Unlocking Brain-Inspired AI: How Architecture and Learning Rules Shape Visual Intelligence

      Artificial intelligence (AI), particularly in the realm of computer vision, has made astonishing strides in recent years. Convolutional Neural Networks (CNNs) now rival human performance in many visual tasks. However, the exact mechanisms by which these complex networks emulate human visual processing—and whether their internal workings truly align with our brains—remain a subject of intense scientific inquiry. A crucial question is whether the method an AI uses to "learn" influences its ability to mimic human vision, or if its inherent design is more significant.

      A recent academic paper, "Untrained CNNs Match Backpropagation at V1: A Systematic RSA Comparison of Four Learning Rules Against Human fMRI," sheds light on this very question. This research, available on arXiv, delves into how different AI learning rules compare to the human brain's visual cortex, even including a critical "untrained" baseline. The findings offer profound insights into designing more efficient, brain-inspired AI systems, especially for applications demanding practical deployment like edge computing.

Decoding Visual Intelligence: How AI Learns to 'See'

      At its core, a Convolutional Neural Network (CNN) is a specialized type of deep neural network designed to process pixel data in images. It mimics aspects of the animal visual cortex, using layers of filters to detect features such as edges, textures, and eventually more complex objects. The traditional method for training these networks is called backpropagation (BP), a powerful algorithm that iteratively adjusts the network's internal "weights" by propagating error signals backward through its layers. While highly effective, backpropagation is considered biologically implausible because it requires complex, non-local information flow and symmetrical weight updates that don't appear to happen in the human brain.

      This biological implausibility has spurred research into alternative learning rules that might be more aligned with how biological brains learn. These alternatives aim to achieve similar performance while adhering to more realistic constraints, such as localized weight updates or simplified error propagation. Understanding these different learning paradigms is key to building next-generation AI that is not only powerful but also potentially more robust, energy-efficient, and understandable.

The Brain's Blueprint: Early vs. Late Visual Processing

      The human visual system is a marvel of hierarchical processing. When we see an image, information first enters the primary visual cortex (V1), where basic features like lines, edges, and orientations are detected. From there, it flows through secondary visual areas (V2) to higher-order regions such as the Lateral Occipital Complex (LOC) and Inferior Temporal (IT) cortex. As information moves from V1 to IT, the representations become progressively more complex and abstract, evolving from simple features to intricate object recognition. This hierarchical processing allows us to identify objects regardless of their size, position, or lighting.

      To compare AI models with this biological hierarchy, researchers employ Representational Similarity Analysis (RSA). RSA quantifies the degree to which the patterns of neural activity (in the brain) or feature activation (in the AI) are similar when processing different stimuli. By comparing these "similarity matrices," RSA reveals how well an AI model's internal representations align with those of the human brain across different visual areas, providing a crucial bridge between computational models and neurobiology.

Unpacking AI Learning Rules: Beyond Backpropagation

      The study systematically compared four distinct learning rules, alongside a crucial untrained baseline, all applied to identical CNN architectures trained on the CIFAR-10 image dataset. The visual alignment of these AI models was then evaluated against human fMRI data from the THINGS-fMRI dataset, which features responses to 720 everyday objects. This rigorous comparison aimed to isolate the impact of the learning mechanism.

      Here’s a brief look at the learning rules examined:

• Backpropagation (BP): The industry standard, known for its high performance but considered biologically implausible due to requirements like symmetric forward and feedback weights and global error signals.
• Feedback Alignment (FA): A biologically more plausible alternative to BP that replaces symmetric feedback weights with fixed random matrices. It addresses the "weight transport problem" where biological neurons would need to precisely match forward and backward connection strengths.
• Predictive Coding (PC): Rooted in neuroscience, this approach frames perception as a continuous process of minimizing prediction errors across a hierarchy of neural layers. Learning is driven by local Hebbian updates, making it highly plausible for biological systems.
• Spike-Timing-Dependent Plasticity (STDP): A highly biologically realistic learning rule where the strength of synaptic connections is adjusted based on the precise timing of "spikes" (neural impulses) between neurons. This rule directly mimics known mechanisms of synaptic plasticity in the brain.
• Random Weights Baseline: A network initialized with random weights but never trained. This condition is vital for determining how much of the brain alignment is due simply to the CNN's architectural design (its layers, filters, pooling structure) rather than any learning process.

