Deep Learning News, Latest Updates and Recent Announcements on Techminis

A naukri.com initiative

New

Home

Deep Learning News

Medium

21h

338

Image Credit: Medium

Understanding Markov Decision Processes

Markov Decision Processes (MDPs) are used to model decision-making problems in which outcomes are partly random and partly under the control of a decision maker.
MDPs consist of a set of states, a set of actions, transition probabilities, and rewards.
The aim of an MDP is to obtain good policies to obtain the best value.
An optimal policy is calculated using the Bellman equation, and the optimal value for the game is obtained by evaluating the optimal policy.
The value of a state depends on the values of the actions possible and the current policy.
The Q-value, or the value of a state-action pair, depends on the expected next reward and the expected sum of the remaining rewards.
The discount factor is a parameter that specifies the importance of future rewards relative to immediate rewards.
When specifying the MDP, the transition probabilities should sum to 1.
MDPs help in decision making processes that involve random outcomes and decisions to be made by the decision maker.
Resources to further understand MDPs include the book on Reinforcement Learning by Sutton and Barto and the lecture on this topic by Dorsa Sadigh in CS 221 at Stanford.

Read Full Article

20 Likes

Medium

22h

224

Working with Multi-View Diffusion Models part7(Machine Learning )

The paper proposes an approach called Grounded-Dreamer for generating 3D assets based on complex text prompts using a pre-trained multi-view diffusion model.
The approach leverages text-guided 4-view images and introduces an attention refocusing mechanism to improve text-aligned 4-view image generation.
A hybrid optimization strategy is also proposed to optimize synergy between the score distillation sampling (SDS) loss and sparse RGB reference images.
The proposed approach consistently outperforms previous state-of-the-art methods in generating accurate and high-quality compositional 3D assets.

Read Full Article

13 Likes

Medium

22h

104

Working with Multi-View Diffusion Models part5(Machine Learning )

Researchers propose Grounded-Dreamer, a two-stage approach for generating high-fidelity 3D assets.
The approach utilizes a pre-trained multi-view diffusion model and text-guided 4-view images.
An attention refocusing mechanism is introduced to align 4-view image generation with the text prompt.
The method outperforms previous state-of-the-art methods in quality and accuracy of generating 3D assets.

Read Full Article

6 Likes

Medium

22h

Working with Multi-View Diffusion Models part4(Machine Learning )

This paper presents a neural architecture MVDiffusion++ for 3D object reconstruction.
MVDiffusion++ synthesizes dense and high-resolution views of an object given one or a few images without camera poses.
MVDiffusion++ achieves superior flexibility and scalability with a pose-free architecture and a view dropout strategy.
It significantly outperforms the current state of the arts in novel view synthesis and 3D reconstruction.

Read Full Article

3 Likes

Discover more

Medium

22h

168

Working with Multi-View Diffusion Models part3(Machine Learning )

The paper proposes DiffPoint, an architecture that combines vision transformers (ViT) and diffusion models for point cloud reconstruction.
DiffPoint divides noisy point clouds into patches and uses a ViT backbone to predict target points based on input images.
The architecture achieves state-of-the-art results in both single-view and multi-view reconstruction tasks.
Additionally, a feature fusion module is introduced to aggregate image features from single or multiple input images.

Read Full Article

10 Likes

Medium

22h

216

Working with Multi-View Diffusion Models part1(Machine Learning )

A new spatial-temporal consistent diffusion framework called DrivingDiffusion has been proposed.
DrivingDiffusion generates realistic multi-view videos controlled by a 3D layout.
The framework ensures cross-view consistency and cross-frame consistency in the synthesized videos.
DrivingDiffusion can generate large-scale realistic multi-camera driving videos in complex urban scenes.

Read Full Article

13 Likes

Medium

23h

362

Working with Modern deep learning part8

Natural Language Processing (NLP) models have achieved phenomenal success in linguistic and semantic tasks.
Large models like BERT and GPT have achieved impressive performances but come with high computational costs.
To address this, transfer learning, pruning, quantization, and knowledge distillation techniques have been used to achieve smaller models with similar performances.
Knowledge Retrievers and efficient attention mechanisms have been developed to extract data and improve inference.

