# parameter_prediction_for_unseen_deep_architectures__351bf4e5.pdf Parameter Prediction for Unseen Deep Architectures Boris Knyazev1,2 Michal Drozdzal4, Graham W. Taylor1,2,3, Adriana Romero-Soriano4,5, 1 University of Guelph 2 Vector Institute for Artificial Intelligence 3 Canada CIFAR AI Chair 4 Facebook AI Research 5 Mc Gill University equal advising https://github.com/facebookresearch/ppuda Deep learning has been successful in automating the design of features in machine learning pipelines. However, the algorithms optimizing neural network parameters remain largely hand-designed and computationally inefficient. We study if we can use deep learning to directly predict these parameters by exploiting the past knowledge of training other networks. We introduce a large-scale dataset of diverse computational graphs of neural architectures DEEPNETS-1M and use it to explore parameter prediction on CIFAR-10 and Image Net. By leveraging advances in graph neural networks, we propose a hypernetwork that can predict performant parameters in a single forward pass taking a fraction of a second, even on a CPU. The proposed model achieves surprisingly good performance on unseen and diverse networks. For example, it is able to predict all 24 million parameters of a Res Net-50 achieving a 60% accuracy on CIFAR-10. On Image Net, top-5 accuracy of some of our networks approaches 50%. Our task along with the model and results can potentially lead to a new, more computationally efficient paradigm of training networks. Our model also learns a strong representation of neural architectures enabling their analysis. 1 Introduction Consider the problem of training deep neural networks on large annotated datasets, such as Image Net [1]. This problem can be formalized as finding optimal parameters for a given neural network a, parameterized by w, w.r.t. a loss function L on the dataset D = {(xi, yi)}N i=1 of inputs xi and targets yi: i=1 L(f(xi; a, w), yi), (1) where f(xi; a, w) represents a forward pass. Equation 1 is usually minimized by iterative optimization algorithms e.g. SGD [2] and Adam [3] that converge to performant parameters wp of the architecture a. Despite the progress in improving the training speed and convergence [4 7], obtaining wp remains a bottleneck in large-scale machine learning pipelines. For example, training a Res Net-50 [8] on Image Net can take many GPU hours [9]. With the ever growing size of networks [10] and necessity of training the networks repeatedly (e.g. for hyperparameter or architecture search), the classical process of obtaining wp is becoming computationally unsustainable [11 13]. A new parameter prediction task. When optimizing the parameters for a new architecture a, typical optimizers disregard past experience gained by optimizing other nets. However, leveraging this past experience can be the key to reduce the reliance on iterative optimization and, hence the high computational demands. To progress in that direction, we propose a new task where iterative optimization is replaced with a single forward pass of a hypernetwork [14] HD. To tackle the task, 35th Conference on Neural Information Processing Systems (Neur IPS 2021). Part of the work was done while interning at Facebook AI Research. HD is expected to leverage the knowledge of how to optimize other networks F. Formally, the task is to predict the parameters of an unseen architecture a / F using HD, parameterized by θp: ˆwp = HD(a; θp). The task is constrained to a dataset D, so ˆwp are the predicted parameters for which the test set performance of f(x; a, ˆwp) is similar to the one of f(x; a, wp). For example, we consider CIFAR-10 [15] and Image Net image classification datasets D, where the test set performance is classification accuracy on test images. Approaching our task. A straightforward approach to expose HD to the knowledge of how to optimize other networks is to train it on a large training set of {(ai, wp,i)} pairs, however, that is prohibitive2. Instead, we follow the bi-level optimization paradigm common in meta-learning [16 18], but rather than iterating over M tasks, we iterate over M training architectures F = {ai}M i=1: i=1 L f xj; ai, HD(ai; θ) , yj . (2) By optimizing Equation 2, the hypernetwork HD gradually gains knowledge of how to predict performant parameters for training architectures. It can then leverage this knowledge at test time when predicting parameters for unseen architectures. To approach the problem in Equation 2, we need to design the network space F and HD. For F, we rely on the previous design spaces for neural architectures [19] that we extend in two ways: the ability to sample distinct architectures and an expanded design space that includes diverse architectures, such as Res Nets and Visual Transformers [20]. Such architectures can be fully described in the form of computational graphs (Fig. 1). So, to design the hypernetwork HD, we rely on recent advances in machine learning on graph-structured data [21 24]. In particular, we build on the Graph Hyper Networks method (GHNs) [24] that also optimizes Equation 2. However, GHNs do not aim to predict large-scale performant parameters as we do in this work, which motivates us to improve on their approach. By designing our diverse space F and improving on GHNs, we boost the accuracy achieved by the predicted parameters on unseen architectures to 77% (top-1) and 48% (top-5) on CIFAR-10 [15] and Image Net [1], respectively. Surprisingly, our GHN shows good out-of-distribution generalization and predicts good parameters for architectures that are much larger and deeper compared to the ones seen in training. For example, we can predict all 24 million parameters of Res Net-50 in less than a second either on a GPU or CPU achieving 60% on CIFAR-10 without any gradient updates (Fig 1, (b)). Overall, our framework and results pave the road toward a new and significantly more efficient paradigm for training networks. Our contributions are as follows: (a) we introduce the novel task of predicting performant parameters for diverse feedforward neural networks with a single hypernetwork forward pass; (b) we introduce DEEPNETS-1M a standardized benchmark with in-distribution and out-of-distribution architectures to track progress on the task ( 3); (c) we define several baselines and propose a GHN model ( 4) that performs surprisingly well on CIFAR-10 and Image Net ( 5.1); (d) we show that our model learns a strong representation of neural network architectures ( 5.2), and our model is useful for initializing neural networks ( 5.3). Our DEEPNETS-1M dataset, trained GHNs and code is available at https://github.com/facebookresearch/ppuda. Meta-batching Virtual Parameter normalization Computational graph Backward pass Forward pass Example of evaluating on an unseen architecture a / F (Res Net-50) Vanilla GHN Our GHN 1 fw pass SGD 50 epochs CIFAR-10 accuracy (%) (a) (b) Figure 1: (a) Overview of our GHN model ( 4) trained by backpropagation through the predicted parameters ( ˆwp) on a given image dataset and our DEEPNETS-1M dataset of architectures. Colored captions show our key improvements to vanilla GHNs ( 2.2). The red one is used only during training GHNs, while the blue ones are used both at training and testing time. The computational graph of a1 is visualized as described in Table 1. (b) Comparing classification accuracies when all the parameters of a Res Net-50 are predicted by GHNs versus when its parameters are trained with SGD (see full results in 5). 2Training a single network ai can take several GPU days and thousands of trained networks may be required. 2 Background We start by providing a brief background about the network design spaces leveraged in the creation of our DEEPNETS-1M dataset of neural architectures described in 3. We then cover elements of graph hypernetworks that we leverage when designing our specific GHN HD in 4. 2.1 Network Design Space of DARTS DARTS [19] is a differentiable NAS framework. For image classification tasks such as those considered in this work, its networks are defined by four types of building blocks: stems, normal cells, reduction cells, and classification heads. Stems are fixed blocks of convolutional operations that process input images. The normal and reduction cells are the main blocks of architectures and are composed of: 3 3 and 5 5 separable convolutions, 3 3 and 5 5 dilated separable convolutions, 3 3 max pooling, 3 3 average pooling, identity and zero (to indicate the absence of connectivity between two operations). Finally, the classification head defines the network output and is built with a global pooling followed by a single fully connected layer. Typically, DARTS networks have one stem block, 14-20 cells, and one classification head, altogether forming a deep computational graph. The reduction cells, placed only at 1/3 and 2/3 of the total depth, decrease the spatial resolution and increase the channel dimensionality by a factor of 2. Summation and concatenation are used to aggregate outputs from multiple operations within each cell. To make the channel dimensionalities match, 1 1 convolutions are used as needed. All convolutional operations use the Re LU-Conv-Batch Norm (BN) [7] order. Overall, DARTS enables defining strong architectures that combine many principles of manual [25, 8, 26, 27] and automatic [24, 28 33] design of neural architectures. While DARTS learns the optimal task-specific cells, the framework can be modified to permit sampling randomly-structured cells. We leverage this possibility for the DEEPNETS-1M construction in 3. Please see A.1 for further details on DARTS. 