound for NN with ReLU activation when the network size doesn’t grow with the training set size. This paper provides a weakened version of uniform stability and proves the generalization guarantees of SGD for training over locally Lipschitz and smooth objectives... | Recommendation: 5: marginally below the acceptance threshold | Area: Theory (eg, control theory, learning theory, algorithmic game theory) | Review:
This paper aims to establish a sharp generalization error bound for NN with ReLU activation when the network size doesn’t grow with the training set size. This paper provides a weakened version of uniform stability and proves the generalization guarantees of SGD for training over locally Lipschitz and smooth ob... |
In this paper, the authors found that using a variant of DEMA to replace the EMA used in past optimizers and using a dynamic lookahead optimizer can achieve a better result. Motivated by that, the author proposed a novel optimizer framework Admeta, and implement it on SGD and Radam. Empirical results demonstrate the su... | Recommendation: 6: marginally above the acceptance threshold | Area: Optimization (eg, convex and non-convex optimization) | Review:
In this paper, the authors found that using a variant of DEMA to replace the EMA used in past optimizers and using a dynamic lookahead optimizer can achieve a better result. Motivated by that, the author proposed a novel optimizer framework Admeta, and implement it on SGD and Radam. Empirical results demonstrat... |
This paper presents a learning from demonstration approach to generate diverse policies representing the different preferences of the expert. The authors seek to learn a GMM in the latent space, and match the state-visitation frequency from state observations only. Experiments on simulated reaching tasks with robot arm... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper presents a learning from demonstration approach to generate diverse policies representing the different preferences of the expert. The authors seek to learn a GMM in the latent space, and match the state-visitation frequency from state observations only. Experiments on simulated reaching tasks with r... |
The authors extend ICA for multi-view settings, building on similar latent variable models that have been used for modelling correlations between multiple views. This paper extends the formulation for supporting non-gaussian latent distributions to enable identification of individual sources, following the standard ICA... | Recommendation: 3: reject, not good enough | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The authors extend ICA for multi-view settings, building on similar latent variable models that have been used for modelling correlations between multiple views. This paper extends the formulation for supporting non-gaussian latent distributions to enable identification of individual sources, following the stan... |
This paper proposes a dynamic semantic prototype learning method that aligns empirical semantic prototypes with actual semantic prototypes in order to synthesize accurate visual features. The method consists of a generative model that generates visual features from semantic prototypes, a network that maps visual featur... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a dynamic semantic prototype learning method that aligns empirical semantic prototypes with actual semantic prototypes in order to synthesize accurate visual features. The method consists of a generative model that generates visual features from semantic prototypes, a network that maps visua... |
This work considers geometric graphs having both non-uniform sampling density, as well varying neighborhood radius. Under this model, a GSO can be though of as a discretization of the latent continuous Laplacian. In order for this GSO to approximate the continuous laplacian, the adjacency matrix needs to be normalized ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This work considers geometric graphs having both non-uniform sampling density, as well varying neighborhood radius. Under this model, a GSO can be though of as a discretization of the latent continuous Laplacian. In order for this GSO to approximate the continuous laplacian, the adjacency matrix needs to be nor... |
This work presents a reinforcement temporal logic rule learning
algorithm to jointly learn temporal logic rules (before-like kind of
rules) and their weights from event data. The proposed learning
algorithm alternates between a rule generator stage and a rule
evaluator stage, where a neural search policy is learned by
... | Recommendation: 6: marginally above the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
This work presents a reinforcement temporal logic rule learning
algorithm to jointly learn temporal logic rules (before-like kind of
rules) and their weights from event data. The proposed learning
algorithm alternates between a rule generator stage and a rule
evaluator stage, where a neural search policy is lea... |
The paper proposes a new method to defend against backdoor attacks. The proposed method is basically a new aggregation rule. Edge-case backdoor attacks are considered, and multiple simple baselines are evaluated.
Strengths
- FL is vulnerable to backdoor attacks
- A new aggregation rule is proposed.
Weaknesses
- M... | 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 proposes a new method to defend against backdoor attacks. The proposed method is basically a new aggregation rule. Edge-case backdoor attacks are considered, and multiple simple baselines are evaluated.
Strengths
- FL is vulnerable to backdoor attacks
- A new aggregation rule is proposed.
Weaknes... |
The authors introduce an algorithm that manipulates images using diffusion models.
They do so by manipulating the representation in the bottle-neck U-net layer of the diffusion model over several (but not all) timesteps. The loss function that they optimize is the same as for DiffusionCLIP.
### Strengths
The visual re... | Recommendation: 8: accept, good paper | Area: Generative models | Review:
The authors introduce an algorithm that manipulates images using diffusion models.
They do so by manipulating the representation in the bottle-neck U-net layer of the diffusion model over several (but not all) timesteps. The loss function that they optimize is the same as for DiffusionCLIP.
