Saliency-Guided Hidden Associative Replay for Continual Learning

Thank you for your time and dedication to this review process. We are grateful for the opportunity to benefit from your expertise, and we believe your contributions will enhance the overall quality of the SHARC.

"The paper lacks a comprehensive overview of both regularization-based and dynamic architecture-based continual learning methods. Notably, there is a complete absence of references to the latter category."

A1. Architect-based:

• DER: Dynamically Expandable Representation for Class Incremental Learning CVPR 2021 \
• Progressive Neural Networks 2016 \
• Lifelong Learning with Dynamically Expandable Networks 2017 \
• Progress & Compress: A scalable framework for continual learning PMLR 2018 \
• DyTox: Transformers for Continual Learning With DYnamic TOken eXpansion CVPR 2022

Regularization-based:

• Uncertainty-based Continual Learning with Adaptive Regularization NerulPS 2019 \
• CPR: Classifier-Projection Regularization for Continual Learning ICLR 2020 \
• Adversarial Targeted Forgetting in Regularization and Generative-Based Continual Learning Models IJCNN 2021 \
• Continually Learning Self-Supervised Representations with Projected Functional Regularization CVPRW 2022

"The absence of a thorough examination of other continual learning methods employing brain-inspired memory systems, such as [4-5], leaves a critical gap that requires addressing through comprehensive review, discussion, and comparison."

A2. [4] CLSER
The method described in quote [4], called CLSER, has already been extensively discussed in this paper. We applied our own method, known as SHARC, to the CLSER framework and conducted evaluations on three datasets: CIFAR-10, CIFAR-100, and TinyImageNet. The results clearly demonstrate a significant improvement in performance, particularly when the buffer size is relatively smaller.

[5] SCoMMER
When comparing our method to CLSER, we observe that our method achieves similar results to SCoMMER on the CIFAR-100 dataset. However, our method surpasses SCoMMER by one percentage point in accuracy when the buffer size is set to 200. Conversely, SCoMMER underperforms CLSER on the CIFAR-10 dataset. In contrast, our SHARC method achieves an even higher accuracy, approximately three percentage points higher than CLSER, particularly when the buffer size is 500. This stark contrast fully illustrates the robustness of our method across different datasets and underscores its superiority.

"It is unclear whether methods without SHARC in Tables I and II are also pre-trained on ImageNet."

A3. In our approach, SHARC, we adopt a consistent starting point by replacing the original network structure, such as ER, MER or GEM, with a trained and partially parameterized ResNet. This ResNet serves as the backbone for processing all image data in our framework.

One notable advantage of SHARC is its flexibility in accommodating different methods. Regardless of the specific method employed in combination with SHARC, the main differentiating factors lie in the fixed structure of the original methods and their respective optimal training parameters. By adhering to this consistent framework, we ensure that the primary distinctions among the methods stem from their unique configurations and parameter choices.

This approach allows for a fair and meaningful comparison of different methods within the SHARC framework, as the focus shifts toward analyzing the impact of their specific configurations and parameter settings on overall performance. By isolating these factors, we can better understand the strengths and weaknesses of each method and make informed decisions when applying them to different datasets or problem domains.

"By "combining" SHARC with other methods, what's the network size increased? In general, we can expect improved performance for many tasks simply by increasing the network size. So it's important to know the increase of network size, after combining with SHARC."

A3. Take the method "erbpcn" as an example. In the specific case of theta=0.2, where only 20% of the original image is retained, if we were to deposit all the complete images inside the episodic memory, it would require approximately 7.5 million parameters. However, with our SHARK method, the episodic memory is represented by a single feature map, resulting in a reduction to only about 1.3 million parameters. This optimization effectively drops about 6 million parameters, and the computation involved is only 17% of the original.

Comparatively, the ResNet18 architecture, which is commonly used as a backbone in deep learning models, consists of around 11 million parameters. In contrast, the associative memory network that we introduced contains only about 4 million parameters. By adopting our approach, we achieve a memory consumption reduction of approximately 11% compared to the original tasker, while also obtaining higher accuracy.

These results clearly demonstrate the efficiency and accuracy of our method. By optimizing the episodic memory representation and introducing the associative memory network, we significantly reduce the parameter count and computational load while still achieving superior performance.

"The novelty of the paper is limited as it combines existing methods such as "saliency" and "associative memory" for the CL problem."

A4. In response to the reviewer's comment on the perceived limited novelty of the paper due to the combination of existing methods - saliency and associative memory - for the Continual Learning (CL) problem, we would like to highlight the unique contributions of our work:

1. Innovative Integration for CL: While it's true that both saliency methods and associative memory have been individually explored in the image retrieval literature, their integration in the context of Continual Learning (CL) represents a significant advancement. Continual Learning, with its unique challenges such as catastrophic forgetting and memory efficiency, requires a nuanced approach. Our work is not just a simple combination of these two methods; it's an innovative adaptation to address the specific needs and challenges of CL.