2.2 Graph Hyper Network: GHN-1 Representation of architectures. GHNs [24] directly operate on the computational graph of a neural architecture a. Specifically, a is a directed acyclic graph (DAG), where nodes V = {vi}|V | i=1 are operations (e.g. convolutions, fully-connected layers, summations, etc.) and their connectivity is described by a binary adjacency matrix A {0, 1}|V | |V |. Nodes are further characterized by a matrix of initial node features H0 = [h0 1, h0 2, ..., h0 |V |], where each h0 v is a one-hot vector representing the operation performed by the node. We also use such a one-hot representation for H0, but in addition encode the shape of parameters associated with nodes as described in detail in B.1. Design of the graph hypernetwork. In [24], the graph hypernetwork HD consists of three key modules. The first module takes the input node features H0 and transforms them into d-dimensional node features H1 R|V | d through an embedding layer. The second module takes H1 together with A and feeds them into a specific variant of the gated graph neural network (Gated GNN) [34]. In particular, their Gated GNN mimics the canonical order π of node execution in the forward (fw) and backward (bw) passes through a computational graph. To do so, it sequentially traverses the graph and performs iterative message passing operations and node feature updates as follows: t [1, ..., T] : h π [fw, bw] : v π : mt v = X u N π v MLP(ht u), ht v = GRU(ht v, mt v) i , (3) where T denotes the total number of forward-backward passes; ht v corresponds to the features of node v in the t-th graph traversal; MLP( ) is a multi-layer perceptron; and GRU( ) is the update function of the Gated Recurrent Unit [35]. In the forward propagation (π = fw), N π v corresponds to the incoming neighbors of the node defined by A, then in the backward propagation (π = bw) it similarly corresponds to the outgoing neighbors of the node. The last module uses the Gated GNN output hidden states h T v to condition a decoder that produces the parameters ˆwv p (e.g. convolutional weights) associated with each node. In practice, to handle different parameter dimensionalities per operation type, the output of the hypernetwork is reshaped and sliced according to the shape of parameters in each node. We refer to the model described above as GHN-1 (Fig. 1). Further subtleties of implementing this model in the context of our task are discussed in B.1. Table 1: Examples of computational graphs (visualized using Network X [44]) in each split and their key statistics, to which we add the average degree and average shortest path length often used to measure local and global graph properties respectively [45, 46]. In the visualized graphs, a node is one of the 15 primitives coded with markers shown at the bottom, where they are sorted by the frequency in the training set. For visualization purposes, a blue triangle marker differentiates a 1 1 convolution (equivalent to a fully-connected layer over channels) from other convolutions, but its primitive type is still just convolution. *Computed based on CIFAR-10. IN-DISTRIBUTION OUT-OF-DISTRIBUTION TRAIN VAL/TEST WIDE DEEP DENSE BN-FREE RESNET/VIT #graphs 106 500/500 100 100 100 100 1/1 #cells 4-18 4-18 4-18 10-36 4-18 4-18 16/12 #channels 16-128 32-128 128-1216 32-208 32-240 32-336 64/128 #nodes (|V |) 21-827 33-579 33-579 74-1017 57-993 33-503 161/114 % w/o BN 3.5% 4.1% 4.1% 2.0% 5.0% 100% 0%/100% #params(M)* 0.01-3.1 2.5-35 39-101 2.5-15.3 2.5-8.8 2.5-7.7 23.5/1.0 avg degree 2.3 0.1 2.3 0.1 2.3 0.1 2.3 0.1 2.4 0.1 2.4 0.1 2.2/2.3 avg path 14.5 4.8 14.5 4.9 14.7 4.9 26.2 9.3 15.1 4.1 10.0 2.8 11.2/10.7 primitive conv BN sum bias group conv concat dilated gr. conv LN max pool avg pool MSA SE input glob avg pos enc fraction in TRAIN (%) 36.3 25.5 11.1 6.5 5.1 3.8 2.5 2.5 1.8 1.7 1.2 1.0 0.5 0.5 0.2 3 DEEPNETS-1M The network design space of DARTS is limited by the number of unique operations that compose cells, and the low variety of stems and classification heads. Thus, many architectures are not realizable within this design space, including: VGG [25], Res Nets [8], Mobile Net [33] or more recent ones such as Visual