A naukri.com initiative
ML News
Arxiv
1d
337
Image Credit: Arxiv
MAGE: Model-Level Graph Neural Networks Explanations via Motif-based Graph Generation
- Graph Neural Networks (GNNs) have shown remarkable success in molecular tasks, yet their interpretability remains challenging.
- To address limitations in traditional model-level explanation methods, an innovative approach called MAGE (Motif-based GNN Explainer) has been introduced.
- MAGE uses motifs as fundamental units for generating explanations and incorporates critical substructures into the explanations.
- The effectiveness of MAGE has been demonstrated through quantitative and qualitative assessments on real-world molecular datasets.
Read Full Article
20 Likes
Arxiv
1d
116
Image Credit: Arxiv
On the Benefits of Memory for Modeling Time-Dependent PDEs
- Data-driven techniques have emerged as a promising alternative to traditional numerical methods for solving PDEs.
- In this work, the benefits of using memory for modeling time-dependent PDEs are investigated.
- The Memory Neural Operator (MemNO) architecture effectively models memory in PDEs.
- Empirical demonstrations show that MemNO outperforms baselines without memory, with up to 6x reduction in test error.
Read Full Article
6 Likes
Arxiv
1d
116
Image Credit: Arxiv
Regressing the Relative Future: Efficient Policy Optimization for Multi-turn RLHF
- Large Language Models (LLMs) such as GPT-3 have achieved great success in single-turn tasks like summarization.
- However, they struggle with multi-turn tasks like dialogue that require long-term planning.
- To address this, researchers have introduced REgressing the RELative FUture (REFUEL), an efficient policy optimization approach for multi-turn reinforcement learning from human feedback (RLHF) in LLMs.
- REFUEL outperforms state-of-the-art methods like DPO and REBEL, and can match the performance of any policy covered by the training set.
Read Full Article
6 Likes
Arxiv
1d
194
Image Credit: Arxiv
A Simple and Efficient Approach to Batch Bayesian Optimization
- Extending Bayesian optimization to batch evaluation can enable the designer to make the most use of parallel computing technology.
- A simple and efficient approach is proposed to extend Bayesian optimization to large-scale batch evaluation.
- The approach involves drawing a batch of axis-aligned subspaces of the original problem and selecting one acquisition point from each subspace.
- Numerical experiments show that the proposed approach significantly speeds up convergence and performs competitively compared to other batch Bayesian optimization algorithms.
Read Full Article
11 Likes
Arxiv
1d
284
Image Credit: Arxiv
Know Unreported Roadway Incidents in Real-time: Early Traffic Anomaly Detection
- This research aims to know traffic anomalies as early as possible.
- The objective is to inform traffic operators of unreported incidents in real time and as early as possible.
- A deep learning framework utilizing prior domain knowledge and model-designing strategies is proposed for early detection of traffic anomalies.
- The experimental results demonstrate the effectiveness of the model in early anomaly detection.
Read Full Article
17 Likes
Arxiv
1d
191
Image Credit: Arxiv
Optimal Rates for Robust Stochastic Convex Optimization
- Machine learning algorithms in high-dimensional settings are highly susceptible to the influence of even a small fraction of structured outliers, making robust optimization techniques essential.
- New novel algorithms have been developed that achieve minimax-optimal excess risk (up to logarithmic factors) under the epsilon-contamination model.
- These algorithms do not require stringent assumptions, such as Lipschitz continuity and smoothness of individual sample functions, improving over existing suboptimal algorithms.
- The developed algorithms can also handle the case of unknown covariance parameter and can be extended to nonsmooth population risks via convolutional smoothing.
Read Full Article
11 Likes
Arxiv
1d
63
Image Credit: Arxiv
Emergent Symbol-like Number Variables in Artificial Neural Networks
- Researchers have investigated the numeric representations that emerge in neural systems and how they can be interpreted through the lens of interpretable Symbolic Algorithms (SAs).
- The study analyzed the raw activity of GRUs, LSTMs, and Transformers trained on numeric tasks and found that the neural activity can be interpreted as simplified SAs when framed in interpretable subspaces.
- The research highlighted the importance of causal interventions for neural network interpretability and demonstrated that recurrent models develop graded, symbol-like number variables in their neural activity.
- Additionally, the study introduced a generalization of Distributed Alignment Search (DAS) and showed that Transformers employ anti-Markovian solutions in the absence of sufficient attention layers.
Read Full Article
3 Likes
Arxiv
1d
326
Image Credit: Arxiv
Spatially-Delineated Domain-Adapted AI Classification: An Application for Oncology Data
- Given multi-type point maps from different place-types (e.g., tumor regions), researchers aim to develop a classifier trained on the source place-type to accurately distinguish between two classes of the target place-type based on their point arrangements.
