Search for: All records

Abstract Understanding the connection between topography and the processes and properties that shape landscapes both helps us better understand landscape evolution, and use topography to infer information relevant for a wide range of human activities. Deep learning methods, like convolutional neural networks (CNNs) have been successfully applied to topographic data in geomorphic contexts but primarily for classification problems. This raises the question: could a CNN be trained to “learn” the parameters responsible for forming a particular kind of terrain? As a first test, we train a convolutional neural network to invert topography generated from a landscape evolution model into the model parameters, and interpret the model's learning. We find that the trained network is able to accurately infer the parameter ratio that governs the degree of landscape dissection, performs better when given topographic derivatives, and has encoded a meaningful relationship between the model parameters and both valley spacing and drainage density. 

Abstract The backpropagation method has enabled transformative uses of neural networks. Alternatively, for energy-based models, local learning methods involving only nearby neurons offer benefits in terms of decentralized training, and allow for the possibility of learning in computationally-constrained substrates. One class of local learning methodscontraststhe desired, clamped behavior with spontaneous, free behavior. However, directly contrasting free and clamped behaviors requires explicit memory. Here, we introduce ‘Temporal Contrastive Learning’, an approach that uses integral feedback in each learning degree of freedom to provide a simple form of implicit non-equilibrium memory. During training, free and clamped behaviors are shown in a sawtooth-like protocol over time. When combined with integral feedback dynamics, these alternating temporal protocols generate an implicit memory necessary for comparing free and clamped behaviors, broadening the range of physical and biological systems capable of contrastive learning. Finally, we show that non-equilibrium dissipation improves learning quality and determine a Landauer-like energy cost of contrastive learning through physical dynamics. 

We introduce a novel framework for analyzing reinforcement learning (RL) in continuous state-action spaces, and use it to prove fast rates of convergence in both off-line and on-line settings. Our analysis highlights two key stability properties, relating to how changes in value functions and/or policies affect system dynamics via the Bellman operator and induced occupation measure. We show that these stability properties can be ensured by sufficient curvature in the function class used to represent state-action value functions. Curvature is a static property, making it easier to verify than the dynamic stability conditions that it implies. Using linear function represetation, we illustrate the connection between curvature and stability with concrete formulations. Overall, the key insight is that function classes with sufficient curvature guarantee dynamic system stability, leading to fast convergence rates in RL. 

Abstract Retreat of continental ice sheets exposes comminuted sediment in disequilibrium with non-glacial conditions. Weathering of this sediment may create climate feedbacks by altering exchange of greenhouse gases between atmosphere and landscapes. Here we show in a partially deglaciated watershed in southwest Greenland that glacial meltwater contains low concentrations of reactive dissolved organic carbon that enhances weathering of freshly comminuted sediment causing net sequestration of carbon dioxide. In contrast, soil water reactions enhance methanogenesis and carbon dioxide production and create greenhouse gas sources as organic carbon is remineralized. We suggest that a change from greenhouse gas sinks in glacial meltwater to greenhouse gas sources in soil water creates a switch from a negative to positive warming feedback during glacial-interglacial transitions, but a negative warming feedback may return with future anthropogenic warming, glacial retreat, and increased meltwater production. We anticipate changing weathering reactions following exposure also alter nutrient and radiogenic isotope exports. 

In this work we show how a network inspired by a coarse-grained description of the cytoskeleton can learn—in a contrastive learning framework—from environmental perturbations if it is endowed with mechanosensitive proteins and motors. Our work is a proof of principle for how force-sensitive proteins and molecular motors can form the basis of a general strategy to learn in biological systems without neurons. Our work identifies a minimal biologically plausible learning mechanism and also explores its implications for commonly occurring phenomenology, such as adaptation and homeostasis.