### Strengths
The v... |
This paper provides an approach for self-supervised denoising of MRI data using diffusion-based generative models. The approach is applied to denoise diffusion-weighted MRI, hence the name DDM2. A three-stage approach is used to train/denoise: 1. learning a denoising function with existing self-supervised approaches; 2... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper provides an approach for self-supervised denoising of MRI data using diffusion-based generative models. The approach is applied to denoise diffusion-weighted MRI, hence the name DDM2. A three-stage approach is used to train/denoise: 1. learning a denoising function with existing self-supervised appro... |
The idea is to learn inverariance to commutative groups. The groups are learned from data through the proposed block, which relies on the bispectrum idea. The proposed block (depicted in Fig. 2) resorts to a third-order polynomial that can be implemented on standard frameworks, as it requires an affine transformation o... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The idea is to learn inverariance to commutative groups. The groups are learned from data through the proposed block, which relies on the bispectrum idea. The proposed block (depicted in Fig. 2) resorts to a third-order polynomial that can be implemented on standard frameworks, as it requires an affine transfor... |
The goal of this paper is to improve the semantic segmentation performance through addressing the boundary-caused class weights confusion issue. The two semantic classes that share more adjacent pixels tend to have more similar class weights in the state-of-the-art semantic segmentation solution such as DeepLabv3+. The... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
The goal of this paper is to improve the semantic segmentation performance through addressing the boundary-caused class weights confusion issue. The two semantic classes that share more adjacent pixels tend to have more similar class weights in the state-of-the-art semantic segmentation solution such as DeepLab... |
This paper focuses on the label skew problem in one-shot federated learning. The author identifies the problem is attributed to the limited scale of certain class in different participants, which brings unreliable voting scores and thus results in ineffective global model. In this paper, authors propose FedOV, which le... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper focuses on the label skew problem in one-shot federated learning. The author identifies the problem is attributed to the limited scale of certain class in different participants, which brings unreliable voting scores and thus results in ineffective global model. In this paper, authors propose FedOV, ... |
The paper extends the idea of ExploreOptions to the deep RL setting. Four different sub-policies are trained off-policy on the same data: an Exploit policy, and 3 Explore policies, which each use a different intrinsic motivation (IM) method to incentivize exploration (e.g. including RND). Results are presented in a gri... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper extends the idea of ExploreOptions to the deep RL setting. Four different sub-policies are trained off-policy on the same data: an Exploit policy, and 3 Explore policies, which each use a different intrinsic motivation (IM) method to incentivize exploration (e.g. including RND). Results are presented ... |
This paper aims to understand “what makes masked language models (MLMs) useful”. The paper claims that it is because MLM pretraining is specifically effective in reducing the reliance on spurious features, and provides the theoretical view of it. Specifically, it first (1) assumes two random variables from the input, X... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper aims to understand “what makes masked language models (MLMs) useful”. The paper claims that it is because MLM pretraining is specifically effective in reducing the reliance on spurious features, and provides the theoretical view of it. Specifically, it first (1) assumes two random variables from the ... |
Empirical Risk Minimization (ERM) is an established technique and is one of the basis of supervised (plus some other types) learning. The ERM objective is to minimize the risk when fitting a set of labeled data points. In the context of Semi Supervised Learning (SSL), as we need to deal with data points with and withou... | Recommendation: 8: accept, good paper | Area: Unsupervised and Self-supervised learning | Review:
Empirical Risk Minimization (ERM) is an established technique and is one of the basis of supervised (plus some other types) learning. The ERM objective is to minimize the risk when fitting a set of labeled data points. In the context of Semi Supervised Learning (SSL), as we need to deal with data points with an... |
This paper proposes a general framework for diffusion models with different types of deterministic degradation process. The framework minimizes a training objective similar to the original diffusion models, which is about minimizing l1 norm between the original clean sample and the predicted clean sample by a restorati... | Recommendation: 5: marginally below the acceptance threshold | Area: Generative models | Review:
This paper proposes a general framework for diffusion models with different types of deterministic degradation process. The framework minimizes a training objective similar to the original diffusion models, which is about minimizing l1 norm between the original clean sample and the predicted clean sample by a r... |
This work presents a benchmark consisting of several graph datasets with different distributional shifts to evaluate the robustness and uncertainty estimation of models for node-level tasks. Furthermore, a series of experiments have been conducted to show the ID/OOD performance gap.
Strengths:
- Distribution shift p... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This work presents a benchmark consisting of several graph datasets with different distributional shifts to evaluate the robustness and uncertainty estimation of models for node-level tasks. Furthermore, a series of experiments have been conducted to show the ID/OOD performance gap.
Strengths:
- Distribution... |
This paper proposes a new module that reduces the quadratic complexity of standard Transformers. The key idea is to slice the input sequence into various groups and then perform local and global attention respectively. Experimental results indicate that the local-to-global approach can indeed accelerate the processing ... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a new module that reduces the quadratic complexity of standard Transformers. The key idea is to slice the input sequence into various groups and then perform local and global attention respectively. Experimental results indicate that the local-to-global approach can indeed accelerate the pro... |
AudioGen is a text to audio model, focusing primarily on non-speech, non-music audio events. It supports rendering multiple overlapping events simultaneously depending on the given text description and can generalize to unseen text/audio pairs by using a pre-trained language model encoder (T5).
The paper does several ... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
AudioGen is a text to audio model, focusing primarily on non-speech, non-music audio events. It supports rendering multiple overlapping events simultaneously depending on the given text description and can generalize to unseen text/audio pairs by using a pre-trained language model encoder (T5).
The paper does ... |
This work proposes to improve automated program translation from unsupervised data by leveraging compiler intermediate representation (IR), a machine-code like format that is unified across many programming languages. Rather than train a model to translate each language into and from this domain, the work argues for us... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This work proposes to improve automated program translation from unsupervised data by leveraging compiler intermediate representation (IR), a machine-code like format that is unified across many programming languages. Rather than train a model to translate each language into and from this domain, the work argue... |
This paper aims to address the problem of learning complex structures from sequences.
They improved an existing method called SyncMap by balancing the number of updates from positive and negative feedback loops.
They demonstrated the effectiveness of the proposed method on multiple synthetic and real-world datasets. ... | Recommendation: 5: marginally below the acceptance threshold | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
This paper aims to address the problem of learning complex structures from sequences.
They improved an existing method called SyncMap by balancing the number of updates from positive and negative feedback loops.
They demonstrated the effectiveness of the proposed method on multiple synthetic and real-world da... |
This paper introduces an algorithm that computes two utile quantities, that they call Q-gradient (g(x, a)) and Bellman-score (s(x, a)), given transitions from a replay buffer, .
The Q-gradient can be used to update a reward function in IRL, and the Bellman-score to predict the adaptation of a policy to a change of the ... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper introduces an algorithm that computes two utile quantities, that they call Q-gradient (g(x, a)) and Bellman-score (s(x, a)), given transitions from a replay buffer, .
