A new study challenges the assumption that a neural network's learning rule is key to mimicking human vision. Researchers from the University of Tübingen compared four learning rules—backpropagation, feedback alignment, predictive coding, and spike-timing-dependent plasticity—using identical convolutional architectures and tested them against human fMRI data from 3 subjects viewing 720 stimuli (THINGS-fMRI dataset).
The Research
Nils Leutenegger and team used Representational Similarity Analysis (RSA) to measure how well each network's internal representations matched human brain activity. Crucially, they included an untrained network with random weights. Results: at early visual areas (V1/V2), the untrained CNN achieved a correlation of rho = 0.071, statistically indistinguishable from backpropagation (rho = 0.072, p = 0.43). At higher areas (LOC/IT), learning rules mattered: backpropagation dominated, and predictive coding with local Hebbian updates matched it (p = 0.18). Feedback alignment consistently impaired representations below the random baseline at V1. Partial RSA confirmed effects were not due to pixel similarity.
Why It Matters
This finding suggests that early visual processing is largely shaped by architecture—the structure of connections—rather than by how the network learns. For your own cognition, it implies that the brain's early stages may be hardwired for efficiency, while higher-order learning depends on experience and objectives. Understanding this hierarchy can guide brain training: strengthening lower-level pattern recognition may require different methods than improving reasoning or memory.
What You Can Do
To boost visual processing, try tasks that challenge your pattern recognition without complex reasoning—like visual puzzles or fast object identification. For higher-level cognition, focus on goal-oriented learning, such as studying a new language or playing strategy games. The study underscores that brain training should target specific levels of processing.
Source: arXiv q-bio.NC
Curious about your own brain? Take our free adaptive IQ test or try 306 brain training levels.
