We consider neural networks (NNs) where the final layer is down-scaled by a fixed hyperparameter $\gamma$. Recent work has identified $\gamma$ as controlling the strength of feature learning. As $\gamma$ increases, network evolution changes from "lazy" kernel dynamics to "rich" feature-learning dynamics, with a host of associated benefits including improved performance on common tasks. In this work, we conduct a thorough empirical investigation of the effect of scaling $\gamma$ across a variety of models and datasets in the online training setting. We first examine the interaction of $\gamma$ with the learning rate $\eta$, identifying several scaling regimes in the $\gamma$-$\eta$ plane which we explain theoretically using a simple model. We find that the optimal learning rate $\eta^*$ scales non-trivially with $\gamma$. In particular, $\eta^* \propto \gamma^2$ when $\gamma \ll 1$ and $\eta^* \propto \gamma^{2/L}$ when $\gamma \gg 1$ for a feed-forward network of depth $L$. Using this optimal learning rate scaling, we proceed with an empirical study of the under-explored "ultra-rich" $\gamma \gg 1$ regime. We find that networks in this regime display characteristic loss curves, starting with a long plateau followed by a drop-off, sometimes followed by one or more additional staircase steps. We find networks of different large $\gamma$ values optimize along similar trajectories up to a reparameterization of time. We further find that optimal online performance is often found at large $\gamma$ and could be missed if this hyperparameter is not tuned. Our findings indicate that analytical study of the large-$\gamma$ limit may yield useful insights into the dynamics of representation learning in performant models.

Online policy optimization (OPO) views policy optimization for sequential decision making as an online learning problem. In this framework, the algorithm designer defines a sequence of online loss functions such that the regret rate in online learning implies the policy convergence rate and the minimal loss witnessed by the policy class determines the policy performance bias. This reduction technique has been successfully applied to solving various policy optimization problems, including imitation learning, structured prediction, and system identification. Interestingly, the policy improvement speed observed in practice is usually much faster than existing theory suggests. In this work, we provide an explanation of this fast policy improvement phenomenon. Let $\epsilon$ denote the policy class bias and assume the online loss functions are convex, smooth, and non-negative. We prove that, after $N$ rounds of OPO with stochastic feedback, the policy converges in $\tilde{O}(1/N + \sqrt{\epsilon/N})$ in both expectation and high probability. In other words, we show that adopting a sufficiently expressive policy class in OPO has two benefits: both the convergence rate increases and the performance bias decreases, as the policy class becomes reasonably rich. This new theoretical insight is further verified in an online imitation learning experiment.