TLDR: This research introduces two novel Convolutional Neural Network (CNN) models that integrate lateral connections, a key feature of the mammalian visual cortex. The first model, LC-CNN, uses recurrent activation within layers, leading to improved accuracy and emergent biological properties like sparse activation and power-law-like response distributions. The second model, LCEI-CNN, separates excitatory and inhibitory lateral connections using custom loss functions, also enhancing performance. The study demonstrates that incorporating these biological principles not only boosts CNN performance but also helps align AI models closer to the functional and structural characteristics of the brain.
Convolutional Neural Networks (CNNs) have revolutionized computer vision, drawing significant inspiration from the mammalian visual system. These powerful AI models mimic how our brains process visual information, incorporating concepts like afferent connections (from the retina to the visual cortex) and long-range projections (connections between different visual cortical areas). However, a crucial architectural feature present in the biological visual system has largely been overlooked in CNN design: lateral, or horizontal, connections.
Lateral connections are intricate networks within each visual cortical area, allowing neurons to communicate locally. This paper, titled “INCORPORATING VISUAL CORTICAL LATERAL CONNECTION PROPERTIES INTO CNN: RECURRENT ACTIVATION AND EXCITATORY-INHIBITORY SEPARATION,” by Jin Hyun Park, Cheng Zhang, and Yoonsuck Choe from Texas A&M University, addresses this gap. The researchers propose a novel approach to integrate these lateral connections into the standard CNN framework, focusing on two key properties: recurrent activation and the separation of excitatory and inhibitory connections.
Recurrent Activation in Laterally Connected CNN (LC-CNN)
The first model, LC-CNN, introduces recurrent activation through lateral connections. Unlike traditional CNNs where information flows strictly forward, LC-CNN allows information to loop back within the same convolutional layer. This is achieved by sharing weights across recurrent loops, meaning the model learns to refine its feature representations over several ‘passes’ within a layer, much like how biological neurons might process information iteratively.
The results were compelling. LC-CNN consistently outperformed baseline CNNs in classification accuracy across various datasets, with performance improving as the number of recurrent loops increased. More importantly, the model exhibited emergent properties strikingly similar to those observed in the biological visual cortex. For instance, the neural activations became sparser with more loops, meaning the network focused on more relevant features while suppressing background noise. This sparsification is a known characteristic of efficient neural coding in the brain.
Furthermore, the study found that the response distribution in LC-CNN shifted from a normal distribution towards a power-law-like distribution, a pattern also observed in the visual cortex and linked to perceptual saliency. The analysis of connection properties revealed that feature maps with similar characteristics tended to have similar outgoing lateral connection patterns, mirroring the biological observation that lateral connections in the visual cortex link regions with similar orientation preferences, which is thought to aid in contour detection.
Excitatory and Inhibitory Separation in Laterally Connected CNN (LCEI-CNN)
The second model, LCEI-CNN, delves into another fundamental aspect of biological neural networks: the distinct roles of excitatory and inhibitory connections. Inspired by Dale’s law, which states that a neuron typically releases only one type of neurotransmitter (either excitatory or inhibitory), the researchers designed LCEI-CNN to have separate pathways for excitatory and inhibitory lateral weights.
To achieve this, custom loss functions were developed to encourage these weights to become predominantly positive (excitatory) or negative (inhibitory). The LCEI-CNN models, with their separated excitatory and inhibitory connections, generally showed improved classification accuracy compared to a baseline without this separation. The weight distributions clearly demonstrated the effectiveness of the penalty terms, with one group of weights becoming positive and the other negative.
While the model successfully separated excitatory and inhibitory weights, the observed ratio of excitatory to inhibitory connections was closer to 0.5:0.5, rather than the biological ratios of 8:2 or 7:3. The authors suggest this warrants further investigation, noting that their count is based on connections, which may not directly translate to the number of neurons.
Also Read:
- Cognitive-Inspired AI: A New Method for Attention Management in Transformers
- EvoBrain: A New Model for Understanding Dynamic Brain Networks in Seizure Detection
Conclusion and Future Directions
This research marks a significant step towards building more biologically plausible and powerful CNN architectures. By incorporating lateral connections, the models not only achieve better performance but also exhibit functional and structural properties that align closely with our understanding of the visual cortex. The emergent sparsification and power-law-like responses, along with the specific patterns of lateral connectivity, highlight the computational significance of these biological features.
The authors plan to combine these two models into a single, more comprehensive architecture and explore their application in spiking CNNs. This work deepens our understanding of the computational roles of lateral connections in the visual cortex and paves the way for a new generation of biologically inspired AI. You can read the full paper here: Incorporating Visual Cortical Lateral Connection Properties into CNN.