The Q-gradient can be used to update a reward function in IRL, and the Bellman-score to predict the adaptation of a policy to a change... |
The paper studies the problem of phylogenetic inference. The authors proposed a structural representation method for phylogenetic inference based on learnable topological features. The learnable topological features are constructed by minimizing the Dirichlet energy. The authors use the combined node features to infer ... | Recommendation: 6: marginally above the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper studies the problem of phylogenetic inference. The authors proposed a structural representation method for phylogenetic inference based on learnable topological features. The learnable topological features are constructed by minimizing the Dirichlet energy. The authors use the combined node features t... |
This paper studies the robust memorization problem in adversarial training, and explains that why adversarial training suffers from severe over-fitting problem.
Strength:
It is an interesting idea to connect the adversarial training to some local estimate to explain over-fitting.
Weakness:
Main concern:
[1] The aut... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper studies the robust memorization problem in adversarial training, and explains that why adversarial training suffers from severe over-fitting problem.
Strength:
It is an interesting idea to connect the adversarial training to some local estimate to explain over-fitting.
Weakness:
Main concern:
[1]... |
The bulk of existing work on imitation learning focuses on settings in which the expert demonstrator and learner operate under the **same observation space**; however, in many real-world settings, this isn’t realistic. The work formalizes the problem of heterogeneously observable imitation learning (HOIL) and present a... | Recommendation: 8: accept, good paper | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The bulk of existing work on imitation learning focuses on settings in which the expert demonstrator and learner operate under the **same observation space**; however, in many real-world settings, this isn’t realistic. The work formalizes the problem of heterogeneously observable imitation learning (HOIL) and p... |
The authors present a new approach to type inference by using T5 architecture to generate types in a seq2seq fashion. Their proposed model, TypeT5 is CodeT5 fine-tuned on a subset of ManyTypes4Py named BetterTypes4Py. TypeT5 uses static analysis to compose its contextual input with a preamble, main code, users, and use... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The authors present a new approach to type inference by using T5 architecture to generate types in a seq2seq fashion. Their proposed model, TypeT5 is CodeT5 fine-tuned on a subset of ManyTypes4Py named BetterTypes4Py. TypeT5 uses static analysis to compose its contextual input with a preamble, main code, users,... |
Inspired by gradient flows as (solution curves to) differential equations that minimize an energy functional, the authors view GNNs as a gradient flow equation of a parametric energy. The authors show that in graph convolutional models (GCN), the positive/negative eigenvalues of the channel mixing matrix correspond to ... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
Inspired by gradient flows as (solution curves to) differential equations that minimize an energy functional, the authors view GNNs as a gradient flow equation of a parametric energy. The authors show that in graph convolutional models (GCN), the positive/negative eigenvalues of the channel mixing matrix corres... |
The authors suggest an approach towards directly exploiting originally noisy speech w/o the need for a clean reference.
The algorithm is presented and code will be released. A combination with "common" denoising by having a reference available is possible.
The method is simple, yet appears effective. It is well describ... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The authors suggest an approach towards directly exploiting originally noisy speech w/o the need for a clean reference.
The algorithm is presented and code will be released. A combination with "common" denoising by having a reference available is possible.
The method is simple, yet appears effective. It is well... |
The paper presents DROP, a non-iterative framework for policy optimization in offline RL, which separates the model training on offline data from deploying the model for testing/adaptation. The approach first splits the data into N subsets to learn distinct behavior policies, which are further used to learn a contextua... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper presents DROP, a non-iterative framework for policy optimization in offline RL, which separates the model training on offline data from deploying the model for testing/adaptation. The approach first splits the data into N subsets to learn distinct behavior policies, which are further used to learn a c... |
The paper presents an approach for augmenting Q-learning to take advantage of domain randomization, name DIQL. By producing observations from two different domains (here defined as visual observations with different distracting perturbations, as in the distracting control suite), DIQL can train its q-network to be inv... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper presents an approach for augmenting Q-learning to take advantage of domain randomization, name DIQL. By producing observations from two different domains (here defined as visual observations with different distracting perturbations, as in the distracting control suite), DIQL can train its q-network t... |
This paper rethinks the skip connection model as a learnable Markov chain and proposes a penal connection to convert a residual-like model to a Markov chain for more efficient training. Experiments on MLP and CV demonstrate the superiority of this plug-in method.
Pros
- This paper formulated models with skip connectio... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper rethinks the skip connection model as a learnable Markov chain and proposes a penal connection to convert a residual-like model to a Markov chain for more efficient training. Experiments on MLP and CV demonstrate the superiority of this plug-in method.
