The Big Picture

Biological systems demonstrate the ability to adapt in spite of time-varying environments, limited data, and scarce resources by using examples presented from the environment to learn robust solutions. Such learning takes places in a variety of biological systems and across many time- and length-scales, depending on the context; evolution requires hundreds of years to induce population shifts while neural systems may only require a few days to wire retinal connections to capture correlation structures. In spite of the individual complexities, many biological systems are modeled using basic statistical physics models, as these models can effectively capture and reproduce the statistics observed in biological data despite being minimal in terms of structure. In addition to preserving the stationary distributions of biological data, statistical physics models can also provide insight into the dynamics of population drift, making it particularly suitable to study learning in contexts where the input may be structured. In our work, we ask several questions about scales in learning systems.