The AI That Learns When There Isn’t Enough to Learn From

The easiest photographs for artificial intelligence are the ones the world has already made millions of times. Cats. Cars. Faces. Street signs. The hard photographs are often the ones that matter to somebody in a very specific place: a hairline defect on a factory component, two fish species separated by a small fin pattern, a medically important feature that appears in only a few hundred scans, or an abnormal arrangement of otherwise ordinary objects.

That is the practical problem behind a body of computer-vision work at the University of Fukui led by associate professor Tatsuhito Hasegawa. The modern AI boom has been powered by abundance—more data, larger models, more accelerators, longer training runs. Fukui's work asks what happens when the application lives on the other side of that equation. What if the labels are expensive? What if the rare class is genuinely rare? What if the model has to fit on an edge device? What if a small manufacturer cannot build a data center just to inspect a part?

The university's current research program, “Fine-Grained Image Recognition Using Foundation Models and Generative AI,” organizes the challenge across three connected fronts: deep-learning theory including industrial anomaly detection and ensemble learning; implementation for extreme edge AI; and fine-grained recognition in primary industries such as agriculture and fisheries. Around that program, Hasegawa's laboratory has developed a set of complementary methods that attack scarcity from different directions.

The bargain that made modern vision powerful

Computer vision did not begin with deep learning. For decades, researchers hand-designed features that told computers to look for edges, corners, textures and shapes. The approach could work extremely well, but engineers had to decide in advance what kinds of visual evidence mattered.

The deep-learning turn changed the bargain. Instead of explicitly describing every useful feature, researchers trained multilayer neural networks to learn representations from examples. The most famous early demonstration was AlexNet in the 2012 ImageNet competition. It trained on roughly 1.2 million labeled images, used two NVIDIA GTX 580 GPUs and ran for five to six days. At the time that sounded enormous. By the standards of the foundation-model era, it looks almost quaint.

The pattern kept scaling. Bigger datasets and faster accelerators made it possible to train systems whose learned representations transferred to many tasks. In 2021, CLIP—Contrastive Language-Image Pre-training—showed what happens when images are paired with natural-language descriptions at web scale. The original paper reported training on 400 million image-text pairs. Instead of being built only to classify a fixed list of labels, the model learned a broad correspondence between visual and linguistic concepts and could transfer to many downstream tasks with little or no task-specific training.

That created a new kind of capital in AI: a pretrained representation. A small laboratory or company may never reproduce the dataset and compute used to create a foundation model, but it may be able to borrow what that model already learned.

The scarce-data problem is therefore no longer only “How do we collect a million examples?” It can also become “How do we extract the most useful knowledge from models that have already seen a world of examples?”

Route one: manufacture variation—but do not manufacture truth

The oldest trick in the modern deep-learning toolbox is data augmentation. If a training set is small, an image can be cropped, flipped, rotated, recolored or otherwise transformed to create additional examples. The hope is not to fool the model with duplicates, but to teach it which changes should not alter the answer.

Generative AI expands that idea dramatically. Instead of rotating a photograph of one fish, a model can attempt to create a new fish image with a different background, pose or lighting condition. In theory, this is exactly what fine-grained recognition needs. The rare examples remain rare in the physical world, but the training set becomes more varied.

Fukui researchers tested that premise in fine-grained image recognition, where the target classes can differ by extremely small visual details. Takuya Kato, Shion Serizawa, Mitsuki Okayama, Yuta Nakano and Hasegawa investigated image-generation models for data augmentation. They varied backgrounds, used multimodal models to put visual features into prompts, and built a background-replacement pipeline.

The result is more useful because it was not uniformly positive. Text-to-image generation struggled to reproduce the detailed features needed to distinguish similar classes. It could nevertheless improve accuracy in one-shot learning, where only one real training example is available for a class. Background replacement also outperformed baseline approaches under some conditions.

That is the central discipline of synthetic data: variation is valuable only if the generated image preserves the label. A model trained to distinguish two closely related fish cannot learn from an image generator that quietly changes the fin, stripe or body shape that defines the species. More images can make the dataset larger while making the knowledge worse.

