Module 6 · Topology of the Loss Landscape
Cusp Catastrophe Theory
Catastrophe theory, developed by René Thom (1972) and popularised by Christopher Zeeman, classifies the stable singularities of smooth families of potential functions. The cusp catastrophe is the second in Thom’s classification and is the universality class most relevant to conformational switching and loss-landscape bifurcations in MAML.
6.1 The Cusp Normal Form
The cusp catastrophe is defined by the family of potential functions:
\[ V(x;\, a,\, b) = x^4 + a\,x^2 + b\,x \tag{6.1}\]
with state variable \(x \in \mathbb{R}\) and control parameters \((a, b) \in \mathbb{R}^2\). The equilibrium states are critical points of \(V\): \(\partial V/\partial x = 0\), i.e. \(4x^3 + 2ax + b = 0\).
6.2 The Catastrophe Set and Bifurcation Diagram
The catastrophe set in control-parameter space — where the number of equilibria changes — is determined by the simultaneous conditions:
\[ \frac{\partial V}{\partial x} = 0 \;\text{ and }\; \frac{\partial^2 V}{\partial x^2} = 0 \quad \Rightarrow \quad 8a^3 + 27b^2 = 0 \tag{6.2}\]
This cusp-shaped curve in \((a, b)\) space encloses the bistable region where two stable equilibria coexist. Crossing the fold set causes a discontinuous jump — a catastrophe — in the equilibrium state. The tip of the cusp at \((a, b) = (0, 0)\) is the degenerate point where three equilibria merge.
6.3 MAML Loss Landscape as a Cusp Surface
We identify the cusp parameters with MAML quantities:
- Splitting factor \(a\): determined by the meta-initialisation \(\theta^*\). Negative \(a\) (bistable landscape) arises when the meta-training task distribution contains energetically degenerate conformations. Positive \(a\) (monostable) corresponds to a homogeneous task distribution with a unique preferred geometry.
- Normal factor \(b\): determined by the task-specific gradient \(\nabla_\theta \mathcal{L}(\theta; \mathcal{S}_i)\). The support set of the new task tilts the potential landscape toward one basin, breaking the symmetry.
- State variable \(x\): the projection of the parameter update \(\phi - \theta\) onto the direction of maximum curvature change (the leading eigenvector of the inner-loop Hessian).
6.4 Cusp Bifurcation Detection via Hessian Spectroscopy
The catastrophe condition (Eq. 6.2) translates directly to a condition on the inner-loop Hessian:
\[ \lambda_1^{(k)} \equiv \lambda_{\min}\!\left(\nabla^2_\phi \mathcal{L}\!\left(\phi^{(k)};\, \mathcal{S}_i\right)\right) = 0 \tag{6.3}\]
When the minimum eigenvalue \(\lambda_1^{(k)}\) passes through zero during the inner loop, the MAML adaptation trajectory crosses the catastrophe set — a computational bifurcation. Monitoring \(\lambda_1^{(k)}\) vs. \(k\) therefore provides a spectroscopy of the loss-landscape topologythat maps directly onto the physical free-energy surface topology of the protein conformational landscape.
Universality
The cusp is a universality class: any smooth potential with a degenerate critical point of codimension 2 is locally equivalent (by a smooth change of coordinates) to Eq. (6.1). This means the cusp analysis applies not only to the toy potential \(V = x^4 + ax^2 + bx\) but to any bistable molecular energy landscape — provided it is smooth and has exactly two control parameters. The relevance to protein conformational switching is therefore universal, not system-specific.