Transformer (Vi T) [20] and Normalization-free networks [36, 37]. Furthermore, DARTS does not define a procedure to sample random architectures. By addressing these two limitations we aim to expose our hypernetwork to diverse training architectures and permit its evaluation on common architectures, such as Res Net-50. We hypothesize that increased training diversity can improve hypernetworks generalization to unseen architectures making it more competitive to iterative optimizers. Extending the network design space. We extend the set of possible operations with non-separable 2D convolutions3, Squeeze&Excite4 (SE) [40] and Transformer-based operations [41, 20]: multihead self-attention (MSA), positional encoding and layer norm (LN) [42]. Each node (operation) in our graphs has two attributes: primitive type (e.g. convolution) and shape (e.g. 3 3 512 512). Overall, our extended set consists of 15 primitive types (Table 1). We also extend the diversity of the generated architectures by introducing VGG-style classification heads and Vi T stems. Finally, to further increase architectural diversity, we allow the operations to not include batch norm (BN) [7] and permit networks without channel width expansion (e.g. as in [20]). Architecture generation process. We generate different subsets of architectures (see the description of each subset in the next two paragraphs and in Table 1). For each subset depending on its purpose, we predefine a range of possible model depths (number of cells), widths and number of nodes per cell. Then, we sample a stem, a normal and reduction cell and a classification head. The internal structure of the normal and reduction cells is defined by uniformly sampling from all available operations. Due to a diverse design space it is extremely unlikely to sample the same architecture multiple times, but we ran a sanity check using the Hungarian algorithm [43] to confirm that (see Figure 6 in A.2 for details). In-distribution (ID) architectures. We generate a training set of |F| = 106 architectures and validation/test sets of 500/500 architectures that follow the same generation rules and are considered to be ID samples. However, training on large architectures can be prohibitive, e.g. in terms of GPU memory. Thus, in the training set we allow the number of channels and, hence the total number of parameters, to be stochastically defined given computational resources. For example, to train 3Non-separable convolutions have weights of e.g. shape 3 3 512 512 as in Res Net-50. NAS works, such as DARTS and GHN, avoid such convolutions, since the separable ones [38] are more efficient. Non-separable convolutions are nevertheless common in practice and can often boost the downstream performance. 4The Squeeze&Excite operation is common in many efficient networks [39, 12]. our models we upper bound the number of parameters in the training architectures to around 3M by sampling fewer channels if necessary. In the evaluation sets, the number of channels is fixed. Therefore, this pre-processing step prior to training results in some distribution shift between the training and the validation/test sets. However, the shift is not imposed by our dataset. Out-of-distribution (OOD) architectures. We generate five OOD test sets that follow different generation rules. In particular, we define WIDE and DEEP sets that are of interest due the stronger downstream performance of such nets in large-scale tasks [47, 48, 10]. These nets are often more challenging to train for fundamental [49, 50] or computational [51] reasons, so predicting their parameters might ease their subsequent optimization. We also define the DENSE set, since networks with many operations per cell and complex connectivity are underexplored in the literature despite their potential [27]. Next, we define the BN-FREE set that is of interest due to BN s potential negative side-effects [52, 53] and the difficulty or unnecessity of using it in some cases [54 56, 36, 37]. We finally add the RESNET/VIT set with two predefined image classification architectures: commonlyused Res Net-50 [8] and a smaller 12-layer version of the Visual Transformer (Vi T) [20] that has recently received a lot of attention in the vision community. Please see A.1 and A.2 for further details and statistics of our DEEPNETS-1M dataset. 4 Improved Graph Hyper Networks: GHN-2 In this section, we introduce our three key improvements to the baseline GHN-1 described in 2.2 (Fig. 1). These components are essential to predict stronger parameters on our task. For the empirical validation of the effectiveness of these components see ablation studies in 5.1 and C.2.1. 