Expert Gate: Lifelong Learning with a Network of Experts (1611.06194v2)

Abstract: In this paper we introduce a model of lifelong learning, based on a Network of Experts. New tasks / experts are learned and added to the model sequentially, building on what was learned before. To ensure scalability of this process,data from previous tasks cannot be stored and hence is not available when learning a new task. A critical issue in such context, not addressed in the literature so far, relates to the decision which expert to deploy at test time. We introduce a set of gating autoencoders that learn a representation for the task at hand, and, at test time, automatically forward the test sample to the relevant expert. This also brings memory efficiency as only one expert network has to be loaded into memory at any given time. Further, the autoencoders inherently capture the relatedness of one task to another, based on which the most relevant prior model to be used for training a new expert, with finetuning or learning without-forgetting, can be selected. We evaluate our method on image classification and video prediction problems.

• The paper presents the Expert Gate framework that sequentially integrates new tasks without the need to store previous task data.
• It employs gating autoencoders to automatically forward test samples to the most relevant expert based on learned task relatedness.
• Empirical evaluations on image classification and video prediction demonstrate efficient performance and scalability over joint model training.

Introduction

Lifelong learning in AI has accelerated the evolution of models capable of adapting to new tasks sequentially. The traditional approach, where models are trained on multiple tasks jointly or each task is trained with separate specialized models, comes with significant trade-offs. It is in this regard that the concept of a Network of Experts (NoE) is introduced, leveraging prior knowledge without encumbering the model with the storage of massive amounts of data from previous tasks.

Network of Experts

The crux of the NoE model lies in its capability to assimilate new tasks or experts sequentially, constructing a knowledge base that is built upon what was learned previously. A unique characteristic of this methodology is that it negates the necessity to store data from previous tasks, which is a practical step towards scalability. A pivotal aspect unaddressed in literature, and the focus of this paper, is the decision-making process regarding the selection of the appropriate expert at test time.

Gating Autoencoders

To automate the selection of the appropriate expert for a given test sample, the researchers introduce a set of gating autoencoders. Their key function is to learn a representation for the current task. At test time, these autoencoders automatically forward the test sample to the relevant expert. This gating mechanism not only ensures memory efficiency but also implicitly captures task relatedness — an essential factor in determining the selection of a relevant prior model for training the new expert.

Methodology and Evaluation

The proposed lifetime learning method, referred to as 'Expert Gate,' incorporates strategies such as under-complete autoencoding for task recognition, leveraging task representations to evaluate relatedness, and subsequently choosing the most relevant expert model for a new task. Evaluated on image classification and video prediction problems, Expert Gate circumvents the issues faced with joint model training and demonstrates superior performance in expert selection. Critically, it automatically assigns test samples to the relevant tasks, akin to a discriminative classifier trained on all past task data — a notable achievement considering Expert Gate does not access previous task data.

In conclusion, the presented approach exhibits the potential to streamline the process of lifelong learning in AI systems, providing a scalable and efficient method for integrating and utilizing prior knowledge in a sequence of learning tasks. The findings and methodologies outlined open pathways for future research aiming to refine mechanisms of knowledge transfer and optimal utilization of task relatedness.