2. Context-Specific Adaptation: The application of these mechanisms to CL is not trivial. It required significant adjustments and innovations to make them suitable for the dynamic and evolving nature of CL. This adaptation includes the development of a novel method for efficient and selective memory encoding and retrieval, which is critical for managing the complex data streams inherent in CL.

3. Unique Problem-Solving Approach: The problem set in CL is fundamentally different from traditional image retrieval. In CL, the model continuously learns from a stream of data while retaining previously acquired knowledge. This poses unique challenges, such as avoiding catastrophic forgetting, which are not present in conventional image retrieval scenarios. Our method addresses these challenges by leveraging saliency and associative memory in a novel way, specifically tailored for CL.

4. Enhanced Performance and Efficiency: The integration of saliency-guided memory encoding and associative memory retrieval in our SHARC framework enhances memory efficiency and reduces the computational overhead, which is a significant contribution to the field of CL. This efficiency is crucial in CL where models need to be highly adaptive yet resource-efficient.

5. Extending Beyond Existing Literature: While prior works have explored these mechanisms separately, our work synergizes them in a novel framework specifically for CL. This not only advances the state-of-the-art in CL but also opens up new avenues for future research in combining these techniques for other complex learning scenarios.

In summary, while our work builds upon existing methods, its novelty lies in the unique integration and adaptation of these methods to address the specific challenges of Continual Learning. This is not merely an application of existing techniques to a new problem but a thoughtful and innovative reimagining of these techniques to create a solution that is greater than the sum of its parts.

Thank you for your time and dedication to this review process. We are grateful for the opportunity to benefit from your expertise, and we believe your contributions will enhance the overall quality of the SHARC.

"The implementation details about how SHARC combines with the other baseline methods are not given. However, this is somewhat mitigated by the inclusion of anonymous codes. I did not inspect the codes, however."

A1. First, we start by replacing the backbone in the original method with a pre-trained and parameter-frozen ResNet. This allows us to leverage the pre-trained weights and architecture of ResNet, which is a popular convolutional neural network. By doing this, we obtain the feature map from the modified ResNet.

Next, we incorporate a simple MLP as the prediction head. The MLP takes the feature map as input and performs the task of making predictions based on the extracted features.

When a batch of data is entered into the system, the first step is to obtain the feature importance using our modified Grad-CAM which is a technique that highlights the important regions in an image that contribute to the predictions made by the model. By applying Grad-CAM on the modified ResNet, we obtain a feature map based on the input data.

After obtaining the feature map, we update the gradients on the current minibatch using this modified Grad-CAM. This step helps us to fine-tune the model based on the specific data in the current batch.

Next, we utilize Grad-CAM masks to mask the feature maps of the images within the batch. Grad-CAM masks allow us to focus on the most relevant parts of the feature maps, enabling us to extract more meaningful information.

Once the feature maps are masked, we process and recover them using BayesPCN. Finally, the recovered feature maps are stored in the episodic memory, which serves as a repository for important information that can be later accessed and utilized for various purposes.

"How does the proposed method combine with the 6 replay-based methods? Does the SHARC Memory Replay module directly replace the respective methods' original replay mechanism? What about the SHARC prediction head? How does it combine with the other methods?"

A2. First, we start by replacing the backbone in the original method with a pretrained and parameter-frozen ResNet. This allows us to leverage the pre-trained weights and architecture of ResNet, which is a popular convolutional neural network. By doing this, we obtain the feature map from the modified ResNet.

Next, we incorporate a simple MLP as the prediction head. The MLP takes the feature map as input and performs the task of making predictions based on the extracted features.

When a batch of data is entered into the system, the first step is to obtain the feature importance using our modified Grad-CAM which is a technique that highlights the important regions in an image that contribute to the predictions made by the model. By applying Grad-CAM on the modified ResNet, we obtain a feature map based on the input data.

After obtaining the feature map, we update the gradients on the current minibatch using this modified Grad-CAM. This step helps us to fine-tune the model based on the specific data in the current batch.

Next, we utilize Grad-CAM masks to mask the feature maps of the images within the batch. Grad-CAM masks allow us to focus on the most relevant parts of the feature maps, enabling us to extract more meaningful information.

Once the feature maps are masked, we process and recover them using BayesPCN. Finally, the recovered feature maps are stored in the episodic memory, which serves as a repository for important information that can be later accessed and utilized for various purposes.

In summary, we made several modifications to the original methods. Firstly, we replaced the backbone of the network with a different architecture. This change allows us to leverage the strengths of the new backbone and potentially improve the performance of the network. Additionally, we modified the form of the data in the episodic memory. This alteration enables us to store and retrieve information in a more efficient and effective manner. By adapting the data structure to better suit our needs, we can enhance the overall performance of the network. Furthermore, we introduced BayesPCN as the associative memory network to achieve strong memory recall performance.

Overall, these changes contribute to improving the performance and memory capabilities of the network, allowing it to handle complex tasks better and make more accurate predictions.

"The experiments in this paper do not seem to be reasonable. The method in this paper introduces an additional associative memory network to store more information for replay, which definitely makes the original method perform better."

