View-tolerant face recognition and Hebbian learning imply mirror-symmetric neural tuning to head orientation

The primate brain contains a hierarchy of visual areas, dubbed the ventral stream, which rapidly computes object representations that are both specific for object identity and relatively robust against identity-preserving transformations like depth-rotations. Current computational models of object recognition, including recent deep learning networks, generate these properties through a hierarchy of alternating selectivity-increasing filtering and tolerance-increasing pooling operations, similar to simple-complex cells operations. While simulations of these models recapitulate the ventral stream’s progression from early view-specific to late view-tolerant representations, they fail to generate the most salient property of the intermediate representation for faces found in the brain: mirror-symmetric tuning of the neural population to head orientation. Here we prove that a class of hierarchical architectures and a broad set of biologically plausible learning rules can provide approximate invariance at the top level of the network. While most of the learning rules do not yield mirror-symmetry in the mid-level representations, we characterize a specific biologically-plausible Hebb-type learning rule that is guaranteed to generate mirror-symmetric tuning to faces tuning at intermediate levels of the architecture.

💡 Research Summary

The paper addresses a striking neural phenomenon observed in the macaque ventral visual stream: neurons in the intermediate face‑selective area (AL) exhibit mirror‑symmetric tuning to head orientation, responding similarly to left‑profile and right‑profile views but not to the frontal view. Early areas (ML/MF) are view‑specific, while later areas (AM) are largely view‑invariant. Existing computational models of object and face recognition—typically hierarchical feed‑forward architectures that alternate filtering (simple‑cell‑like) and pooling (complex‑cell‑like) operations—can reproduce the progression from view‑specific to view‑invariant representations but fail to generate the mirror‑symmetric intermediate tuning.

The authors propose a biologically plausible hierarchical model that combines such feed‑forward processing with Hebbian‑type learning at the level of simple cells. They focus on Oja’s rule, a normalized Hebbian update that converges to the principal components (PCs) of the input distribution. When an immature neuron is exposed to a sequence of face images rotating from left to right profile, the training set is bilaterally symmetric: for every view there is a mirrored counterpart. Under Oja learning, the resulting PCs are either even (symmetric) or odd (antisymmetric) functions of the view angle. Consequently, when the complex‑cell layer pools the squared responses of these PCs, the intermediate layer’s units inherit even‑function tuning curves, i.e., mirror symmetry.

The authors prove that this mechanism also yields approximate invariance to depth rotation at the top layer: pooling over the set of PCs (instead of discrete view templates) produces a signature that changes little as the face rotates, because the PCs span the subspace of the observed view manifold. They extend the proof to other unsupervised learning rules (Foldiak’s trace rule, ICA) and supervised back‑propagation, showing that most such rules can achieve invariance, but only Oja’s rule guarantees the even/odd structure needed for mirror symmetry.

Empirical validation uses HMAX C1 features as input. Two models are compared: a classic view‑based model that stores discrete view templates, and the Oja‑based model that learns PCs from the same training videos. Performance is measured on a same‑different face matching task with increasing angular ranges (invariance radius). With a sufficient number of PCs (e.g., 15 or more), the Oja model matches the view‑based model and even surpasses it when all 39 PCs are used. Both outperform the raw C1 baseline, confirming that the learned representation is both view‑tolerant and capable of supporting mirror‑symmetric intermediate tuning.

In summary, the paper demonstrates that a feed‑forward hierarchical architecture combined with Oja‑type Hebbian learning can simultaneously explain two key aspects of face processing in the primate brain: (1) the emergence of mirror‑symmetric orientation tuning in an intermediate cortical stage, and (2) the development of robust, view‑invariant face representations at higher stages. This work bridges neurophysiological observations with computational theory, suggesting that simple, biologically plausible learning rules may underlie the sophisticated transformation of visual information observed in the ventral stream.