TL;DR
Instead of training a model to solve a specific task, this paper proposes a "divergent search" strategy that discovers a wide repertoire of ways to partition unlabeled data. By using a meta-learning framework to accumulate these diverse behaviors into a single model, the system learns generic, robust features that can adapt to new, unseen few-shot classification tasks without ever seeing a human-provided label during pre-training.
What Problem Does This Paper Address?
Most unsupervised learning methods for images rely on "proxy tasks"—like predicting rotations or clustering—to create artificial labels. These methods have a strong "convergent" bias: they try to force the model to learn features that look like human-defined classes.
However, in the real world, we often don't know the target task a priori. An image of a dog in a park could be classified by breed, pose, the type of trees in the background, or the weather. Convergent algorithms usually settle on one of these and ignore the others. The authors tackle the challenge of Divergent Discriminative Feature Accumulation (DDFA) in deep networks—essentially, how do we systematically search for and "archive" as many different ways of looking at data as possible?
Methods
The authors build their method on Reptile, a first-order meta-learning algorithm, but replace the supervised "tasks" with a search for novelty. The framework operates in two loops:
- The Inner Loop (The Search):
- The Archive: A model that represents everything the system already knows how to do. Its final layer acts as an "index" to past behaviors.
- The Teacher: A model tasked with finding a "New Behavior"—a way of partitioning the current batch of images that the Archive cannot currently replicate. It does this by maximizing the Jensen-Shannon (JS) divergence between its outputs and the Archive's outputs.
- The Learner: Once a novel behavior is found, a Learner model is trained to master that specific behavior.
- The Outer Loop (The Integration):
- The gradients from multiple Learners (each having found a different novel behavior) are aggregated.
- Crucially, the authors use a weighted average based on Novelty (how hard it was for the archive to match the behavior) and Quality/Robustness (how well the behavior survives input noise/dropout).
- This update moves the global model parameters toward a state that can easily adapt to any of these discovered behaviors.
Datasets
The method was evaluated on two standard few-shot learning benchmarks:
- Omniglot: A "transpose" of MNIST containing 1,623 handwritten characters from 50 alphabets.
- Mini-ImageNet: A subset of the massive ImageNet dataset, containing 100 classes with 600 images each, featuring complex natural scenes.
Key Insights and Ideas
- Novelty as an Objective: Unlike clustering, which seeks a "correct" partition, divergent search treats any partition that is different from what has been seen before as potentially valuable.
- Archive vs. Teacher: The tension between the Archive (trying to match the teacher) and the Teacher (trying to escape the archive) creates a dynamic "frontier" of learning.
- Generic vs. Specific Features: While supervised models learn to group images by class (e.g., "all lions together"), the divergent search model learns more generic features like spatial frequency, pose, or image style.
- Avoiding the "Random Label" Trap: Simply training on random labels actually hurts performance because it forces the model to memorize noise. Divergent search avoids this because the "tasks" it creates are grounded in the model's current capabilities—they are "hard but possible" to learn.
Key Results
The authors compared their "Proposed" divergent search against several baselines in a 5-way, 5-shot setting:
- Mini-ImageNet: The proposed method achieved 43.41% accuracy, significantly outperforming Random Initialization (36.52%) and Fixed Random Labels (32.49%). It performed nearly identically to a "Low-Augmentation" version of UMTRA (an unsupervised meta-learning baseline).
- Omniglot: The proposed method reached 77.85% accuracy, a massive jump over Random Initialization (64.29%).
- Qualitative Findings: t-SNE visualizations (Figure 4 in the paper) show that while supervised models cluster by object category, the divergent search model clusters by visual properties like the "busyness" of an image or the orientation of an object.
Limitations and Future Work
- Performance Gap: As expected, divergent search is still outperformed by models that have a "strong bias" toward the evaluation tasks (like fully supervised models or UMTRA with heavy data augmentation).
- Behavior Space Complexity: In high-dimensional spaces (like deep conv-nets), many regions of the behavior space correspond to "nonsense" partitions. Refining the "quality" metric to better distinguish between "novel-but-useful" and "novel-but-garbage" is a key next step.
- Model Capacity: The current study keeps the model size small for fair comparison. Future work could explore how larger architectures might better "store" a more massive repertoire of behaviors.
Summary updated