Pros
- This paper formulated models with skip c... |
The submission proposes TAMiL, a continual learning method inspired by aspects of the global workspace theory that trains task-specific attention modules to select which features from a shared representation are relevant to the current task. The approach uses replay in combination with functional regularization to avoi... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The submission proposes TAMiL, a continual learning method inspired by aspects of the global workspace theory that trains task-specific attention modules to select which features from a shared representation are relevant to the current task. The approach uses replay in combination with functional regularization... |
The paper proposes to use MPNNs to reconstruct the (time-averaged) pressure and velocity fields around an airfoil based on measurements of the pressure distribution at the airfoil's surface. Some global parameters of the flow (farfield velocity, angle of attack, turbulence intensity) are also predicted. The authors use... | Recommendation: 3: reject, not good enough | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
The paper proposes to use MPNNs to reconstruct the (time-averaged) pressure and velocity fields around an airfoil based on measurements of the pressure distribution at the airfoil's surface. Some global parameters of the flow (farfield velocity, angle of attack, turbulence intensity) are also predicted. The aut... |
The paper introduces a novel technique for detecting which model has been compromised (i.e. an adversarial example has been created targeting that specific model) from a set of deployed models that have been watermarked. Thus, each model deployed has a watermark so that some of the inputs are masked. When an attacker t... | 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 novel technique for detecting which model has been compromised (i.e. an adversarial example has been created targeting that specific model) from a set of deployed models that have been watermarked. Thus, each model deployed has a watermark so that some of the inputs are masked. When an at... |
The paper proposes a method to train neural networks with 4-bit matrix multiplications. The authors propose logarithm unbiased quantization (LUQ), which combines stochastic rounding and logarithm quantization, both of which are known to be required for extreme low-bit training. The proposed training algorithm involves ... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The paper proposes a method to train neural networks with 4-bit matrix multiplications. The authors propose logarithm unbiased quantization (LUQ), which combines stochastic rounding and logarithm quantization, both of which are known to be required for extreme low-bit training. The proposed training algorithm i... |
This paper studies how different numbers of spurious training examples (from one to thousands) affect neural nets. They measure how the predictions of test examples change when the spurious features are added. They show even a few spurious training examples can make the model outputs be affected by the spurious feature... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
This paper studies how different numbers of spurious training examples (from one to thousands) affect neural nets. They measure how the predictions of test examples change when the spurious features are added. They show even a few spurious training examples can make the model outputs be affected by the spurious... |
This paper presents a new low-rank method, called DBA, to approximate the attention mechanism used in transformer layers. It has two main differences from previous approaches:
1. DBA maps the input sequences $Q, K \in \mathbb{R}^{n \times d}$ to shorter sequences $Q', K' \in \mathbb{R}^{d_p \times d}$ via a linear tra... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper presents a new low-rank method, called DBA, to approximate the attention mechanism used in transformer layers. It has two main differences from previous approaches:
1. DBA maps the input sequences $Q, K \in \mathbb{R}^{n \times d}$ to shorter sequences $Q', K' \in \mathbb{R}^{d_p \times d}$ via a li... |
This paper proposes a new domain adaptation setting where multi-modalities are provided in source domain while only visual images are provided in unseen target domain. Specifically, the authors focus on the BEV prediction problem where perspective image is given as input and model would output Bev semantics. Assuming t... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a new domain adaptation setting where multi-modalities are provided in source domain while only visual images are provided in unseen target domain. Specifically, the authors focus on the BEV prediction problem where perspective image is given as input and model would output Bev semantics. As... |
The paper analyzes various intrinsic reward generation algorithms in the MiniGrid benchmark and presents that episodic curiosity is helpful for training while lifelong curiosity is not.
The authors compare four different curiosity algorithms (ICM, RIDE, RND, BeBold) with and without two episodic curiosity (episodic vi... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The paper analyzes various intrinsic reward generation algorithms in the MiniGrid benchmark and presents that episodic curiosity is helpful for training while lifelong curiosity is not.
The authors compare four different curiosity algorithms (ICM, RIDE, RND, BeBold) with and without two episodic curiosity (epi... |
1) The paper proposed BINDER to incorporate symbolic component into large language models, and the main benefit of the method is not requiring any fine-tuning.
2) BINDER outperforms state-of-the-art results on WIKITABLEQUESTIONS and TABFACT datasets
3) useful analysis is presented to help readers understand the propos... | Recommendation: 8: accept, good paper | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
1) The paper proposed BINDER to incorporate symbolic component into large language models, and the main benefit of the method is not requiring any fine-tuning.
2) BINDER outperforms state-of-the-art results on WIKITABLEQUESTIONS and TABFACT datasets
3) useful analysis is presented to help readers understand th... |
This paper investigates an important problem of interpreting why distributional RL outperforms conventional RL. Specifically, the author separates the action value function into the expectation part and regularization part, and attributes the superiority to the regularization part. In addition, the author proposes a ne... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper investigates an important problem of interpreting why distributional RL outperforms conventional RL. Specifically, the author separates the action value function into the expectation part and regularization part, and attributes the superiority to the regularization part. In addition, the author propo... |
This paper proposes Cognitive Continual Learner (CCL) which is composed of three modules. An explicit module that learns from the input and two implicit modules (inductive biases and semantic memories) that share indirect contextual knowledge. CCL is evaluated under a number of continual learning settings, including on... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes Cognitive Continual Learner (CCL) which is composed of three modules. An explicit module that learns from the input and two implicit modules (inductive biases and semantic memories) that share indirect contextual knowledge. CCL is evaluated under a number of continual learning settings, incl... |
The authors propose a systematic method of creating variations of a convolutional architecture by reducing the number of filters at certain groups of layers. Subsequently training and evaluating these new architectures shows that some are significantly more efficient that the initial "seed" architecture achieving simil... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The authors propose a systematic method of creating variations of a convolutional architecture by reducing the number of filters at certain groups of layers. Subsequently training and evaluating these new architectures shows that some are significantly more efficient that the initial "seed" architecture achievi... |
This paper proposes a few-shot incremental semantic segmentation methods via guided copy-paste synthesis. To achieves this, authors provide three kinds of guidance, i.e., diversity-guide, context-guide, and frequency-guide. Ablation study and comparison with baseline model show clear performance improvement with copy-... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a few-shot incremental semantic segmentation methods via guided copy-paste synthesis. To achieves this, authors provide three kinds of guidance, i.e., diversity-guide, context-guide, and frequency-guide. Ablation study and comparison with baseline model show clear performance improvement wi... |
This paper studies why and how adaptive learning methods help training of GANs. The authors postulate that the adaptive magnitudes of gradients of Adam is a key. This postulate leads them to propose, with the step-size grafting approach by Agarwal et al., two algorithms, Ada-nSGDA and nSGDA, the former combining the Ad... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper studies why and how adaptive learning methods help training of GANs. The authors postulate that the adaptive magnitudes of gradients of Adam is a key. This postulate leads them to propose, with the step-size grafting approach by Agarwal et al., two algorithms, Ada-nSGDA and nSGDA, the former combinin... |
This paper does a thorough simulations to clarify and verify the existing hypothesis about why small batch SGD outperform its large batch counterpart. They derive many interesting results. First, they show that adding regularizer based on gradient norm or the Fisher Information Matrix trace can recover small-batch SGD ... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
This paper does a thorough simulations to clarify and verify the existing hypothesis about why small batch SGD outperform its large batch counterpart. They derive many interesting results. First, they show that adding regularizer based on gradient norm or the Fisher Information Matrix trace can recover small-ba... |
By defining exogeneous information as information irrelevant for control, this work works to learn a representation that removes these features in the context of offline RL. These representations are learned by taking a latent space learned through inverse dynamics modeling. It also provides a set of benchmarks for off... | Recommendation: 3: reject, not good enough | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
By defining exogeneous information as information irrelevant for control, this work works to learn a representation that removes these features in the context of offline RL. These representations are learned by taking a latent space learned through inverse dynamics modeling. It also provides a set of benchmarks... |
This paper considers a probabilistic gradient method (PAGE). The authors invoke different kinds of sampling schemes and claim that the new result is shaper than the original one.
