Scheduled Sampling: A technique to improve sequence generation in machine learning models by mitigating discrepancies between training and testing phases.
Scheduled Sampling is a method used in sequence generation problems, particularly in auto-regressive models, which generate output sequences one discrete unit at a time. During training, these models use a technique called teacher-forcing, where the ground-truth history is provided as input. However, at test time, the ground-truth is replaced by the model's prediction, leading to discrepancies between training and testing. Scheduled Sampling addresses this issue by randomly replacing some discrete units in the history with the model's prediction, bridging the gap between training and testing conditions.
Recent research in Scheduled Sampling has focused on various aspects, such as parallelization, optimization of annealing schedules, and reinforcement learning for efficient scheduling. For instance, Parallel Scheduled Sampling enables parallelization across time, leading to improved performance in tasks like image generation and dialog response generation. Another study proposes an algorithm for optimal annealing schedules, which outperforms conventional scheduling schemes. Furthermore, Symphony, a scheduling framework, leverages domain-driven Bayesian reinforcement learning and a sampling-based technique to reduce training data and time requirements, resulting in better scheduling policies.
Practical applications of Scheduled Sampling can be found in various domains. In image generation, it has led to significant improvements in Frechet Inception Distance (FID) and Inception Score (IS). In natural language processing tasks, such as dialog response generation and translation, it has resulted in higher BLEU scores. Scheduled Sampling can also be applied to optimize scheduling in multi-source systems, where samples are taken from multiple sources and sent to a destination via a channel with random delay.
One company case study involves Symphony, which uses a domain-driven Bayesian reinforcement learning model for scheduling and a sampling-based technique to compute gradients. This approach reduces both the amount of training data and the time required to produce scheduling policies, significantly outperforming black-box approaches.
In conclusion, Scheduled Sampling is a valuable technique for improving sequence generation in machine learning models by addressing discrepancies between training and testing phases. Its applications span various domains, and ongoing research continues to enhance its effectiveness and efficiency.

Scheduled Sampling
Scheduled Sampling Further Reading
1.AutoSampling: Search for Effective Data Sampling Schedules http://arxiv.org/abs/2105.13695v1 Ming Sun, Haoxuan Dou, Baopu Li, Lei Cui, Junjie Yan, Wanli Ouyang2.REX: Revisiting Budgeted Training with an Improved Schedule http://arxiv.org/abs/2107.04197v1 John Chen, Cameron Wolfe, Anastasios Kyrillidis3.Parallel Scheduled Sampling http://arxiv.org/abs/1906.04331v2 Daniel Duckworth, Arvind Neelakantan, Ben Goodrich, Lukasz Kaiser, Samy Bengio4.Bilateral Teleoperation of Multiple Robots under Scheduling Communication http://arxiv.org/abs/1804.04290v1 Yuling Li, Kun Liu, Wei He, Yixin Yin, Rolf Johansson, Kai Zhang5.Variational Optimization of Annealing Schedules http://arxiv.org/abs/1502.05313v2 Taichi Kiwaki6.Feedback Scheduling of Priority-Driven Control Networks http://arxiv.org/abs/0806.0130v1 Feng Xia, Youxian Sun, Yu-Chu Tian7.Inductive-bias-driven Reinforcement Learning For Efficient Schedules in Heterogeneous Clusters http://arxiv.org/abs/1909.02119v2 Subho S Banerjee, Saurabh Jha, Zbigniew T. Kalbarczyk, Ravishankar K. Iyer8.Age-optimal Sampling and Transmission Scheduling in Multi-Source Systems http://arxiv.org/abs/1812.09463v3 Ahmed M. Bedewy, Yin Sun, Sastry Kompella, Ness B. Shroff9.Scheduling for Cellular Federated Edge Learning with Importance and Channel Awareness http://arxiv.org/abs/2004.00490v2 Jinke Ren, Yinghui He, Dingzhu Wen, Guanding Yu, Kaibin Huang, Dongning Guo10.Smart Sampling for Lightweight Verification of Markov Decision Processes http://arxiv.org/abs/1409.2116v2 Pedro D'Argenio, Axel Legay, Sean Sedwards, Louis-Marie TraonouezScheduled Sampling Frequently Asked Questions
What is scheduled sampling?
Scheduled sampling is a technique used in sequence generation problems, particularly in auto-regressive models, to improve the performance of machine learning models by mitigating discrepancies between training and testing phases. It addresses the issue of teacher-forcing by randomly replacing some discrete units in the input history with the model's prediction, bridging the gap between training and testing conditions.
Why was scheduled sampling introduced?
Scheduled sampling was introduced to address the discrepancies between training and testing phases in sequence generation problems. During training, auto-regressive models use teacher-forcing, where the ground-truth history is provided as input. However, at test time, the ground-truth is replaced by the model's prediction, leading to a mismatch between training and testing conditions. Scheduled sampling helps to reduce this mismatch and improve the model's performance.
What is teacher forcing in deep learning?
Teacher forcing is a technique used in training auto-regressive models, where the ground-truth history is provided as input to the model during the training phase. This approach helps the model learn the correct output sequence by using the actual data as a guide. However, it can lead to discrepancies between training and testing conditions, as the ground-truth history is not available during testing.
How does scheduled sampling improve sequence generation?
Scheduled sampling improves sequence generation by gradually exposing the model to its own predictions during training. By randomly replacing some discrete units in the input history with the model's prediction, the model learns to generate sequences more accurately under testing conditions, where it must rely on its own predictions instead of the ground-truth history.
What are some recent advancements in scheduled sampling research?
Recent research in scheduled sampling has focused on parallelization, optimization of annealing schedules, and reinforcement learning for efficient scheduling. Parallel Scheduled Sampling enables parallelization across time, leading to improved performance in tasks like image generation and dialog response generation. Optimal annealing schedules have been proposed to outperform conventional scheduling schemes. Symphony, a scheduling framework, leverages domain-driven Bayesian reinforcement learning and a sampling-based technique to reduce training data and time requirements, resulting in better scheduling policies.
What are some practical applications of scheduled sampling?
Practical applications of scheduled sampling can be found in various domains, such as image generation, natural language processing tasks like dialog response generation and translation, and optimization of scheduling in multi-source systems. In image generation, it has led to significant improvements in Frechet Inception Distance (FID) and Inception Score (IS). In natural language processing tasks, it has resulted in higher BLEU scores.
Can you provide a case study of a company using scheduled sampling?
One company case study involves Symphony, which uses a domain-driven Bayesian reinforcement learning model for scheduling and a sampling-based technique to compute gradients. This approach reduces both the amount of training data and the time required to produce scheduling policies, significantly outperforming black-box approaches.
Explore More Machine Learning Terms & Concepts