The laboratory has pushed the idea further into fish identification. Hasegawa and Yuta Nakano presented foundation-model-based augmentation for fine-grained fish species recognition at IEEE GCCE 2024, work that received a presentation award. The subject is not accidental. Fisheries are a natural example of the mismatch between AI's hunger and the real world's data: catches vary by region and season, uncommon species may have few labeled photographs, and practical systems may need to work close to the dock rather than in a cloud laboratory.

Route two: let the giant teacher educate a smaller student

A second route attacks the model rather than the dataset. Knowledge distillation is an old idea by deep-learning standards. A large “teacher” model contains useful information in the pattern of its outputs and internal representations. A smaller “student” can be trained not only on hard labels—“this is class A”—but on the teacher's richer judgments about similarity and uncertainty.

Geoffrey Hinton, Oriol Vinyals and Jeff Dean popularized the modern formulation in 2015, partly as a way to make powerful ensembles easier to deploy. The attraction has only grown as foundation models have become larger. Training a huge model once is one problem. Carrying it everywhere, on every camera, scanner or edge computer, is another.

Fukui's unCLIP-KD work combines that idea with generation. Tatsuya Sasaki, Shunsuke Sakai and Hasegawa use representations from CLIP to condition a generative model, expanding the training material available for distillation. In the paper's reported experiments, the method consistently beat common augmentation and knowledge-distillation approaches across several benchmarks; in one CLIP-to-ResNet experiment on STL-10, accuracy improved by about 7.5 percentage points over the baseline.

The conceptual move is elegant. A foundation model has already spent enormous resources learning a semantic map of visual space. The smaller student does not need to recreate the entire journey. The generator helps expose the student to additional examples shaped by the teacher's representation.

But again, the word “distillation” matters. A bottle of distilled liquid is not the original mixture. A smaller network is useful precisely because it leaves things behind. The engineering question is whether it preserves the information needed for the downstream task while shedding enough parameters, memory use and inference cost to become practical.

Route three: build a committee without training every member from birth

Ensemble learning rests on a simple intuition: several models can make a better decision together than one model can make alone. If independently trained networks make different mistakes, averaging their predictions can improve accuracy and calibration.

The price is obvious. Five independently trained networks can require roughly five training processes. That may be reasonable in a large research center. It is less attractive when the goal is a practical, resource-constrained system.

Noisy Deep Ensemble, developed by Shunsuke Sakai, Shunsuke Tsuge and Hasegawa, changes the construction process. The method trains a parent model to convergence, perturbs its weights in different ways, and then uses those perturbed starting points to create child models. The perturbation is intended to push the children toward different nearby solutions without forcing every ensemble member to begin from random initialization.

On CIFAR-10 and CIFAR-100 with multiple convolutional architectures, the authors reported test accuracy comparable to standard deep ensembles while substantially reducing the training burden, and better performance than other efficient-ensemble baselines in their experiments.

This may sound like a technical optimization, but it expresses the same philosophy as the other Fukui work. Do not spend computation twice if information from the first computation can be reused. Diversity still matters—the committee is useless if every member repeats the same error—but diversity can be engineered more deliberately.

From fish to a factory line: why “fine-grained” is a real-world word

Fine-grained recognition is easiest to understand by thinking about categories that humans regard as almost the same. A generic classifier may only need to recognize “bird.” A fine-grained classifier may need to distinguish species whose silhouette is nearly identical. A factory system may need to distinguish a normal component from one with a subtle assembly error. In medicine, visually small differences can carry large consequences.

University of Fukui's current project explicitly lists industrial-product anomaly detection and primary-industry recognition among its research groups, and it describes future cross-disciplinary connections including medical-engineering collaboration. The university's broader AI program has already worked across fisheries, meal-intake measurement, sensing and medical-support problems.

One 2026 Fukui paper illustrates the industrial side directly. Sakai, Hasegawa and Makoto Koshino proposed LADMIM, a masked-image-modeling approach for “logical” anomaly detection. Many anomaly systems are good at finding a scratch or unusual texture. Logical anomalies are stranger: the individual objects may all look normal, but one is missing, misplaced or combined incorrectly. LADMIM masks parts of an image and asks the model to infer what belongs there, forcing it to learn relationships across the scene rather than only local surface patterns.