The new sampling assumption is justified by considering several representative sampling schemes. However, the technique is a direct combinati... | Recommendation: 3: reject, not good enough | Area: Optimization (eg, convex and non-convex optimization) | Review:
This paper considers a probabilistic gradient method (PAGE). The authors invoke different kinds of sampling schemes and claim that the new result is shaper than the original one.
The new sampling assumption is justified by considering several representative sampling schemes. However, the technique is a direct c... |
First, the paper introduces a dataset for "difference-aware medical visual question answering". It has chest x-ray pairs from the same patient, with associated questions and answer labels for a VQA task. It is scraped from the existing MIMIC dataset. Though this is similar to "image captioning" task, the authors argue ... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
First, the paper introduces a dataset for "difference-aware medical visual question answering". It has chest x-ray pairs from the same patient, with associated questions and answer labels for a VQA task. It is scraped from the existing MIMIC dataset. Though this is similar to "image captioning" task, the author... |
The paper addresses the universum class (negative class) problem in NLP tasks, by modelling the class boundaries with Gaussian mixture models, and learning thresholds with a boundary learning loss on misclassified examples. They showed that their approach outperforms vanilla classifiers for 3 different NLP tasks.
St... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper addresses the universum class (negative class) problem in NLP tasks, by modelling the class boundaries with Gaussian mixture models, and learning thresholds with a boundary learning loss on misclassified examples. They showed that their approach outperforms vanilla classifiers for 3 different NLP task... |
The authors propose a novel multi-agent architecture and intrinsic reward scheme for avoiding the issue of revisiting previously seen states in multi-agent reinforcement learning. They demonstrate improved performance in a didactic domain as well as in Google Research Football.
# Strengths
* The challenge of revisitat... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
The authors propose a novel multi-agent architecture and intrinsic reward scheme for avoiding the issue of revisiting previously seen states in multi-agent reinforcement learning. They demonstrate improved performance in a didactic domain as well as in Google Research Football.
# Strengths
* The challenge of r... |
The paper proposes a novel two-stage knowledge distillation approach using a teacher network and multiple students which, in essence, combines offline response-based, feature-based and structural knowledge distillation using an online scheme. The authors do so by evaluating for online schemes, namely PCL, ONE, FC, and ... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper proposes a novel two-stage knowledge distillation approach using a teacher network and multiple students which, in essence, combines offline response-based, feature-based and structural knowledge distillation using an online scheme. The authors do so by evaluating for online schemes, namely PCL, ONE, ... |
This work propose a new SNN model named the d-block model, with stochastic absolute refractory periods and recurrent conductance latencies, which reduces the number of sequential computations using fast vectorised operations. Input spikes are processed by d equal length blocks, where every block is a single-spike SNN. ... | Recommendation: 3: reject, not good enough | Area: General Machine Learning | Review:
This work propose a new SNN model named the d-block model, with stochastic absolute refractory periods and recurrent conductance latencies, which reduces the number of sequential computations using fast vectorised operations. Input spikes are processed by d equal length blocks, where every block is a single-spi... |
This paper proposes a low-rank matrix to approximate the coefficient matrix which describes the aggregations of neighborhood information globally. The paper designed a new coefficient matrix with the form of a low rank matrix $UV^t$ which is the solution of the minimization problem inspired by the subspace clustering p... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a low-rank matrix to approximate the coefficient matrix which describes the aggregations of neighborhood information globally. The paper designed a new coefficient matrix with the form of a low rank matrix $UV^t$ which is the solution of the minimization problem inspired by the subspace clus... |
This paper aims to achieve disentangled, interpretable and controllable text-guided image manipulation with CLIP-based models. The authors propose CLIP projection-augmentation embedding (PAE) to replace the original multimodal embedding as optimizing target. PAE is simple and can be easily to injected into other CLIP-b... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper aims to achieve disentangled, interpretable and controllable text-guided image manipulation with CLIP-based models. The authors propose CLIP projection-augmentation embedding (PAE) to replace the original multimodal embedding as optimizing target. PAE is simple and can be easily to injected into othe... |
This paper point out two challenges in the downstream inference of pre-training vision-language models: expressive sensitivity and conceptual sensitivity. To handle the problems, the paper proposes a new dual-model feature prompting methd, named as Decomposed Feature Prompting (DeFo). By providing an independent set o... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper point out two challenges in the downstream inference of pre-training vision-language models: expressive sensitivity and conceptual sensitivity. To handle the problems, the paper proposes a new dual-model feature prompting methd, named as Decomposed Feature Prompting (DeFo). By providing an independe... |
This work proposes a novel episodic memory model for sequential learning, termed an Eigen Memory Tree (EMT). EMT adds to an important but under-studied topic in online learning: efficient online memory models. To evaluate the effectiveness of EMT, the authors tested 206 datasets from OpenML and compared with another on... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
This work proposes a novel episodic memory model for sequential learning, termed an Eigen Memory Tree (EMT). EMT adds to an important but under-studied topic in online learning: efficient online memory models. To evaluate the effectiveness of EMT, the authors tested 206 datasets from OpenML and compared with an... |
This paper studies how to effectively fine-tune a pre-trained network for a downstream task. The challenge here is to identify the mismatch between source and target distributions. To this end, this paper proposes context-aware feature compensation, which learns context information among the training data. They first p... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper studies how to effectively fine-tune a pre-trained network for a downstream task. The challenge here is to identify the mismatch between source and target distributions. To this end, this paper proposes context-aware feature compensation, which learns context information among the training data. They... |
This paper proposed two unsupervised metrics for evaluating image qualities: uMSE and uPSNR. The main idea is to use three noisy references to design an estimator for MSE, which is the uMSE. The noisy references can be computed using consecutive frames or applying spatial subsampling. It uses the uMSE to define the uns... | Recommendation: 5: marginally below the acceptance threshold | Area: Unsupervised and Self-supervised learning | Review:
This paper proposed two unsupervised metrics for evaluating image qualities: uMSE and uPSNR. The main idea is to use three noisy references to design an estimator for MSE, which is the uMSE. The noisy references can be computed using consecutive frames or applying spatial subsampling. It uses the uMSE to define... |
This works studies the noise injection node approach and how it improves the noise-resistance to various types of data perturbations.
