Smarter AI, Less Data

Much of mainstream AI development is still dominated by a familiar playbook: collect vast datasets, train ever-larger models, and hope emergent capability outruns the costs. That approach has underpinned many of today’s high-profile systems, but it comes with growing friction—data can be costly to collect and clean, and may be legally or ethically constrained. Even when data is available, the compute required to convert it into capability can be substantial.

A Johns Hopkins University report argues there is a credible alternative to brute-force scale: redesigning AI systems so their “built-in expectations” more closely resemble aspects of biological brains. In the Johns Hopkins summary, researchers report that when model architectures were made more brain-like, some systems produced brain-like patterns of activity without the usual large-scale dataset training—a result that, if replicated, would challenge the idea that cortex-aligned visual representations require massive pre-training primarily through more data and more compute.

The underlying idea is not new—engineers have long borrowed from biology—but the claim here is pointed: better design may, in some cases, substitute for huge training runs, potentially shifting the economics and energy footprint of AI.

According to Johns Hopkins, the researchers explored what happens when AI is engineered with constraints and dynamics intended to better resemble nervous systems, rather than treating neural networks as largely generic function approximators. In their account, models with more brain-like design produced neural activity patterns that resemble biological signatures—even before learning from data. That suggests some “brain-like” signals may reflect architecture as well as experience.

It’s important to read that carefully. “Brain-like activity” is not the same as general intelligence, consciousness, or human-level understanding. It typically refers to measurable features—patterns, correlations, rhythms, or representational geometry—used in neuroscience to compare systems. The Johns Hopkins piece frames this as a challenge to today’s data-hungry AI paradigm, suggesting that architectural choices may move researchers closer to neural plausibility more quickly, and could make training more efficient as a result. The summary is available via Johns Hopkins University’s Hub.

If this line of work holds up across tasks and labs, it would strengthen the case for “inductive bias”: the idea that learning systems can perform better with less data when they come with useful assumptions baked in.

In machine learning, inductive bias is the set of assumptions a model uses to generalise beyond what it has seen. Without some bias, learning from limited examples is either impossible or inefficient. A classic illustration is vision: convolutional neural networks historically learned images better than generic multilayer perceptrons because the architecture assumed that nearby pixels relate and patterns repeat across space—an inductive bias that often matches real-world imagery. A general primer on the concept is available at the inductive bias overview.

Brain-inspired work can be seen as taking that logic further: if biological brains are exceptionally data-efficient learners, some of that efficiency may come from structure—such as recurrence, sparse activity, local learning rules, and temporal prediction—rather than from exposure to extremely large datasets.

This isn’t just academic hair-splitting. Inductive bias is one of the few levers that can reduce the need for training data without simply shifting the burden elsewhere. In practice, “smarter design” can mean architectures that embed constraints (for example, causality or object permanence), training objectives that better match an environment, or hardware that supports different computational regimes.

Another reason this debate is intensifying is infrastructure. Training and operating large-scale AI can run into electricity supply, data-centre capacity, and emissions constraints—especially as AI features are embedded into everyday products and services. The International Energy Agency has warned that leading AI models and the data centres behind them can be “energy hungry”, with implications for power systems and climate targets; see IEA commentary on energy demand from top AI models.

If architectural improvements mean a model can learn useful behaviours from fewer examples—or reach a given capability with fewer training steps—that can translate into less compute time, less energy use, and lower cost. It can also affect who can participate. Data-hungry AI tends to favour organisations that can assemble large datasets and buy compute at scale. More data-efficient approaches could, in principle, broaden access to serious research and deployment.

Efficiency claims can be slippery, though. A method that reduces training cost might increase inference cost, or vice versa. Some “brain-like” models may also require specialised simulation approaches or hardware. Even so, the Johns Hopkins framing arrives at a moment when the field is actively looking for credible ways to do more with less.

A common misunderstanding is that brain-inspired AI aims to faithfully reproduce biology. In practice, much of the field is about selective borrowing: taking a principle that seems to matter (for example, sparse firing, event-driven computation, or predictive processing) and translating it into engineering that is tractable, testable, and useful.

Neuromorphic computing is one strand of this effort, aiming to run brain-inspired models efficiently using event-driven or spiking approaches. Intel’s work on neuromorphic systems, for example, describes architectures designed to process information with spikes and local activity rather than dense matrix multiplications, with a focus on energy efficiency and real-time adaptation; see Intel’s neuromorphic computing overview. IBM similarly positions neuromorphic computing as a route to brain-like efficiency and new forms of computation; see IBM Research’s neuromorphic computing explainer.

These programs do not, on their own, validate the specific Johns Hopkins findings, but they are consistent with the broader premise that today’s dominant deep-learning hardware and software stack is not the only possible path to efficient machine intelligence.

Another relevant thread is the idea—contested in the details but influential in the big picture—that brains are fundamentally predictive machines, constantly generating hypotheses about incoming sensory signals and updating those hypotheses when surprised. In neuroscience, this family of theories is often discussed under “predictive coding” and “predictive processing”. A peer-reviewed overview is available in Nature Reviews Neuroscience’s review on predictive coding.

Why does this matter for data? If an AI system is built to anticipate and compress the world through prediction, it may learn from the structure in unlabelled experience rather than relying heavily on human-curated labels. This is broadly consistent with observations from developmental science: infants do not receive millions of hand-annotated examples, yet they develop robust models of objects, motion, language, and social cues.

Researchers disagree about how directly predictive coding maps onto cortical circuits, and “brain-like” can mean different things depending on the metric. However, the design principle—learn by predicting—has influenced practical AI, including sequence modelling and “world model” approaches in robotics.

Even within mainstream deep learning, there has been a steady pivot away from reliance on labelled datasets towards self-supervised learning, where systems learn from the structure of data itself (for example, predicting missing parts of text, future frames in video, or augmentations of images). Meta has argued that self-supervised learning is a foundational route to more general intelligence because it reduces reliance on curated labels.

The Johns Hopkins result, as presented, goes further by implying that architecture alone can produce brain-like signatures even before training. Taken together, these lines of work suggest a possible future where “learning” is a combination of:

• strong architectural priors (inductive biases that better match the world)
• objectives aligned with prediction and control
• smaller, more targeted amounts of data

That would be a marked shift from the idea that scaling datasets and parameters is the primary path to progress. It would also reframe competition: breakthroughs might come as much from careful design and measurement as from bigger budgets.

Johns Hopkins’ recent report adds momentum to a growing argument in AI: data is not the only fuel. If brain-inspired design can reliably produce more brain-like behaviour—and do so with little or no training in some cases—it may point towards faster, cheaper, and less energy-intensive AI. The open question is how far these approaches generalise beyond specific laboratory tasks and benchmarks, and whether the field can turn “brain-like activity” into broadly useful capability without simply shifting costs to other parts of the stack.