Memorization Capacity of Invariant Convolutional Neural Nets

Deep neural networks are capable of both learning generalizing solutions on natural tasks and memorizing arbitrary mappings. We argue that understanding the former requires understanding the latter, which is much less studied. In what ways do networks that generalize or memorize differ in their behavior or their training dynamics? Do memorizing networks simply function as a lookup table where input-output mappings are stored independently? We investigate these questions in the context of CNNs trained on image classification tasks with randomized labels or pixels. We first study the behavior of such memorizing networks on unseen images. We find that independently initialized networks trained on the same random labels learn different predictive functions: they make different predictions on test examples, in contrast with generalizing networks. We then examine the effect of dataset size on memorization through the required network capacity and training time to interpolate the training set. Counter-intuitively, we find that the number of presentations of a given example required for complete memorization can decrease with dataset size even when input-output pairs share no common structure. Although memorizing networks do not learn a universal solution, they exhibit interesting behaviors which call for further investigations into the underlying mechanisms.

This paper studies the memorization capacity of deep neural networks with ReLU activation. Specifically, we investigate the minimal size of such networks to memorize any $N$ data points in the unit ball with pairwise separation distance $δ$ and discrete labels. Most prior studies characterize the memorization capacity by the number of parameters or neurons. We generalize these results by constructing neural networks, whose width $W$ and depth $L$ satisfy $W^2L^2= \mathcal{O}(N\log(δ^{-1}))$, that can memorize any $N$ data samples. We also prove that any such networks should also satisfy the lower bound $W^2L^2=Ω(N \log(δ^{-1}))$, which implies that our construction is optimal up to logarithmic factors when $δ^{-1}$ is polynomial in $N$. Hence, we explicitly characterize the trade-off between width and depth for the memorization capacity of deep neural networks in this regime.

Given a labeled dataset $\{(x_i, y_i)\}_{i = 1}^n$, consider drawing a parameter vector $\hatθ$ uniformly at random from those that interpolate the data within a width-$W$ depth-$D$ ReLU architecture with continuous weights bounded within $[-M, M]$. We study the generalization properties of $\hatθ$ in the case where the dataset can, in fact, be explained by a smaller $w$-by-$d$ ReLU network. Despite the obliviousness of the selection rule for $\hatθ$, it adaptively generalizes at the rate $\tilde O\left(\frac{wdWD}{n}\right)$, rather than being governed by the VC dimension of the full $W$-by-$D$ hypothesis class. The reason is due to a particular inductive bias in ReLU networks: the $w$-by-$d$ parameter witnesses an exponential volume of $W$-by-$D$ interpolating parameters, which dwarfs the volume of poorly generalizing parameters (conditioned on interpolation). We explicitly construct this set of near-equivalent $W$-by-$D$ parameters by a proof technique called deep lifting, and then tie the achievable volume to stability properties of the $w$-by-$d$ witness. Examples show that these techniques allow us to convert approximation-theoretic results into generalization guarantees for nonparametric classes. We further show that generalization of uniform-random interpolators can survive label corruptions.