Read Full Article

21 Likes

Medium

213

Revision Research on Data Pruning part5(Machine Learning)

Improving the efficiency of Neural Architecture Search (NAS) is a challenging task.
This work pioneers an exploration into the critical role of dataset characteristics for DARTS bi-level optimization.
They propose a novel Bi-level Data Pruning (BDP) paradigm that targets the weights and architecture levels of DARTS to enhance efficiency.
Comprehensive evaluations show that BDP reduces search costs by over 50% while achieving superior performance.

Read Full Article

12 Likes

Medium

Revision Research on Data Pruning part4(Machine Learning)

The over-parameterized pre-trained models pose a challenge to fine-tuning with limited computation resources.
A series of training-based scoring functions are proposed to quantify the informativeness of the data subset, but the pruning cost becomes non-negligible.
Adapting geometric-based methods for efficient pruning leads to inferior performance.
The proposed learning complexity scoring function achieves state-of-the-art performance in classification tasks and instruction fine-tuning of large language models.

Read Full Article

4 Likes

Medium

337

Revision Research on Data Pruning part3(Machine Learning)

Data pruning is an attractive field of research due to the increasing size of datasets used for training neural networks.
Most current data pruning algorithms have limitations in preserving accuracy compared to models trained on the full data.
This paper explores the application of data pruning with knowledge distillation (KD) when training on a pruned subset.
Using KD, simple random pruning is shown to be comparable or superior to sophisticated pruning methods across all pruning regimes.

Read Full Article

20 Likes

Medium

254

Revision Research on Data Pruning part2(Machine Learning)

Data pruning is essential to mitigate costs in deep learning models by removing redundant or uninformative samples.
Existing data pruning algorithms can produce highly biased classifiers.
Random data pruning with appropriate class ratios has the potential to improve worst-class performance.
A "fairness-aware" approach to pruning is proposed and empirically demonstrated to improve robustness without significant drop in average performance.

Read Full Article

15 Likes

Medium

Different Distillation methods in Machine Learning research part4

In multi-modal learning, influential modalities are crucial for high accuracy classification/segmentation.
A novel approach called Meta-learned Cross-modal Knowledge Distillation (MCKD) is proposed to address this issue.
MCKD dynamically estimates the importance weight of each modality through meta-learning.
Experimental results show that MCKD outperforms current state-of-the-art models in multiple tasks.

Read Full Article

Medium

Different Distillation methods in Machine Learning research part3

Researchers propose GLiRA, a distillation-guided approach to membership inference attacks on black-box neural networks.
The study explores the connection between vulnerability to membership inference attacks and distillation-based functionality stealing attacks.
Knowledge distillation significantly improves the efficiency of likelihood ratio membership inference attacks, especially in the black-box setting.
The proposed method outperforms the current state-of-the-art membership inference attacks in the black-box setting across multiple image classification datasets and models.

Read Full Article

2 Likes

Medium

348

Research on Stochastic Bridge part6(Machine Learning 2024)

Researchers have developed a sampling method to quantify rare events in stochastic processes.
The method constructs stochastic bridges, which are trajectories with fixed start and end points.
By carefully choosing and weighting these bridges, the method focuses processing power on rare events while preserving their statistics.
The method is compared to the Wentzel-Kramers-Brillouin (WKB) optimal paths and shows accuracy when noise levels are low.

Read Full Article

20 Likes

Medium

Research on Stochastic Bridge part5(Machine Learning 2024)

Stochastic Bridge part5 (Machine Learning 2024) introduces the Stochastically Extended Adversarial (SEA) model as an interpolation between stochastic and adversarial online convex optimization.
They investigate the theoretical guarantees of optimistic online mirror descent (OMD) for the SEA model and derive regret bounds for convex, strongly convex, and exp-concave functions.
For convex and smooth functions, the regret bound is O(σ21:T−−−√+Σ21:T−−−−√), while for strongly convex and smooth functions, it is O((σ2max+Σ2max)log(σ21:T+Σ21:T)).
They also propose novel algorithms to handle non-smooth and convex functions in the SEA model, achieving provable static regret and dynamic regret guarantees without smoothness conditions.

Read Full Article

3 Likes

For uninterrupted reading, download the app