The major problem of this manuscript is its writing, which cites a lot of analytical results from Anonymous (2022). Unfortunately, I cannot find this reference from the web at all. This... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This works studies the noise injection node approach and how it improves the noise-resistance to various types of data perturbations.
The major problem of this manuscript is its writing, which cites a lot of analytical results from Anonymous (2022). Unfortunately, I cannot find this reference from the web at a... |
This paper describes a new approach for domain generalization in cross-species pose estimation using different approaches to partition joints to be estimated into different ‘concepts’, and separately training parts of the network to estimate the position of each joint concept group. Three approaches for separating join... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
This paper describes a new approach for domain generalization in cross-species pose estimation using different approaches to partition joints to be estimated into different ‘concepts’, and separately training parts of the network to estimate the position of each joint concept group. Three approaches for separat... |
This paper proposes a new framework that performs the Named Entity Recognition task from the semantic matching perspective (bi-encoder here). The idea is to map candidate text spans and entity types into the same vector representation space and perform the NER task using distance metric accordingly. One important issue... | Recommendation: 6: marginally above the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper proposes a new framework that performs the Named Entity Recognition task from the semantic matching perspective (bi-encoder here). The idea is to map candidate text spans and entity types into the same vector representation space and perform the NER task using distance metric accordingly. One importa... |
The paper studies some tasks related to programming language (e.g. clone detection, code search). The paper makes two contributions in the domain: firstly, through online OJ code mining, the paper proposes a new benchmark XCD that evaluates model over multiple program languages. Secondly, the paper proposes a new pretr... | Recommendation: 5: marginally below the acceptance threshold | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
The paper studies some tasks related to programming language (e.g. clone detection, code search). The paper makes two contributions in the domain: firstly, through online OJ code mining, the paper proposes a new benchmark XCD that evaluates model over multiple program languages. Secondly, the paper proposes a n... |
This paper studies the bank-loan problem: a binary classification setting where the learner only receives feedback when it assigns an example the positive label. In this setting, training on a dataset consisting only of the accepted examples naturally leads to bias, since the learner has no ability to correct mislabell... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper studies the bank-loan problem: a binary classification setting where the learner only receives feedback when it assigns an example the positive label. In this setting, training on a dataset consisting only of the accepted examples naturally leads to bias, since the learner has no ability to correct m... |
This paper proposes HASTE for estimating the transferability of a source domain to a target domain. Two techniques are introduced, one is class-agnostic and another is class-specific. The techniques achieve state-of-the-art compared to other concurrent baseline metrics.
--post rebuttal--
Dear authors, thank you for g... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes HASTE for estimating the transferability of a source domain to a target domain. Two techniques are introduced, one is class-agnostic and another is class-specific. The techniques achieve state-of-the-art compared to other concurrent baseline metrics.
--post rebuttal--
Dear authors, thank y... |
- This paper investigates the qualities of object-centric representations (OCR) in the context of reinforcement learning (RL) tasks. Does so by empirically benchmarking OCR and non-OCR methods (eg: end-to-end learning (E2E) and variational auto-encoding (VAE)) on 5 tasks from 2 domains (see Fig. 2). Each benchmark is e... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
- This paper investigates the qualities of object-centric representations (OCR) in the context of reinforcement learning (RL) tasks. Does so by empirically benchmarking OCR and non-OCR methods (eg: end-to-end learning (E2E) and variational auto-encoding (VAE)) on 5 tasks from 2 domains (see Fig. 2). Each benchm... |
This paper proposes a simple new ensemble method for protecting data from someone training on it. The work shows the empirical success of their method on multiple datasets and compared to strong baselines like adversarial poisons.
The main strength of the paper is in the empirical success compared to strong baselines ... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a simple new ensemble method for protecting data from someone training on it. The work shows the empirical success of their method on multiple datasets and compared to strong baselines like adversarial poisons.
The main strength of the paper is in the empirical success compared to strong ba... |
This paper reveals an interesting finding that minMPJPE metric would result in miscalibrated distribution in a 2D-to-3D lifting problem, which is particularly unsuitable for using MPJPE to evaluate those distribution-based methods that generate multiple hypotheses to solve the ambiguity. It provides a toy example that ... | Recommendation: 3: reject, not good enough | Area: Applications (eg, speech processing, computer vision, NLP) | Review:
This paper reveals an interesting finding that minMPJPE metric would result in miscalibrated distribution in a 2D-to-3D lifting problem, which is particularly unsuitable for using MPJPE to evaluate those distribution-based methods that generate multiple hypotheses to solve the ambiguity. It provides a toy examp... |
This paper proposes a method called Kcal for the (full) calibration of deep neural networks.
Kcal consists in replacing the last layer (usually a softmax) of a neural network with a learned projection layer for dimensionality reduction, followed by a Kernel Density (KDE)-based estimator to produce output probabilities.... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes a method called Kcal for the (full) calibration of deep neural networks.
Kcal consists in replacing the last layer (usually a softmax) of a neural network with a learned projection layer for dimensionality reduction, followed by a Kernel Density (KDE)-based estimator to produce output probab... |
The paper investigates the role of neural and synaptic heterogeneity in improving the computational capacity and/or energy efficiency of spiking recurrent neural nets. It also provides an optimization procedure for determining the optimal degree of heterogeneity for a problem.
