FAMO: A fast optimization method for multi-task learning (MTL) that uses O(1) space and time to reduce competing gradients.

Multitask learning (MLT) involves training a single model to perform multiple tasks simultaneously, leveraging shared information to improve performance. Although useful, MLT poses challenges in managing large models and optimizing the overall task. Optimizing the average loss may result in suboptimal performance if the task progresses unevenly. Balancing task performance and optimization strategies is important for effective MLT.

Existing solutions to alleviate the underoptimization problem in multitask learning include gradient manipulation techniques. These methods compute a new update vector for the average loss and ensure that all task losses decrease more uniformly. However, while these approaches show performance improvements, they can be computationally expensive for large tasks and model sizes. This is because all task gradients must be computed and stored for each iteration, which adds significant space and time complexity. In contrast, computing the average gradient is more efficient and requires less computational overhead per iteration.

To overcome these limitations, a research team from the University of Texas at Austin, Salesforce AI Research, and Sony AI recently published a new paper. In their research, they introduced Fast Adaptive Multitask Optimization (FAMO). This is a technique designed to address the underoptimization problem in multitask learning without the computational burden associated with existing gradient manipulation techniques.

FAMO dynamically adjusts task weights to ensure balanced loss reduction across tasks by leveraging loss history rather than computing gradients for every task. Key contributions include the introduction of his FAMO, his MTL optimizer with O(1) space and time complexity per iteration, and significant computational efficiency improvements across a variety of his MTL benchmarks. This includes demonstrating performance equal to or better than existing methods.

The proposed approach consists of two main ideas. It's about achieving a balanced loss reduction across tasks and amortizing computations over time.

1. Balanced loss rate improvement:

• FAMO aims to reduce losses for all tasks as evenly as possible. This defines the improvement rate for each task based on the change in loss over time.
• By formulating an optimization problem, FAMO seeks update directions that maximize the worst-case improvement rate across all tasks.

2. Fast approximation with amortization over time:
   • Instead of solving the optimization problem at each step, FAMO performs a single-step gradient descent on parameters representing the task weights and amortizes the computation over the optimization trajectory.
   • This is achieved by updating the task weights based on changes in log loss and approximating the gradient.

In practice, FAMO reparameterizes the task weights to fit within a valid range and introduces regularization to give more weight to recent updates. The algorithm iteratively updates task weights and parameters based on observed losses, finding a balance between task performance and computational efficiency.

Overall, FAMO provides a computationally efficient approach to multitasking optimization by dynamically adjusting task weights and amortizing computations over time. This improves performance without requiring extensive gradient computation.

To evaluate Famo, the authors conducted a demonstration experiment in various experimental settings. They started with a toy two-task problem and demonstrated Famo's ability to efficiently relax competing gradients (CGs). Compared to state-of-the-art methods in the MLT supervised reinforcement learning benchmark, Famo consistently showed good performance. Compared to methods such as NASHMTL, we showed significant efficiency gains, especially in training time. Additionally, ablation studies on the regularization factor γ highlighted Famo's robustness across different settings, with the exception of certain cases like CityScapes, where performance can be stabilized by adjusting γ. This evaluation highlighted Famo's effectiveness and efficiency in a variety of multi-task learning scenarios.

In conclusion, FAMO presents a promising solution to the MLT challenge by dynamically adjusting task weights and amortizing computations over time. This method effectively alleviates the underoptimization problem without the computational burden associated with existing gradient manipulation techniques. Through demonstration experiments, FAMO demonstrated consistent performance improvements across different MLT scenarios, demonstrating its effectiveness and efficiency. With its balanced loss reduction approach and efficient optimization strategy, FAMO provides a valuable contribution to the field of multi-task learning, paving the way for more scalable and effective machine learning models.

Please check paper. All credit for this study goes to the researchers of this project.Don't forget to follow us twitter.Please join us telegram channel, Discord channeland LinkedIn groupsHmm.

If you like what we do, you'll love Newsletter..

Don't forget to join us 41,000+ ML subreddits