4.1 Differentiable Normalization of Predicted Parameters Table 2: Parameter normalizations. Type of node v Normalization Conv./fully-conn. ˆwv p q β/(Cin HW) Norm. weights 2 sigmoid( ˆwv p/T ) Biases tanh( ˆwv p/T ) When training the parameters of a given network from scratch using iterative optimization methods, the initialization of parameters is crucial. A common approach is to use He [57] or Glorot [58] initialization to stabilize the variance of activations across layers of the network. Chang et al. [59] showed that when the parameters of the network are instead predicted by a hypernetwork, the activations in the network tend to explode or vanish. To address the issue of unstable network activations especially for the case of predicting parameters of diverse architectures, we apply operation-dependent normalizations (Table 2). We normalize convolutional and fully-connected weights by following the fan-in scheme of [57] (see the comparison to fan-out in C.2.1): ˆwv p p β/(Cin HW), where Cin, H, W are the number of input channels and spatial dimensions of weights ˆwv p, respectively; and β is a nonlinearity specific constant following the analysis in [57]. The parameters of normalization layers such as BN and LN, as well as biases typically initialized with constants, are normalized by applying a squashing function with temperature T to imitate the empirical distributions of models trained with SGD (see Table 2). These are differentiable normalizations, so that they are applied at training (and testing) time. Further analysis of our normalization and its stabilizing effect on activations is presented in B.2.2. 4.2 Enhancing Long-range Message Propagation Figure 2: Virtual edges (in green) allow for better capture of global context. Computational graphs often take the form of long chains (Table 1) with only a few incoming/outcoming edges per node. This structure might hinder long-range propagation of information between nodes [60]. Different approaches to alleviate the long-range propagation problem exist [61 63], including stacking GHNs in [24]. Instead we adopt simple graph-based heuristics in line with recent works [64, 65]. In particular, we add virtual edges between two nodes v and u and weight them based on the shortest path svu between them (Fig. 2). To avoid interference with the real edges in the computational graph, we introduce a separate MLPsp to transform the features of the nodes connected through these virtual edges, and redefine the message passing of Equation 3 as: u N π v MLP(ht u) + X u N (sp) v 1 svu MLPsp(ht u), (4) where N (sp) v are neighbors satisfying 1 < svu s(max), and s(max) is a hyperparameter. To maintain the same number of trainable parameters as in GHN-1, we decrease MLPs sizes appropriately. Despite its simplicity, this approach is effective (see the comparison to stacking GHNs in C.2.1). 4.3 Meta-batching Architectures During Training GHN-1 updates its parameters θ based on a single architecture sampled for each batch of images (Equation 2). In vanilla SGD training, larger batches of images often speed up convergence by reducing gradient noise and improve model s performance [66]. Therefore, we define a meta-batch bm as the number of architectures sampled per batch of images. Both the parameter prediction and the forward/backward passes through the architectures in a meta-batch can be done in parallel. We then average the gradients across bm to update the parameters θ of HD: θL = 1/bm Pbm i=1 θLi. Further analysis of the meta-batching effect on the training loss and convergence speed is presented in B.2.3. 5 Experiments We focus the evaluation of GHN-2 on our parameter prediction task ( 5.1). In addition, we show beneficial side-effects of i) learning a stronger neural architecture representation using GHN-2 in analyzing networks ( 5.2) and ii) predicting parameters for fine-tuning ( 5.3). We provide further experimental and implementation details, as well as more results supporting our arguments in C. Datasets. We use the DEEPNETS-1M dataset of architectures ( 3) as well as two image classification datasets D1 (CIFAR-10 [15]) and D2 (Image Net [1]). CIFAR-10 consists of 50k training and 10k test images of size 32 32 3 and 10 object categories. Image Net is a larger scale dataset with 1.28M training and 50k test images of variable size and 1000 fine-grained object categories. We resize Image Net images to 224 224 3 following [19, 24]. We use 5k/50k training images as a validation set in CIFAR-10/Image Net and 500 validation architectures of