🤖 AI

ACIL: Active Class Incremental Learning for Image Classification

This paper proposes ACIL, a novel active learning framework for class incremental learning that leverages uncertainty and diversity criteria to select informative samples for annotation, thereby significantly reducing annotation costs while effectively mitigating catastrophic forgetting.

Original authors: Aditya R. Bhattacharya, Debanjan Goswami, Shayok Chakraborty

Published 2026-02-05

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Aditya R. Bhattacharya, Debanjan Goswami, Shayok Chakraborty

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are training a very smart robot to recognize different types of animals. In the real world, you can't just dump a million labeled photos of every animal into the robot's brain at once. Instead, the robot learns in episodes: first, it sees a few pictures of cats; then, a few pictures of dogs; then, birds, and so on.

The problem is that as the robot learns about birds, it tends to "forget" what a cat looks like. This is called catastrophic forgetting.

Furthermore, there's a huge cost to this learning process: human annotation. For the robot to learn, a human has to look at every single photo and say, "Yes, that's a cat." If you have a million photos, that's a million hours of human work. Most existing methods assume you label every single photo in every episode, which is incredibly expensive and wasteful, especially since the robot won't see those specific photos again in the future.

The Solution: ACIL (The "Smart Curator")

The authors of this paper propose a new system called ACIL (Active Class Incremental Learning). Think of ACIL as a Smart Curator for a museum.

Here is how it works, using simple analogies:

1. The "Partial Labeling" Rule

In traditional methods, the curator is forced to label every new exhibit that arrives.
In ACIL, the curator is given a small "budget" of time. They don't label everything. Instead, they look at a huge pile of unlabeled photos and only pick the most important ones to label.

The Analogy: Imagine you are packing a suitcase for a long trip, but you only have room for 10 items. You don't just grab random socks. You carefully pick the 10 items that will be most useful for the whole trip. ACIL picks the "most useful" photos to label.

2. The "Memory Bank" (Exemplars)

When the robot finishes learning about cats, it doesn't throw away the photos. It keeps a tiny, special collection of the best cat photos in a "Memory Bank."

The Twist: When the robot starts learning about dogs, it doesn't just look at the new dog photos. It also looks at its "Memory Bank" of cats.
The Innovation: Most systems only pick new photos from the current pile. ACIL is smart enough to realize: "Wait, I need to keep some of my old cat photos fresh in my memory, too!" So, it splits its labeling budget: some goes to picking the best new dog photos, and some goes to picking the best old cat photos to re-verify.

3. The Selection Strategy (Uncertainty & Diversity)

How does ACIL decide which photos to pick? It uses two rules:

Uncertainty: It looks for photos where the robot is confused. "Is this a cat or a dog?" If the robot is unsure, a human needs to label it so the robot learns.
Diversity: It ensures the selected photos are all different from each other. It doesn't pick 10 photos of the exact same cat sitting in the same pose. It picks a cat sleeping, a cat running, and a cat eating. This gives the robot a well-rounded view.

The Results: Saving Time Without Losing Smarts

The paper tested this on six different image datasets (like MNIST, CIFAR, and Tiny ImageNet). Here is what they found:

Massive Savings: Because ACIL only labels a tiny fraction of the photos (the "exemplars") instead of everything, it reduced the human effort (annotation cost) by a huge amount. For example, on one dataset, it saved about 4 times the amount of work compared to standard methods.
No Forgetting: Despite labeling so few photos, the robot didn't forget the old classes. It performed just as well as the expensive methods that labeled everything.
Better than "Random" Picking: When compared to other "Active Learning" methods that just pick random or simple samples, ACIL was much smarter and more accurate.

In a Nutshell

ACIL is a way to teach a robot new things over time without making humans do all the boring work of labeling every single picture. It acts like a wise librarian who knows exactly which few books to keep on the shelf to remember the whole library, ensuring the robot stays smart about old things while learning new things, all while saving a massive amount of human time and money.

Technical Summary: ACIL – Active Class Incremental Learning for Image Classification

1. Problem Statement

The paper addresses the challenge of Class Incremental Learning (CIL) in computer vision, a scenario where deep neural networks (DNNs) must learn new classes sequentially from episodic data while retaining knowledge of previously learned classes. A primary obstacle in CIL is catastrophic forgetting, where optimizing the network for new classes erases knowledge of old ones.

Existing CIL research primarily focuses on mitigating forgetting, often assuming that all training samples in every episode are fully annotated. The authors argue this assumption is unrealistic and inefficient for two reasons:

High Annotation Cost: Manual annotation is expensive in terms of time, labor, and expertise.
Annotation Wastage: In episodic settings, most data from a current episode is inaccessible in future episodes. Annotating all samples in an episode is therefore a waste of effort, as only a subset of these samples (exemplars) will be needed to prevent forgetting in subsequent steps.

The paper proposes a new setting where, in each episode, only a small portion of the data is annotated, while the majority remains unlabeled. The goal is to develop a framework that selects a minimal set of exemplar samples to be annotated and propagated to future episodes, thereby reducing annotation costs while effectively preventing catastrophic forgetting.

