Vision Transformers Exhibit Human-Like Biases: Evidence of Orientation and Color Selectivity, Categorical Perception, and Phase Transitions

This study explored whether Vision Transformers (ViTs) developed orientation and color biases similar to those observed in the human brain. Using synthetic datasets with controlled variations in noise levels, angles, lengths, widths, and colors, we analyzed the behavior of ViTs fine-tuned with LoRA. Our findings revealed four key insights: First, ViTs exhibited an oblique effect showing the lowest angle prediction errors at 180 deg (horizontal) across all conditions. Second, angle prediction errors varied by color. Errors were highest for bluish hues and lowest for yellowish ones. Additionally, clustering analysis of angle prediction errors showed that ViTs grouped colors in a way that aligned with human perceptual categories. In addition to orientation and color biases, we observed phase transition phenomena. While two phase transitions occurred consistently across all conditions, the training loss curves exhibited delayed transitions when color was incorporated as an additional data attribute. Finally, we observed that attention heads in certain layers inherently develop specialized capabilities, functioning as task-agnostic feature extractors regardless of the downstream task. These observations suggest that biases and properties arise primarily from pre-training on the original dataset which shapes the model's foundational representations and the inherent architectural constraints of the vision transformer, rather than being solely determined by downstream data statistics.

Mechanistic Interpretability of Fine-Tuned Vision Transformers on Distorted Images: Decoding Attention Head Behavior for Transparent and Trustworthy AI

Mechanistic interpretability improves the safety, reliability, and robustness of large AI models. This study examined individual attention heads in vision transformers (ViTs) fine tuned on distorted 2D spectrogram images containing non relevant content (axis labels, titles, color bars). By introducing extraneous features, the study analyzed how transformer components processed unrelated information, using mechanistic interpretability to debug issues and reveal insights into transformer architectures. Attention maps assessed head contributions across layers. Heads in early layers (1 to 3) showed minimal task impact with ablation increased MSE loss slightly ({\mu}=0.11%, {\sigma}=0.09%), indicating focus on less critical low level features. In contrast, deeper heads (e.g., layer 6) caused a threefold higher loss increase ({\mu}=0.34%, {\sigma}=0.02%), demonstrating greater task importance. Intermediate layers (6 to 11) exhibited monosemantic behavior, attending exclusively to chirp regions. Some early heads (1 to 4) were monosemantic but non task relevant (e.g. text detectors, edge or corner detectors). Attention maps distinguished monosemantic heads (precise chirp localization) from polysemantic heads (multiple irrelevant regions). These findings revealed functional specialization in ViTs, showing how heads processed relevant vs. extraneous information. By decomposing transformers into interpretable components, this work enhanced model understanding, identified vulnerabilities, and advanced safer, more transparent AI.

Uncovering the Origins of Instability in Dynamical Systems: How Attention Mechanism Can Help?

The behavior of the network and its stability are governed by both dynamics of individual nodes as well as their topological interconnections. Attention mechanism as an integral part of neural network models was initially designed for natural language processing (NLP), and so far, has shown excellent performance in combining dynamics of individual nodes and the coupling strengths between them within a network. Despite undoubted impact of attention mechanism, it is not yet clear why some nodes of a network get higher attention weights. To come up with more explainable solutions, we tried to look at the problem from stability perspective. Based on stability theory, negative connections in a network can create feedback loops or other complex structures by allowing information to flow in the opposite direction. These structures play a critical role in the dynamics of a complex system and can contribute to abnormal synchronization, amplification, or suppression. We hypothesized that those nodes that are involved in organizing such structures can push the entire network into instability modes and therefore need higher attention during analysis. To test this hypothesis, attention mechanism along with spectral and topological stability analyses was performed on a real-world numerical problem, i.e., a linear Multi Input Multi Output state-space model of a piezoelectric tube actuator. The findings of our study suggest that the attention should be directed toward the collective behaviour of imbalanced structures and polarity-driven structural instabilities within the network. The results demonstrated that the nodes receiving more attention cause more instability in the system. Our study provides a proof of concept to understand why perturbing some nodes of a network may cause dramatic changes in the network dynamics.