Strengths:
+ biological motivation
+ what ... | Recommendation: 8: accept, good paper | Area: Unsupervised and Self-supervised learning | Review:
The paper investigates the role of neural and synaptic heterogeneity in improving the computational capacity and/or energy efficiency of spiking recurrent neural nets. It also provides an optimization procedure for determining the optimal degree of heterogeneity for a problem.
Strengths:
+ biological motivation... |
This paper introduces an efficient method to finetune pre-trained LDM to small data. The problem is formulated as an image-to-image generation task. A conditional image is first encoded by the CLIP image encoder and then injected into the LDM by the cross attention mechanism. To avoid overfitting, only the parameters o... | Recommendation: 6: marginally above the acceptance threshold | Area: Generative models | Review:
This paper introduces an efficient method to finetune pre-trained LDM to small data. The problem is formulated as an image-to-image generation task. A conditional image is first encoded by the CLIP image encoder and then injected into the LDM by the cross attention mechanism. To avoid overfitting, only the para... |
This paper investigates the application of language models to interactive theorem proving by doing a qualitative case study on input-dependent prompt engineering. It shows that Codex is capable of producing partially correct formal proofs (in Lean 4) that can be turned into correct proofs with a moderate amount of modi... | Recommendation: 5: marginally below the acceptance threshold | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper investigates the application of language models to interactive theorem proving by doing a qualitative case study on input-dependent prompt engineering. It shows that Codex is capable of producing partially correct formal proofs (in Lean 4) that can be turned into correct proofs with a moderate amount... |
This paper presents a weakly supervised learning approach to generate labels for unlabeled triples on graphs. Overall, the proposed approach could be regarded as a combination of inductive logic programming and active learning. The ILP module learns a set of (noisy) logic rules to perform relation completion in the gra... | Recommendation: 5: marginally below the acceptance threshold | Area: General Machine Learning | Review:
This paper presents a weakly supervised learning approach to generate labels for unlabeled triples on graphs. Overall, the proposed approach could be regarded as a combination of inductive logic programming and active learning. The ILP module learns a set of (noisy) logic rules to perform relation completion in... |
This paper studies dimension collapse in federated learning with label distribution shift, and methods on more to mitigate it. With the simple observation that the singular values of features are decaying fast for heterogeneous cases, the authors propose to minimize the corresponding variance, or equivalently, the Frob... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper studies dimension collapse in federated learning with label distribution shift, and methods on more to mitigate it. With the simple observation that the singular values of features are decaying fast for heterogeneous cases, the authors propose to minimize the corresponding variance, or equivalently, ... |
Inspired by Global Workspace Theory of conscioussness authors propose TAMiL - a continual learning method that entails task-attention modules to capture task-specific information from the common representation space. Experimental results show that their method outperforms SOTA rehearsal-based and dynamic sparse approac... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
Inspired by Global Workspace Theory of conscioussness authors propose TAMiL - a continual learning method that entails task-attention modules to capture task-specific information from the common representation space. Experimental results show that their method outperforms SOTA rehearsal-based and dynamic sparse... |
This paper proposes to use coordinate-based networks for time-series data representation. This is done by proposing a new architecture which separates a time series into a trend (low-frequency) and seasonal (high-frequency) component by representing each independently with a set of weights and a SIREN, and then summing... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
This paper proposes to use coordinate-based networks for time-series data representation. This is done by proposing a new architecture which separates a time series into a trend (low-frequency) and seasonal (high-frequency) component by representing each independently with a set of weights and a SIREN, and then... |
Self-Supervised Learning (SSL) algorithms like BarlowTwins and BYOL (both contrastive, non-contrastive) are known to suffer from $\textit{dimension collapse}$, where the effective rank of the representations is much smaller than the dimensionality of the representation. Empirical analysis of such algorithms presents co... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
Self-Supervised Learning (SSL) algorithms like BarlowTwins and BYOL (both contrastive, non-contrastive) are known to suffer from $\textit{dimension collapse}$, where the effective rank of the representations is much smaller than the dimensionality of the representation. Empirical analysis of such algorithms pre... |
Authors propose a model for generating layouts of furniture in rooms. Unlike existing works in this area (e.g. ATISS), the proposed method allows conditioning on combinations of attributes (e.g. "place a chair somewhere, and some other object at this particular location"). The proposed model is a novel combination of t... | Recommendation: 5: marginally below the acceptance threshold | Area: Generative models | Review:
Authors propose a model for generating layouts of furniture in rooms. Unlike existing works in this area (e.g. ATISS), the proposed method allows conditioning on combinations of attributes (e.g. "place a chair somewhere, and some other object at this particular location"). The proposed model is a novel combinat... |
The paper provides an improved way for constructing combinatorial optimization networks for problems that involve cardinality constraints.
The main strength of the paper is providing an extension to combinatorial optimization network framework that allows for bounding the constraint violation, for the case of cardinali... | Recommendation: 6: marginally above the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper provides an improved way for constructing combinatorial optimization networks for problems that involve cardinality constraints.
