Concept-Driven Continual Learning

Sin-Han Yang 1, Tuomas Oikarinen 2, Tsui-Wei (Lily) Weng 2
1 National Taiwan University, 2 UC San Diego
TMLR 2024
description Paper code Code

Abstract

This paper introduces two novel solutions to the challenge of catastrophic forgetting in continual learning: Interpretability Guided Continual Learning (IG-CL) and Intrinsically Interpretable Neural Network (IN2). These frameworks bring interpretability into continual learning, systematically managing human-understandable concepts within neural network models to enhance knowledge retention from previous tasks. Our methods are designed to enhance interpretability, providing transparency and control over the continual training process. While our primary focus is to provide a new framework to design continual learning algorithms based on interpretability instead of improving performance, we observe that our methods often surpass existing ones: IG-CL employs interpretability tools to guide neural networks, showing an improvement of up to 1.4% in average incremental accuracy over existing methods; IN2, inspired by the Concept Bottleneck Model, adeptly adjusts concept units for both new and existing tasks, reducing average incremental forgetting by up to 9.1%. Both our frameworks demonstrate superior performance compared to exemplar-free methods, are competitive with exemplar-based methods, and can further improve their performance by up to 18% when combined with exemplar-based strategies. Additionally, IG-CL and IN2 are memory-efficient as they do not require extra memory space for storing data from previous tasks. These advancements mark a promising new direction in continual learning through enhanced interpretability.

Methods

(I) Interpretability Guided Continual Learning (IG-CL)

IG-CL uses an interpretability tool [1][3] to steer the learning process, identifying and preserving previously learned concepts to prevent forgetting. In the IG-CL framework, we propose 3 steps to identify and preserve concepts:
  1. Decipher network - Use the interpretability tool to identify concept units in the last feature extractor layer, where it has high-level concepts related to classes.
  2. Freezing the learned knowledge - Use Breadth First Search (BFS) to find neurons that are connected to concept units, which we can them subnetwork. Afterwards, we freeze the weights of these subnetworks.
  3. Learning without forgetting - Train the model by regularizing the final prediction layer WF, while also encouraging sparsity in the feature extractor Wl. This ensures that the subnetworks remain small, allowing the model to efficiently learn new tasks:

Fig 1: Overview of of our Method 1: IG-CL.


(II) Intrinsically Interpretable Neural Network (IN2)

Our second method, IN2, is a novel framework based on the Concept Bottleneck Model (CBM) [2][4]. IN2 uniquely maintains previously learned concept units while integrating new ones for upcoming tasks. Compared with IG-CL, IN2 does not need external interpretability tools since it inherits CBM's interpretability. We propose 4 steps to transform CBMs into IN2:

  1. Concept set expansion - When learning a new task, add concepts related to new classes into the Concept Bottleneck Layer. Concepts are generated from large language models [2] or provided from the dataset [4].
  2. Learning the concept mapping - To preserve learned concepts, we freeze the weights of the existing concepts from the concept mapping of the previous task.
  3. Learning the prediction layer - Train the model by regularizing the final prediction layer WF.

Fig 2: Overview of our Method 2: IN2.


Experimental Results

Continual Learning Performance

The tables below evaluate our method against continual learning baselines: [5] (EWC), [6] (GEM), [7] (SI), [8] (LwF), [9] (Adam-NSCL) and [10] (MIR). The results show that our methods frequently outperform baselines. IG-CL showing an improvement of up to 1.4% in average incremental accuracy (AT) over existing methods, and IN2 reducing average incremental forgetting (FT) by up to 9.1%. Furthermore, IG-CL and IN2 can improve exemplar-based methods' performance by up to 18% when combined with exemplar-based methods.


Table 1: IG-CL performance, compared with exemplar-free methods and combined with exemplar-based methods.


Table 2: IN2 performance, compared with exemplar-free methods and combined with exemplar-based methods..


Interpretability

1. IG-CL

We analyze the concept units in the last feature extractor layer of IG-CL to evaluate the concept evolution in the continual learning process. The results in Table 3 shows that concepts are preserved after learning unrelated new tasks, which gives us insight on how IG-CL helps avoid catastrophic forgetting.

Table 3: Concept evolution for IG-CL in freeze-all implementation. The blue concept means the concept is related to the current task, while the green concept means it is unrelated. "x" stands for non-interpretable units. The results show that IG-CL can preserve the concepts from previous classes.



2. IN2

We analyze the final prediction weights of IN2 in the continual learning process. Figure 3 demonstrates that IN2 effectively preserves the contributions of concepts from previous tasks, indicating its strength in retaining and utilizing previously learned knowledge. This ability is crucial in mitigating catastrophic forgetting.

Fig 3: Final weight visualization of IN2. Concepts generated from the caterpillar class itself are colored blue, and other concepts from task 1 are colored gray.



Cite this work

S.-H. Yang, T. Oikarinen and T.-W. Weng, Concept-Driven Continual Learning , TMLR 2024. 
            
    @article{yang2024conceptdriven,
        title={Concept-Driven Continual Learning},
        author={Yang, Sin-Han and Oikarinen, Tuomas and Weng, Tsui-Wei},
        journal={Transactions on Machine Learning Research},
        year={2024}
        }
This webpage template was recycled from here.

Accessibility