A naukri.com initiative
ML News
Arxiv
6h
195
Image Credit: Arxiv
Leveraging Generalizability of Image-to-Image Translation for Enhanced Adversarial Defense
- Adversarial attacks pose a critical vulnerability in machine learning models by tricking them using nearly invisible perturbations to images.
- Existing defensive mechanisms for mitigating adversarial attacks often require significant time and computational costs.
- This study presents an improved model that incorporates residual blocks to enhance the generalizability and transferability of the defense method.
- Experimental results demonstrate that the proposed model can restore classification accuracy while maintaining competitive performance compared to state-of-the-art methods.
Read Full Article
11 Likes
Arxiv
6h
101
Image Credit: Arxiv
ToolACE-R: Tool Learning with Adaptive Self-Refinement
- ToolACE-R is a novel method for tool learning with adaptive self-refinement.
- Current tool learning approaches mainly focus on data synthesis for fine-tuning models but overlook fully stimulating model potential.
- ToolACE-R incorporates an iterative training procedure that progressively incorporates more training samples based on the model's evolving capabilities.
- Experimental results show that ToolACE-R achieves competitive performance and can further improve through adaptive self-refinement.
Read Full Article
6 Likes
Arxiv
6h
80
Image Credit: Arxiv
Multi-convex Programming for Discrete Latent Factor Models Prototyping
- Discrete latent factor models (DLFMs) are widely used in various domains such as machine learning, economics, neuroscience, psychology, etc.
- A generic framework based on CVXPY is proposed to specify and solve the fitting problem of a wide range of DLFMs, including regression and classification models.
- The framework allows for the integration of regularization terms and constraints on the DLFM parameters and latent factors, providing flexibility for customization.
- An open-source Python implementation is introduced, and several examples illustrate the usage of the framework.
Read Full Article
4 Likes
Arxiv
6h
299
Image Credit: Arxiv
BiSeg-SAM: Weakly-Supervised Post-Processing Framework for Boosting Binary Segmentation in Segment Anything Models
- Accurate segmentation of polyps and skin lesions is crucial for diagnosing colorectal and skin cancers.
- The pixel-level annotation of medical images is time-consuming and costly.
- The paper proposes BiSeg-SAM, a weakly supervised post-processing framework for the segmentation of polyps and skin lesions.
- BiSeg-SAM demonstrates superior performance compared to state-of-the-art methods in polyp and skin cancer segmentation.
Read Full Article
18 Likes
Arxiv
6h
296
Image Credit: Arxiv
Incorporating Coupling Knowledge into Echo State Networks for Learning Spatiotemporally Chaotic Dynamics
- Machine learning methods have shown promise in learning chaotic dynamical systems.
- In large-scale, spatiotemporally chaotic systems, data-driven machine learning methods suffer from inefficiencies.
- Clustered echo state networks, incorporating the spatial coupling structure, outperform existing models in learning chaotic systems.
- The approach remains effective even with imperfect prior coupling knowledge and noise in training data.
Read Full Article
17 Likes
Arxiv
6h
24
Image Credit: Arxiv
Density estimation via mixture discrepancy and moments
- Researchers have proposed density estimation via mixture discrepancy and moments as alternative methods to the star discrepancy for generalizing histogram statistics to higher dimensional cases.
- The density estimation via mixture discrepancy based sequential partition (DSP-mix) and density estimation via moments based sequential partition (MSP) are computationally tractable and exhibit reflection and rotation invariance.
- Both DSP-mix and MSP run approximately ten times faster than density estimation via discrepancy based sequential partition (DSP) while maintaining the same accuracy.
- Numerical experiments demonstrate the efficiency of DSP-mix and MSP in reconstructing the d-D mixture of Gaussians and Betas for various dimensions.
Read Full Article
1 Like
Arxiv
6h
27
Image Credit: Arxiv
Pro-DG: Procedural Diffusion Guidance for Architectural Facade Generation
- Pro-DG is a framework for procedurally controllable photo-realistic facade generation that combines a procedural shape grammar with diffusion-based image synthesis.
- The framework reconstructs the facade layout from a single input image and allows user-defined transformations to edit the structure.
- A hierarchical matching procedure aligns facade structures at different levels and control maps guide a generative diffusion pipeline, enabling large-scale edits while preserving local appearance fidelity.
- Pro-DG outperforms inpainting-based baselines and synthetic ground truths in terms of preservation of architectural identity and edit accuracy.
Read Full Article
1 Like
Arxiv
6h
48
Image Credit: Arxiv
A Causal Inference Framework for Data Rich Environments
- A formal model for counterfactual estimation with unobserved confounding in data-rich environments has been proposed.