Key Findings: Architecture Leads Early, Learning Rules Define Later

      The study yielded several compelling results that challenge conventional thinking about AI learning and brain alignment. The most striking finding was regarding the early visual areas, V1 and V2:

• Architecture Dominates Early Vision: An untrained CNN, with merely its random initial weights, achieved a statistical alignment with human V1/V2 that was indistinguishable from a network trained with the powerful backpropagation algorithm. This suggests that the inherent convolutional architecture of CNNs—its layered structure, filters, and pooling mechanisms—is the primary driver for representations in the earliest stages of visual processing, even before any explicit learning occurs.
• Learning Rules Matter for Higher-Level Vision: The impact of learning rules became significant only in higher visual areas like LOC and IT cortex, which are responsible for more abstract object recognition. Here, backpropagation demonstrated a clear advantage over the random baseline.
• Predictive Coding Rivals Backpropagation: Remarkably, Predictive Coding (PC), a biologically plausible learning rule relying only on local Hebbian updates, achieved IT alignment statistically indistinguishable from backpropagation. This indicates that complex, global error signals might not be necessary for high-level visual alignment, opening doors for more brain-like and efficient AI.

Feedback Alignment Underperforms: Surprisingly, Feedback Alignment (FA) consistently degraded representations, performing worse* than the random baseline at V1. This suggests that while FA aims for biological plausibility, its current implementation may not always lead to effective brain alignment in CNNs.

• Robust Results: The researchers confirmed that these effects were robust, even after controlling for pixel-level similarity between stimuli, ensuring the observed alignments were due to deeper representational structures.

Implications for Next-Gen AI and Edge Computing

      These findings carry significant implications for the future of AI development, particularly in areas where efficiency, biological inspiration, and real-world deployment are crucial. The fact that architectural design is so potent in early visual processing suggests that optimizing network structures, rather than just learning algorithms, could lead to more inherently capable AI. For instance, in AI video analytics, an optimally designed network can begin to extract meaningful low-level features even before extensive training, leading to faster deployment and reduced computational overhead.

      The strong performance of Predictive Coding at higher visual areas is particularly exciting. Its reliance on local Hebbian updates makes it far more biologically plausible and potentially more suitable for edge AI applications. Edge devices, such as those in ARSA's AI Box Series, benefit immensely from learning rules that require minimal computational resources and operate without constant cloud connectivity. Local updates reduce data transfer, enhance privacy, and enable real-time processing directly on the device, aligning perfectly with the demands of autonomous systems in industrial, retail, and smart city environments.

      Understanding which learning rules contribute effectively to brain alignment also informs the development of more robust and interpretable AI. If an AI system processes information similarly to the human brain, it might also share some of its strengths, such as generalization capabilities and robustness to noise. This could lead to more reliable computer vision systems for critical applications like security, safety monitoring, and healthcare. For organizations seeking tailored AI solutions that marry high performance with practical deployment realities, designing custom AI systems that leverage these insights can lead to significant competitive advantages. ARSA Technology, with its expertise in developing and deploying custom AI solutions, focuses on engineering intelligence into operations for mission-critical enterprises.

      This research underscores that effective AI doesn't always require the most computationally intensive or globally coordinated learning rules. Sometimes, smarter architecture and biologically inspired local learning can yield equally powerful results, paving the way for a new generation of AI systems that are both highly capable and elegantly efficient.

      For businesses looking to implement advanced AI and IoT solutions, understanding these fundamental principles is key to making informed strategic decisions. To explore how brain-inspired AI and edge computing can transform your operations, please contact ARSA for a free consultation.

      Source: Leutengger, N. (2026). Untrained CNNs Match Backpropagation at V1: A Systematic RSA Comparison of Four Learning Rules Against Human fMRI. arXiv preprint arXiv:2604.16875.