2. Methodology: The ACIL Framework

The authors propose ACIL (Active Class Incremental Learning), a framework that integrates Active Learning (AL) strategies into the CIL pipeline. The core mechanism involves selecting an exemplar set of a fixed size $k$ in each episode to be manually annotated and appended to the data of the next episode.

2.1 Problem Setup

Data Flow: Training data arrives in $N$ sequential episodes. Each episode $n$ contains a small labeled set ( $X^L_n$ ), a large unlabeled set ( $X^U_n$ ), and an exemplar set ( $E_{n-1}$ ) from the previous episode.
Constraint: The model is trained on $X^L_n$ and $E_{n-1}$ . From the current episode, the system must select a new exemplar set $E_n$ of size $k$ to be annotated.
Selection Source: $E_n$ is constructed by selecting samples from $X^U_n$ (unlabeled current data), $E_{n-1}$ (previous exemplars), and potentially $X^L_n$ (though the authors note $X^U_n$ is prioritized due to size).

2.2 Budget Splitting Strategy

To balance learning new classes and retaining old knowledge, ACIL splits the annotation budget $k$ between the current unlabeled set ( $X^U_n$ ) and the previous exemplar set ( $E_{n-1}$ ).

The split is proportional to the number of classes present in each set ( $|C_{episode}|$ vs. $|C_{exemplar}|$ ).
Formula:
$k_{unlabeled} = \left( \frac{|C_{episode}|}{|C_{episode}| + |C_{exemplar}|} \right) \cdot k$
$k_{exemplar} = \left( \frac{|C_{exemplar}|}{|C_{episode}| + |C_{exemplar}|} \right) \cdot k$
This ensures that the exemplar set $E_n$ maintains a representation of all classes seen so far, mitigating forgetting.

2.3 Active Sampling Strategy

The selection of specific samples within the allocated budget is driven by a criterion combining uncertainty and diversity.

Uncertainty: Measured using Shannon's entropy of the model's predictions.
Diversity: Achieved by partitioning the candidate set into $B$ clusters (where $B$ is the budget) such that the variance within each cluster is minimized.
Optimization: The authors formulate a weighted set partitioning problem to minimize the weighted variance within partitions, where the weight is the prediction uncertainty.
$\text{argmin}_P \sum_{b=1}^{B} \sum_{x \in X_b} I(x) ||F(x) - c_b||^2$
Where $I(x)$ is the uncertainty (entropy) and $F(x)$ is the feature embedding.
Algorithm: An NP-hard problem approximated using Weighted K-Means. The sample closest to the weighted centroid in each partition is selected for annotation.

2.4 Loss Function

The DNN is trained using a combination of:

Weighted Cross-Entropy Loss ( $L_{WCE}$ ): Applied to the labeled data ( $X^L_n \cup E_{n-1}$ ) to handle class imbalance, with weights inversely proportional to class frequency.
Knowledge Distillation Loss ( $L_D$ ): Applied to the exemplar set $E_{n-1}$ to transfer knowledge from the previous model ( $M_{n-1}$ ) to the current model ( $M_n$ ), preventing forgetting.
$L = L_{WCE} + \lambda L_D$

3. Key Contributions

Novel Learning Paradigm: Introduction of a CIL setting where full annotation of episodic data is not assumed, addressing the high cost and inefficiency of current CIL benchmarks.
ACIL Framework: A specific algorithm that selects exemplars based on a dual criterion of uncertainty and diversity, utilizing a budget-splitting strategy to ensure coverage of both new and old classes.
Efficiency vs. Performance: Demonstration that ACIL can achieve performance comparable to state-of-the-art CIL methods (which assume full annotation) while drastically reducing the number of samples requiring human annotation.

4. Experimental Results

The framework was evaluated on six vision datasets: MNIST, SVHN, CIFAR-10, CIFAR-100, COIL, and Tiny ImageNet.

Accuracy: ACIL outperformed Active Learning baselines (Random, Coreset, BADGE) significantly, as those methods failed to retain knowledge from previous episodes by only sampling from the current unlabeled set. ACIL achieved accuracy comparable to, and in some cases slightly better than, full-annotation CIL baselines (iCaRL, GDumb, Rainbow Memory).
Annotation Effort: ACIL resulted in the lowest annotation effort among all methods. For example, on CIFAR-10, ACIL required approximately 4.13 times fewer annotations per episode compared to CIL baselines while maintaining similar accuracy.
Retention: ACIL demonstrated strong retention of knowledge from the first episode, comparable to replay-based CIL methods, validating its ability to mitigate catastrophic forgetting.
Robustness: Experiments varying the exemplar set size (budget) showed that ACIL maintains performance trends consistent with CIL baselines across different memory constraints.

5. Significance and Claims

The paper claims that ACIL offers a practical solution for real-world incremental learning applications where:

Memory is constrained: The exemplar set size is fixed, simulating limited storage.
Annotation is expensive: The framework drastically reduces the human labor required to train models over time.
Data is ephemeral: The ability to learn from limited annotated data in each episode aligns with privacy and storage constraints where past data cannot be fully retained.

The authors conclude that by intelligently selecting a small subset of informative samples to propagate, ACIL successfully balances the trade-off between annotation cost and model performance, making continual learning more feasible for resource-constrained environments. Future work is noted to potentially extend this framework to regression tasks.