- The model combines the structural causal model view with the latent factor model view of causal inference.
- Classic models for potential outcomes and treatment assignments fit within this framework.
- The study establishes consistency of estimators for various causal parameters.
Read Full Article
2 Likes
Arxiv
6h
69
Image Credit: Arxiv
TransforMerger: Transformer-based Voice-Gesture Fusion for Robust Human-Robot Communication
- TransforMerger is a transformer-based reasoning model that infers a structured action command for robotic manipulation based on fused voice and gesture inputs.
- It merges multimodal data into a single unified sentence and employs probabilistic embeddings to handle uncertainty.
- The model integrates contextual scene understanding to resolve ambiguous references and is robust to noise, misalignment, and missing information.
- TransforMerger outperforms deterministic baselines, demonstrating its effectiveness in enabling more robust and flexible human-robot communication.
Read Full Article
4 Likes
Arxiv
6h
13
Image Credit: Arxiv
Learning with Imperfect Models: When Multi-step Prediction Mitigates Compounding Error
- Compounding error, where small prediction mistakes accumulate over time, presents a major challenge in learning-based control.
- Mitigating compounding error is important in model-based reinforcement learning and imitation learning.
- Training multi-step predictors directly can help reduce compounding error and improve performance.
- In the context of linear dynamical systems, well-specified single-step models achieve lower asymptotic prediction error, while direct multi-step predictors perform better in case of misspecified models with partial observability.
Read Full Article
Like
Arxiv
6h
229
Image Credit: Arxiv
BlenderGym: Benchmarking Foundational Model Systems for Graphics Editing
- BlenderGym is the first comprehensive benchmark for vision-language models (VLM) in 3D graphics editing.
- It evaluates VLM systems through code-based 3D reconstruction tasks.
- BlenderGym reveals that even state-of-the-art VLM systems struggle with tasks that are relatively easy for human Blender users.
- The study also explores how inference scaling techniques and distribution of inference compute impact VLM's performance on graphics editing tasks.
Read Full Article
13 Likes
Arxiv
6h
344
Image Credit: Arxiv
A Novel Approach To Implementing Knowledge Distillation In Tsetlin Machines
- The Tsetlin Machine (TM) is a propositional logic based model that uses conjunctive clauses to learn patterns from data.
- A novel approach to implementing knowledge distillation in Tsetlin Machines is proposed.
- Utilizing probability distributions of each output sample in the teacher provides additional context to the student model.
- The proposed algorithm improves the performance of the student model without negatively impacting latency in image recognition and text classification.
Read Full Article
20 Likes
Arxiv
6h
278
Image Credit: Arxiv
Barrier Certificates for Unknown Systems with Latent States and Polynomial Dynamics using Bayesian Inference
- Certifying safety in dynamical systems is crucial, but barrier certificates typically require explicit system models.
- A novel approach is proposed for synthesizing barrier certificates for unknown systems with latent states and polynomial dynamics.
- A Bayesian framework is employed, updating a prior in state-space representation using input-output data via a targeted marginal Metropolis-Hastings sampler.
- The resulting samples are used to construct a candidate barrier certificate through a sum-of-squares program, providing probabilistic guarantees for the unknown system.
Read Full Article
16 Likes
Arxiv
6h
247
Image Credit: Arxiv
Autonomous optical navigation for DESTINY+: Enhancing misalignment robustness in flyby observations with a rotating telescope
- The upcoming JAXA Epsilon medium-class mission, DESTINY+, aims to flyby multiple asteroids including Phaethon.
- The mission's flyby observation instrument, TCAP, is a telescope capable of single-axis rotation.
- TCAP is also used as a navigation camera for autonomous optical navigation during the closest-approach phase.
- A proposed algorithm utilizing the unscented Kalman filter can mitigate misalignment-induced degradation of the optical navigation accuracy.
Read Full Article
14 Likes
Arxiv
6h
261
Image Credit: Arxiv
A Randomized Zeroth-Order Hierarchical Framework for Heterogeneous Federated Learning
- A novel framework is proposed for heterogeneous federated learning (FL) to address client heterogeneity and improve model performance.
- The framework captures local and global training processes through a bilevel formulation.
- It includes personalized learning, pre-training on the server's side, nonstandard aggregation, nonidentical local steps, and clients' local constraints.
- The proposed method, ZO-HFL, achieves nonasymptotic and asymptotic convergence guarantees without relying on standard assumptions in heterogeneous FL.
Read Full Article
15 Likes
For uninterrupted reading, download the app