A2. The role of our associative memory network is to reduce the memory consumption of whole images by converting them into individual feature maps and depositing them into episodic memory after going through our GradCAM method. This reduction in memory usage is significant.

However, it is important to note that even with the introduction of an associative memory network, the amount of information we introduce is much less compared to methods such as ER and MER, which store images directly into episodic memory.

The reduction in information is due to the transformation of the original image into a feature map representation. Unlike the complete image, which includes all the pixel-level details and color information, a feature map represents a condensed version that highlights specific visual features and patterns. This reduction in information allows for more efficient storage and processing, making it an effective strategy for reducing memory consumption in our associative memory network.

In addition, our associate memory, or BayesPCN, is a compact network architecture comprising only a few layers of Multi-Layer Perceptron (MLP). While BayesPCN plays a crucial role, it is important to note that the memory it provides is relatively limited when compared to the vast amount of information present in the complete image dataset.

"How much memory and computing resources does associative memory take up? Wouldn't it be better if those extra storage resources were used to store more exemplars? Please make a comparison."

A3. Take the method ER with SHARC as an example. In the specific case of theta=0.2, where only 20% of the original image is retained, if we were to deposit all the complete images inside the episodic memory, it would require approximately 7.5 million parameters. However, with our SHARC method, the episodic memory is represented by a single feature map, resulting in a reduction to only about 1.3 million parameters. This optimization effectively drops about 6 million parameters, and the computation involved is only 17% of the original.

Comparatively, the ResNet18 architecture, which is commonly used as a backbone in deep learning models, consists of around 11 million parameters. In contrast, the associative memory network that we introduced contains only about 4 million parameters. By adopting our approach, we achieve a memory consumption reduction of approximately 11% compared to the original tasker, while also obtaining higher accuracy.

These results clearly demonstrate the efficiency and accuracy of our method. By optimizing the episodic memory representation and introducing the associative memory network, we significantly reduce the parameter count and computational load while still achieving superior performance.

“Lack of detailed network and hyper-parameter configuration, especially, for associative memory networks.“

A1. By expanding on the previous list of hyperparameters, we aim to offer a more comprehensive illustration of the parameter settings for our associative memory network. This additional information will enable a deeper analysis and evaluation of the different methods employed. We believe that a thorough parameterization is essential for replicability, comparability, and overall confidence in the results obtained. With this in mind, we present the following comprehensive parameterization for your consideration.

For the different methods, we use the same parameter settings for our associative memory network:

Bpcn_args:

Model Configs

Training Configs

Methods_args:

The authors have not provided a rebuttal, but I did review the comments from the other reviewers. My rating is unchanged.

Q1. The paper does not provide a comprehensive comparison with existing replay-based methods for Continual Learning, making it difficult to assess the superiority of the proposed framework.

A1. Thank you for your suggestion. Additional results have been added to the general response above.

Q2. While the paper introduces a content-focused memory retrieval mechanism, it lacks detailed explanation and analysis of how this mechanism works and its impact on recall performance.

A2. Associative memory is based on attractor dynamics, which refers to neuronal network dynamics dominated by groups of persistently active neurons. Generally, such persistent activation is associated with an attractor state of the dynamics, often in the form of a fixed point. This type of network can be used to implement associative memory by allowing the network's attractors to correspond to the vectors we want to store. This approach supports memory for specific items and differs from semantic memory in that it stores items quickly but does not represent the semantic structure of the data. Instead, attractor dynamics resembles working and episodic memory. Like episodic memory, it acts as an associative memory, returning stored values when triggered with the right cues. The Hopfield network, originally proposed in 1982 (Hopfield, 1982), is a recurrent neural network that implements associative memory using fixed points as attractors. The function of the associative memory is to recognize previously learned input vectors, even in cases where some noise has been added. To achieve this function, every neuron in the network is connected to all the others (see Fig. 2.4(a)). Each neuron outputs discrete values, normally 1 or -1, according to the following equation: x_i(t+1) = sign(Σ_{j=1}^{N} w_{ij}x_j(t))
where x_i(t) is the state of the i-th neuron at time t, and N is the number of neurons. The Hopfield network has a scalar value associated with the state of all neurons x, referred to as the "energy" or Lyapunov function: E(x) = -1/2 Σ_{i=1}^{N} Σ_{j=1}^{N} w_{ij}x_i x_j
If we want to store Q patterns x_p, p=1,2,...,Q, we can use the Hebbian learning rule (Hebb, 1962) to assign the values of the weights as follows: w_{ij} = Σ_{p=1}^{Q} x_p^i x_p^j
This is equivalent to setting the weights to the elements of the correlation matrix of the patterns. Upon presentation of an input to the network, the activity of the neurons can be updated (asynchronously) according to Eq. (2.22) until the energy function has been minimized (Hopfield, 1982). Hence, repeated updates would eventually lead to convergence to one of the stored patterns. However, the network may possibly converge to spurious patterns (different from the stored patterns) as the energy in these spurious patterns is also a local minimum.

I appreciate the author's thorough response to my questions. After revisiting the paper and the rebuttal, I have decided to maintain my current rating.