Information geometry for multiparameter models: new perspectives on the origin of simplicity

The training process of many deep networks explores the same low-dimensional manifold

We develop information-geometric techniques to analyze the trajectories of the predictions of deep networks during training. By examining the underlying high-dimensional probabilistic models, we reveal that the training process explores an effectively low-dimensional manifold. Networks with a wide range of architectures, sizes, trained using different optimization methods, regularization techniques, data augmentation techniques, and weight initializations lie on the same manifold in the prediction space. We study the details of this manifold to find that networks with different architectures follow distinguishable trajectories, but other factors have a minimal influence; larger networks train along a similar manifold as that of smaller networks, just faster; and networks initialized at very different parts of the prediction space converge to the solution along a similar manifold.

Sloppy model analysis identifies bifurcation parameters without normal form analysis

Bifurcation phenomena are common in multidimensional multiparameter dynamical systems. Normal form theory suggests that bifurcations are driven by relatively few combinations of parameters. Models of complex systems, however, rarely appear in normal form, and bifurcations are controlled by nonlinear combinations of the bare parameters of differential equations. Discovering reparameterizations to transform complex equations into a normal form is often very difficult, and the reparameterization may not even exist in a closed form. Here we show that information geometry and sloppy model analysis using the Fisher information matrix can be used to identify the combination of parameters that control bifurcations. By considering observations on increasingly long timescales, we find those parameters that rapidly characterize the system's topological inhomogeneities, whether the system is in normal form or not. We anticipate that this novel analytical method, which we call time-widening information geometry (TWIG), will be useful in applied network analysis.

Playing it safe: information constrains collective betting strategies

Every interaction of a living organism with its environment involves the placement of a bet. Armed with partial knowledge about a stochastic world, the organism must decide its next step or near-term strategy, an act that implicitly or explicitly involves the assumption of a model of the world. Better information about environmental statistics can improve the bet quality, but in practice resources for information gathering are always limited. We argue that theories of optimal inference dictate that "complex" models are harder to infer with bounded information and lead to larger prediction errors. Thus, we propose a principle of playing it safe where, given finite information gathering capacity, biological systems should be biased towards simpler models of the world, and thereby to less risky betting strategies. In the framework of Bayesian inference, we show that there is an optimally safe adaptation strategy determined by the Bayesian prior. We then demonstrate that, in the context of stochastic phenotypic switching by bacteria, implementation of our principle of "playing it safe" increases fitness (population growth rate) of the bacterial collective. We suggest that the principle applies broadly to problems of adaptation, learning and evolution, and illuminates the types of environments in which organisms are able to thrive.

Far from Asymptopia: Unbiased High-Dimensional Inference Cannot Assume Unlimited Data

Inference from limited data requires a notion of measure on parameter space, which is most explicit in the Bayesian framework as a prior distribution. Jeffreys prior is the best-known uninformative choice, the invariant volume element from information geometry, but we demonstrate here that this leads to enormous bias in typical high-dimensional models. This is because models found in science typically have an effective dimensionality of accessible behaviors much smaller than the number of microscopic parameters. Any measure which treats all of these parameters equally is far from uniform when projected onto the sub-space of relevant parameters, due to variations in the local co-volume of irrelevant directions. We present results on a principled choice of measure which avoids this issue and leads to unbiased posteriors by focusing on relevant parameters. This optimal prior depends on the quantity of data to be gathered, and approaches Jeffreys prior in the asymptotic limit. However, for typical models, this limit cannot be justified without an impossibly large increase in the quantity of data, exponential in the number of microscopic parameters.

How Occam's razor guides human decision-making

Occam's razor is the principle that, all else being equal, simpler explanations should be preferred over more complex ones. This principle is thought to play a role in human perception and decision-making, but the nature of our presumed preference for simplicity is not understood. Here we use preregistered behavioral experiments informed by formal theories of statistical model selection to show that, when faced with uncertain evidence, human subjects exhibit preferences for particular, theoretically grounded forms of simplicity of the alternative explanations. These forms of simplicity can be understood in terms of geometrical features of statistical models treated as manifolds in the space of the probability distributions, in particular their dimensionality, boundaries, volume, and curvature. The simplicity preferences driven by these features, which are also exhibited by artificial neural networks trained to optimize performance on comparable tasks, generally improve decision accuracy, because they minimize over-sensitivity to noisy observations (i.e., overfitting). However, unlike for artificial networks, for human subjects these preferences persist even when they are maladaptive with respect to the task training and instructions. Thus, these preferences are not simply transient optimizations for particular task conditions but rather a more general feature of human decision-making. Taken together, our results imply that principled notions of statistical model complexity have direct, quantitative relevance to human and machine decision-making and establish a new understanding of the computational foundations, and behavioral benefits, of our predilection for inferring simplicity in the latent properties of our complex world.