The main strength of the paper is providing an extension to combinatorial optimization network framework that allows for bounding the constraint violation, for the case of c... |
The paper proposes using an empirical Bayesian framework to simultaneously learn an embedding model ($f_\phi$) and a probabilistic model (parameterized w/ $\theta$) for the continual learning setting. To that end, $f_\phi$ is updated using a variant of experience replay through a running memory of samples, where the me... | Recommendation: 5: marginally below the acceptance threshold | Area: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes) | Review:
The paper proposes using an empirical Bayesian framework to simultaneously learn an embedding model ($f_\phi$) and a probabilistic model (parameterized w/ $\theta$) for the continual learning setting. To that end, $f_\phi$ is updated using a variant of experience replay through a running memory of samples, wher... |
This paper tackles the problem of source-free domain adaptation (SFDA), where a pre-trained source model is adapted using unlabeled target domain data without accessing any source domain data. While previous SFDA algorithms only considered a narrow set of domain shifts in computer vision tasks, the authors put them on ... | Recommendation: 3: reject, not good enough | Area: Unsupervised and Self-supervised learning | Review:
This paper tackles the problem of source-free domain adaptation (SFDA), where a pre-trained source model is adapted using unlabeled target domain data without accessing any source domain data. While previous SFDA algorithms only considered a narrow set of domain shifts in computer vision tasks, the authors put ... |
The paper studies the stability properties of the hard thresholding operator in gradient descent framework and introduces several new concepts such as HT-stable and HT-unstable stationary points. Moreover, the paper aims to answer four fundamental questions regarding the local/global minimizers and accumulation points ... | Recommendation: 3: reject, not good enough | Area: Optimization (eg, convex and non-convex optimization) | Review:
The paper studies the stability properties of the hard thresholding operator in gradient descent framework and introduces several new concepts such as HT-stable and HT-unstable stationary points. Moreover, the paper aims to answer four fundamental questions regarding the local/global minimizers and accumulation... |
This paper develops a new optimization algorithm named PGrad for domain generalization. The proposed method attempts to filter out domain-specific noise and provide an update direction that maximally benefits all training domains. The experimental results show improved domain generalization performance on multiple data... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper develops a new optimization algorithm named PGrad for domain generalization. The proposed method attempts to filter out domain-specific noise and provide an update direction that maximally benefits all training domains. The experimental results show improved domain generalization performance on multi... |
This paper studies sequential selection of laboratory test panels, formulated as a cost-sensitive adaptive feature selection/acquisition problem. The two main innovations are:
1. For diagnostic performance, instead of maximizing accuracy, this work proposes to maximize F1 score to address high class imbalance (how hig... | Recommendation: 8: accept, good paper | Area: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability ) | Review:
This paper studies sequential selection of laboratory test panels, formulated as a cost-sensitive adaptive feature selection/acquisition problem. The two main innovations are:
1. For diagnostic performance, instead of maximizing accuracy, this work proposes to maximize F1 score to address high class imbalance ... |
The authors present a framework/tool for understanding and evaluating multimodal networks. This includes four main parts: unimodal, cross modal, multimodal (representations) and multimodal (predictions). They also present a novel interpretability methods for a subset of these. Finally, they evaluate this framework with... | 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 authors present a framework/tool for understanding and evaluating multimodal networks. This includes four main parts: unimodal, cross modal, multimodal (representations) and multimodal (predictions). They also present a novel interpretability methods for a subset of these. Finally, they evaluate this framew... |
In this paper, the authors look at the problem of recovering the DAG structure from samples from a DAG-structured distribution. Despite having seen a lot of important works spanning the last several decades, the problem has remained open and is an active area of research. The proposed solution in this paper consists of... | Recommendation: 6: marginally above the acceptance threshold | Area: General Machine Learning | Review:
In this paper, the authors look at the problem of recovering the DAG structure from samples from a DAG-structured distribution. Despite having seen a lot of important works spanning the last several decades, the problem has remained open and is an active area of research. The proposed solution in this paper con... |
The authors reveal that Neural Module Networks (NMNs), i.e., question-specific compositions of modules that tackle a sub-task, achieve better or similar systematic generalization performance than the conventional Transformers, even though NMNs’ modules are CNN-based.
To address this shortcoming of Transformers with re... | Recommendation: 5: marginally below the acceptance threshold | Area: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces) | Review:
The authors reveal that Neural Module Networks (NMNs), i.e., question-specific compositions of modules that tackle a sub-task, achieve better or similar systematic generalization performance than the conventional Transformers, even though NMNs’ modules are CNN-based.
To address this shortcoming of Transformers... |
This paper proposes an offline learning approach for path planning. The proposed Dual Gradient Fields (DualGF) model the probability distributions of (1) target states (e.g. goal states) and (2) free space. Then, the problem of finding a path toward a target state while avoiding collision becomes simply following the g... | Recommendation: 5: marginally below the acceptance threshold | Area: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics) | Review:
This paper proposes an offline learning approach for path planning. The proposed Dual Gradient Fields (DualGF) model the probability distributions of (1) target states (e.g. goal states) and (2) free space. Then, the problem of finding a path toward a target state while avoiding collision becomes simply followi... |
The paper aims to tackle the out-of-domain generalization challenge for instance segmentation of 2D images and 3D point clouds. The proposed method combines the recent slot attention-based model for segmentation and the standard test-time-augmentation approach with the auxiliary reconstruction task. The proposed method... | Recommendation: 5: marginally below the acceptance threshold | Area: Deep Learning and representational learning | Review:
The paper aims to tackle the out-of-domain generalization challenge for instance segmentation of 2D images and 3D point clouds. The proposed method combines the recent slot attention-based model for segmentation and the standard test-time-augmentation approach with the auxiliary reconstruction task. The propose... |
This paper proposes some modifications to the execution flow of BNNs for better performance on low-end hardware, specifically ARM CPUs. Firstly, the authors reduce 32-bit intermediate representations after the XNOR popcount in the BNN to 8-bit values to reduce overhead. Secondly, they replace BN with a comparison opera... | Recommendation: 3: reject, not good enough | Area: Deep Learning and representational learning | Review:
This paper proposes some modifications to the execution flow of BNNs for better performance on low-end hardware, specifically ARM CPUs. Firstly, the authors reduce 32-bit intermediate representations after the XNOR popcount in the BNN to 8-bit values to reduce overhead. Secondly, they replace BN with a comparis... |
The paper address the scalability issue of forward gradient learning by employing many different local greedy loss functions (blockwise, patch-wise, and group-wise local losses, and a combination of all three). The paper shows good performance on MNIST and CIFAR-10, and also outperforms other backprop free algorithms ... | Recommendation: 8: accept, good paper | Area: Deep Learning and representational learning | Review:
The paper address the scalability issue of forward gradient learning by employing many different local greedy loss functions (blockwise, patch-wise, and group-wise local losses, and a combination of all three). The paper shows good performance on MNIST and CIFAR-10, and also outperforms other backprop free alg... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.