Abstract

Authors survey overlaps between deep learning systems and biological neural computation, emphasizing where metaphors help — and where they mislead. The review argues that productive exchange between machine learning and neuroscience requires clearer separation of engineering performance, mechanistic hypothesis, and rhetorical flourish.

Across the papers surveyed, three patterns recur: (1) useful shared vocabulary around representation and learning dynamics; (2) frequent slippage from analogy into identity claims; and (3) uneven empirical standards when “brain-inspired” is used as a selling point rather than a testable design constraint.

Cortex extracts claim boundaries only — no invented effect sizes or DOIs beyond the article’s own identifiers.

1. Introduction

AI systems increasingly borrow language from neuroscience. Terms such as attention, memory, and “neural” architectures travel between communities with different standards of evidence. This review asks which borrowings are mechanistic, which are rhetorical, and which remain untested.

The introduction situates the review against two decades of brain-inspired computing narratives — from early connectionism to modern transformers — and notes that public product language often outruns the cited neuroscience.

A working distinction is proposed early: engineering metaphors (useful for design intuition) versus biological claims (requiring measurement against neural or behavioral data).

2. Representation learning

Distributed representations in artificial networks are contrasted with population codes in biological systems. Similarity metrics (RSA-style comparisons, probing classifiers) appear throughout the cited literature, but rarely justify equating a layer with a cortical area.

The authors summarize evidence that artificial networks can develop latent geometries that correlate with neural recordings under controlled tasks — while stressing that correlation is not circuit identity.

Practical takeaway logged for Cortex: when marketing or research notes say “like the brain,” prefer citations that specify the measurement (task, species, recording modality) over loose metaphor.

3. Attention and routing

Transformer attention is compared to selective routing hypotheses in cortex. The paper stresses analogy limits: attention weights are not synaptic traces, and multi-head attention is not a literal map of cortical columns.

Several cited works use “attention” as a bridge term. The review recommends keeping the mathematical definition (query–key–value weighting) distinct from psychological or neuroscientific attention.

Cortex note: extract this caveat whenever product copy or other papers treat attention maps as neural proof.

4. Evaluation gaps

Calls for benchmarks that separate engineering performance from claims about biological plausibility. Leaderboard wins alone do not validate brain models.

Suggested evaluation axes include: task ecological validity, comparison to neural data when claimed, ablation of “brain-inspired” components, and transparent reporting of negative results.

The review criticizes papers that cite neuroscience selectively in introductions while evaluating only on standard ML datasets.

5. Discussion and limitations

Useful for Cortex: cite caveats when product language uses “brain” framing; prefer measured evaluations over metaphor. The authors acknowledge selection bias in any narrative review and call for systematic meta-analyses.

Limitations include incomplete coverage of embodied and neuromorphic lines of work, and rapidly moving transformer literature that may outdate specific citations.

Closing recommendation: interdisciplinary teams should agree upfront whether a project aims for biological insight, engineering gains, or both — and evaluate accordingly.

References (scan)

Reference list scanned for overlap with arXiv brain-aligned evals and Nature AI/neuro pieces already in Cortex’s research graph. High-citation neuroscience primers flagged for a later deep pass.