That matters for inspection. A perfect screw in the wrong location is still an error. A package containing individually correct parts in the wrong combination is still defective. The problem is not merely appearance. It is structure.

Medical imaging creates a parallel constraint, but the stakes require caution. Foundation models and low-data methods can be valuable when expert labels are expensive and patient data cannot simply be multiplied at will. Synthetic images may help research, and compact models can enable local deployment. But clinical validation, bias, calibration, privacy and regulatory evidence do not disappear because training became more efficient. Fukui's project identifies medical-engineering collaboration as a direction; the specific methods described here should not be read as already validated diagnostic products.

Fukui is part of the story

There is a geographic reason this work is interesting. The global AI narrative tends to point toward giant technology companies, hyperscale data centers and a race for ever larger clusters. Fukui offers a different setting: a national university in a regional manufacturing and primary-industry economy asking how advanced AI can be made usable outside the places that can purchase unlimited compute.

Hasegawa's laboratory was established at the University of Fukui in 2017. His own route to academia included work in medical systems at Fujitsu Hokuriku Systems before a doctorate at Kanazawa University and appointments in medical information and engineering. The laboratory's portfolio has ranged from activity recognition and learning-support systems to fish recognition, anomaly detection and foundation-model research.

The university's earlier cross-disciplinary AI project stated the adoption problem plainly: many organizations still cannot use deep learning because preparing datasets and tuning models is complicated, while people capable of applying the technology are scarce. Its proposed answer was not only a new algorithm but a “Virtual Laboratory” linking AI specialists with researchers in other fields.

That is an important corrective to the language of democratization. AI does not become democratic merely because a model is downloadable. Somebody still has to define the problem, collect and audit the examples, understand failure, maintain the system and decide what happens when confidence is low.

The dangerous shortcut: synthetic confidence

The most seductive version of the low-data story is that generative AI makes scarcity disappear. It does not. If the original dataset is biased, a generator may reproduce the bias. If the prompt omits the feature that matters, the synthetic image may teach the wrong distinction. If a rare defect has never been observed, an attractive generated defect may be physically implausible. If a model learns the generator's visual fingerprints, evaluation can look better without real-world performance improving.

The Fukui fine-grained augmentation work is valuable precisely because its own results expose this tension. Image generation could help when only one example was available, but text-to-image generation had difficulty reproducing detailed class features. That is not a reason to abandon synthetic data. It is a reason to measure it.

There is another scarcity that algorithms cannot solve: trustworthy evaluation data. A team can augment a training set, distill a teacher and compress an ensemble, but if the final test set does not represent the deployment environment, efficiency simply delivers the wrong answer faster.

The next frontier may be smaller

AI progress is usually narrated by addition: more parameters, more tokens, more images, more GPUs. That path has produced astonishing systems. It also creates a risk that “state of the art” becomes confused with “deployable everywhere.”

The University of Fukui work suggests a second axis of progress. A system can improve because it uses more resources, but it can also improve because it wastes less. A generator can create the variation that a camera never happened to capture. A foundation model can lend representations to a smaller specialist. An ensemble can share a trained ancestor instead of making every member learn from zero.

None of those ideas eliminates scarcity. They negotiate with it.

That may be the more important form of intelligence for the next wave of applied AI in Japan. A factory in Fukui does not need a machine that can recognize everything on the internet. It needs one that notices the wrong part at the right time. A fisheries worker does not need a universal visual oracle. The model needs to tell two nearly identical species apart. A medical researcher does not need infinite synthetic patients. The system needs to learn from limited, carefully governed evidence without becoming confidently wrong.

The spectacular question in AI is how large a model can become. The practical question is how little a useful model can need.

In that sense, Fukui's low-data, low-compute research is not a rebellion against foundation models. It depends on them. It treats their accumulated knowledge as infrastructure—something expensive that has already been built—and asks how much of it can be carried into smaller rooms.