- The challenge lies in the spatial variability and inherent heterogeneity observed in spatial properties or arrangements across different place-types.
- The proposed multi-task self-learning framework targets spatial arrangements, utilizing techniques such as spatial mix-up masking and spatial contrastive predictive coding.
- Experimental results on real-world datasets, specifically oncology data, demonstrate that the framework provides higher prediction accuracy compared to baseline methods.
Read Full Article
19 Likes
Arxiv
1d
359
Image Credit: Arxiv
Weak-to-Strong Diffusion with Reflection
- The Weak-to-Strong Diffusion (W2SD) framework is proposed to reduce the gap between generated outputs and real data in diffusion generative models.
- W2SD utilizes the estimated difference between weak and strong models to bridge the gap and align latent variables with the real data distribution.
- The W2SD framework is highly flexible and applicable to various model pairs and modalities, achieving state-of-the-art performance.
- Experiments demonstrate significant improvements in human preference, aesthetic quality, and prompt adherence with W2SD.
Read Full Article
21 Likes
Arxiv
1d
221
Image Credit: Arxiv
Efficient Model Editing with Task Vector Bases: A Theoretical Framework and Scalable Approach
- Efficient Model Editing with Task Vector Bases: A Theoretical Framework and Scalable Approach
- Task vectors enable flexible task adaptation and model merging through arithmetic operations.
- Existing approaches often have limited theoretical support and performance gaps compared to direct task fine tuning.
- A new framework, task vector bases, reduces memory usage and maintains compositional advantage for large-scale task arithmetic.
Read Full Article
13 Likes
Arxiv
1d
101
Image Credit: Arxiv
Efficient Neural Network Approaches for Conditional Optimal Transport with Applications in Bayesian Inference
- Two neural network approaches have been developed to approximate the solutions of conditional optimal transport (COT) problems.
- The approaches enable conditional sampling and density estimation, which are important in Bayesian inference.
- The methods represent the target conditional distribution as a transformation of a tractable reference distribution.
- The algorithms use neural networks to parameterize candidate maps and exploit the structure of the COT problem.
Read Full Article
6 Likes
Arxiv
1d
187
Image Credit: Arxiv
Towards Spatially-Lucid AI Classification in Non-Euclidean Space: An Application for MxIF Oncology Data
- Researchers have developed a spatially-lucid classifier for analyzing immune-tumor relationships and designing new immunotherapies in oncology.
- The classifier can distinguish between two classes based on the arrangements of their multi-category point sets.
- Unlike previous techniques, the proposed framework can handle significant spatial variability within a single place-type.
- Experimental results on real-world MxIF oncology data demonstrate higher prediction accuracy compared to baseline methods.
Read Full Article
11 Likes
Arxiv
1d
56
Image Credit: Arxiv
Variation Due to Regularization Tractably Recovers Bayesian Deep Learning
- Uncertainty quantification is crucial in deep learning for safe and reliable decision-making.
- A new method based on variation due to regularization is proposed for uncertainty quantification in large networks.
- The method adjusts the training loss during fine-tuning and reflects confidence in the output based on all layers of the network.
- Experiments show that the proposed method provides rigorous uncertainty estimates and improves uncertainty quantification quality.
Read Full Article
3 Likes
Arxiv
1d
288
Image Credit: Arxiv
AI in Lung Health: Benchmarking Detection and Diagnostic Models Across Multiple CT Scan Datasets
- Lung cancer remains the leading cause of cancer-related mortality worldwide.
- AI models are being integrated into medical imaging for the early detection of lung cancer.
- Duke Lung Cancer Screening (DLCS) Dataset, a large open-access dataset with over 2,000 scans and 3,000 expert-verified nodules, is introduced.
- Benchmarking of deep learning models for nodule detection and lung cancer classification is conducted across multiple datasets, demonstrating strong generalizability.
Read Full Article
17 Likes
Arxiv
1d
63
Image Credit: Arxiv
RACH Traffic Prediction in Massive Machine Type Communications
- Traffic pattern prediction in massive machine-type communication (mMTC) networks is challenging due to the inherent randomness of events and bursty traffic.
- A machine learning-based framework using long-term short-term memory (LSTM) and DenseNet with feed-forward neural network (FFNN) layers is proposed for forecasting bursty traffic in multi-channel slotted ALOHA networks.
- The framework includes a low-complexity online prediction algorithm that updates the states of the LSTM network using frequently collected data from the mMTC network.
- Simulation results show that the proposed framework achieves a 52% higher accuracy in long-term predictions compared to traditional methods, without imposing additional processing load.
Read Full Article
3 Likes
For uninterrupted reading, download the app