1: \begin{abstract}%

2: Understanding the implicit bias of training algorithms is of crucial importance in

3: order to explain the success of overparametrised neural networks. In this paper, we

4: study the role of the label noise in the training dynamics of a quadratically parametrised model

5: through its continuous time version. We explicitly

6: characterise the solution chosen by the stochastic flow and prove that it implicitly solves a Lasso program. To fully

7: complete our analysis, we provide nonasymptotic convergence guarantees for the dynamics as well as conditions for support recovery. We

8: also give experimental results which support our theoretical claims. Our findings

9: highlight the fact that structured noise can induce better generalisation and

10: help explain the greater performances of stochastic dynamics as observed in practice.

11: \end{abstract}

12: