Achieving generalization of deep learning models in a quick way by adapting T-HTR learning rate scheduler

Vidyabharathi, D; Mohanraj,; Kumar, J Senthil; Suresh, Yeresime

doi:10.1007/s00779-021-01587-4

Cited by 8 publications

(4 citation statements)

References 16 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Another insight for future work is solving the challenge of selecting the optimal learning rate, as excessively high values can cause divergence, while overly low values hinder fast convergence, significantly affecting domain adaptation quality. Moreover, monotonously decreasing learning rates may trap neural networks in suboptimal states, possibly causing premature stagnation [53]. Implementing an attenuation strategy, which involves dynamically adjusting the learning rate based on gradient observations [53] or utilizing reinforcement learning agents to adapt it to observed states [54], is essential for robustness.…”

Section: Future Work and Conclusionmentioning

confidence: 99%

“…Moreover, monotonously decreasing learning rates may trap neural networks in suboptimal states, possibly causing premature stagnation [53]. Implementing an attenuation strategy, which involves dynamically adjusting the learning rate based on gradient observations [53] or utilizing reinforcement learning agents to adapt it to observed states [54], is essential for robustness. Although not employed in our current research due to its relatively uncommon use and tool limitations, we acknowledge its importance.…”

Section: Future Work and Conclusionmentioning

confidence: 99%

See 1 more Smart Citation

Efficient Sim-to-Real Transfer in Reinforcement Learning Through Domain Randomization and Domain Adaptation

Shakerimov,

Alizadeh,

Varol

2023

IEEE Access

View full text Add to dashboard Cite

Reinforcement learning has gained significant interest in modern industries for its advancements in tackling challenging control tasks compared to rule-based programs. However, the robustness aspect of this technique is still under development, limiting its widespread adoption. This problem has become more pronounced as users switch to training simulations to reduce costs, resulting in a reality gap that negatively affects real-world performance. One popular method employed to mitigate this problem is randomizing uncertain parameters of the environment during training. Nevertheless, this approach requires expert knowledge to determine the appropriate range of randomization. On the other hand, there is a technique that introduces fine-tuning of agents by adapting their policies to new environments. However, it is challenging to adapt policies when the new environment requires shifting them off their initial distribution. These obstacles limit the practical utilization and popularization of both techniques. Our study proposes a hybrid approach that handles these issues by fine-tuning agents trained with domain randomization through additional real-world training. To assess the efficacy of our approach, we conducted experiments involving a rotary inverted pendulum, augmented with an extra weight not represented in the simulation. Additionally, we employed simulated environments including a cart pole, a simple pendulum, a quadruped, and an ant robot scenario. These environments were given in two distinct versions with mismatching parameters to imitate a gap between training and testing conditions. The results demonstrate that adding as few as twenty to fifty additional real-world training episodes can significantly enhance the performance of agents trained with domain randomization. Moreover, including fifty to two hundred additional episodes can elevate it to a level comparable to those fully trained in the real world. Our study concludes that achieving efficient simulation-to-reality transfer is feasible with domain randomization and relatively small amounts of realworld training.

show abstract

Section: Future Work and Conclusionmentioning

confidence: 99%

Section: Future Work and Conclusionmentioning

confidence: 99%

Efficient Sim-to-Real Transfer in Reinforcement Learning Through Domain Randomization and Domain Adaptation

Shakerimov,

Alizadeh,

Varol

2023

IEEE Access

View full text Add to dashboard Cite

show abstract

“…The cyclical learning rate includes base learning rate and max learning rate, cycle, step size, batch size, batch, or iteration. Base learning rate and max learning rate define the boundaries of a range, where the learning rate will fluctuate (Vidyabharathi et al, 2021). The value of base learning rate and max learning rate was set as 10 -8 and 10, respectively.…”

Section: Cyclical Learning Ratementioning

confidence: 99%

Applying machine learning to fine classify construction and demolition waste based on deep residual network and knowledge transfer

Lin

Zhao

Ting-ting

et al. 2022

Environ Dev Sustain

View full text Add to dashboard Cite

Few studies reported using the convolutional neural network with transfer learning to finely classify the construction and demolition waste. Objectives: This study aims to develop a highly efficient method to realize the finely sorting the construction and demolition waste, which is a key step for promoting the recycling system to realize carbon neutrality in the waste management sector. Methodology: C&DWNet models, ResNet structures based on knowledge transfer and cyclical learning rate, were proposed to classify ten types of construction and demolition waste. Indexes (confusion metric, accuracy, precision, recall, F1 score, sensitivity, specificity and kappa) were adopted to evaluate the performance of various C&DWNet models. Results: Knowledge transfer can reduce the training time and improve the performance of the C&DWNet model. The average training time is increased with the increase of the layer of C&DWNet architecture from C&DWNet-18 (946.7 s) to C&DWNet-152 (1186.6 s). The accuracy of various C&DWNet models is approximately 72~74%, the best accuracy is 73.6% in C&DWNet-152. C&DWNet-18 is more suitable for the classification of construction and demolition waste in terms of training time, accuracy, precision, and F1 score. Moreover, the t-distributed stochastic neighbor embedding can distinctly separate each type of construction and demolition waste. Improvement: The environmental applications and limitations of the C&DWNet module were also discussed, which could provide a reference for the intelligent management of construction and demolition waste and promote the development of the circular economy.

show abstract

“…Abid et al [6] presented an optimized homomorphic encryption Chinese remainder theorem with a Rivest-Shamir-Adleman algorithm for an efficient and secure communication in the current digital world. Correspondingly, the seventh article "Achieving generalization of deep learning models in a quick way by adapting T-HTR learning rate scheduler" by Vidyabharathi et al [7] implemented a new HTR learning rate scheduler (toggle between hyperbolic tangent decay and triangular mode with restarts) in order to identify an optimal learning rate in deep neural network. Compared to the traditional methods, the proposed learning rate scheduler consumed minimal time for every epoch in many cases.…”

Section: Submission and Summary Of Contributionsmentioning

confidence: 99%

Ultra-low power wearables

Parameshachari

Rocha

Fung

2021

Pers Ubiquit Comput

View full text Add to dashboard Cite

Achieving generalization of deep learning models in a quick way by adapting T-HTR learning rate scheduler

Cited by 8 publications

References 16 publications

Efficient Sim-to-Real Transfer in Reinforcement Learning Through Domain Randomization and Domain Adaptation

Efficient Sim-to-Real Transfer in Reinforcement Learning Through Domain Randomization and Domain Adaptation

Applying machine learning to fine classify construction and demolition waste based on deep residual network and knowledge transfer

Ultra-low power wearables

Contact Info

Product

Resources

About