DEEPNETS-1M for hyperparameter tuning. Baselines. Our baselines include GHN-1 and a simple MLP that only has access to operations, but not to the connections between them. This MLP baseline is obtained by replacing the Gated GNN with an MLP in our GHN-2. Since GHNs were originally introduced for small architectures of 50 nodes and only trained on CIFAR-10, we reimplement5 them and scale them up by introducing minor modifications to their decoder that enable their training on Image Net and on larger architectures of up to 1000 nodes (see B.1 for details). We use the same hyperparameters to train the baselines and GHN-2. Iterative optimizers. In the parameter prediction experiments, we also compare our model to standard optimization methods: SGD and Adam [3]. We use off-the-shelf hyperparameters common in the literature [24, 19, 32, 67 69]. On CIFAR-10, we train evaluation architectures with SGD/Adam, initial learning rate η = 0.025 / η = 0.001, batch size b = 96 and up to 50 epochs. With Adam, we train only 300 evaluation architectures as a rough estimation of an average performance. On Image Net, we train them with SGD, η = 0.1 and b = 128, and, for computational reasons (given 1402 evaluation architectures in total), we limit training with SGD to 1 epoch. We have also considered meta-optimizers, such as [17, 18]. However, we were unable to scale them to diverse and large architectures of our DEEPNETS-1M, since their LSTM requires a separate hidden state for every trainable parameter in the architecture. The scalable variants exist [70, 71], but are hard to reproduce without open source code. Additional experimental details. We follow [24] and train GHNs with Adam, η = 0.001 and batch size of 64 images for CIFAR-10 and 256 for Image Net. We train for up to 300 epochs, except for one experiment in the ablation studies, where we train one GHN with bm = 1 eight times longer, i.e. for 2400 epochs. All GHNs in our experiments use T = 1 propagation (Equation 3), as we found the original T = 5 of [24] to be inefficient and it did not improve the accuracies in our task. GHN-2 uses s(max) = 50 and bm = 8 and additionally uses LN that slightly further improves results (see these ablations in C.2.1). Model selection is performed on the validation sets, but the results in our paper are reported on the test sets to enable their direct comparison. 5.1 Parameter Prediction Experimental setup. We trained our GHN-2 and baselines on the training architectures and training images, i.e. a separate model is trained for CIFAR-10 and Image Net. According to our DEEPNETS-1M benchmark, we assess whether these models can generalize to unseen in-distribution (ID) and out-of-distribution (OOD) test architectures from our DEEPNETS-1M. We measure this generalization by predicting parameters for the test architectures and computing their classification accuracies on the test images of CIFAR-10 (Table 3) and Image Net (Table 4). The evaluation architectures with batch norm (BN) have running statistics, which are not learned by gradient descent [7], and 5While source code for GHNs [24] is unavailable, we appreciate the authors help in implementing some steps. hence are not predicted by our GHNs. To alleviate that, we follow [24] and evaluate the networks with BN by computing per batch statistics with batch size of 64 images. This is further discussed in C.1. Results. Despite GHN-2 never observed the test architectures, GHN-2 predicts good parameters for them making the test networks perform surprisingly well on both image datasets (Tables 3 and 4). Our results are especially strong on CIFAR-10, where some architectures with predicted parameters achieve up to 77.1%, while the best accuracy of training with SGD for 50 epochs is around 15% more. We even show good results on Image Net, where for some architectures we achieve a top-5 accuracy of up to 48.3%. While these results are low for direct downstream applications, they are remarkable for three main reasons. First, to train GHNs by optimizing Equation 2, we do not rely on the prohibitively expensive procedure of training the architectures F by SGD. Second, GHNs rely on a single forward pass to predict all parameters. Third, these results are obtained for unseen architectures, including the OOD ones. Even in the case of severe distribution shifts (e.g. Res Net-506) and underrepresented networks (e.g. Vi T7), our model still predicts parameters that perform better than random ones. On CIFAR-10, generalization of GHN-2 is particularly strong with a 58.6% accuracy on Res Net-50. On both image datasets, our GHN-2 significantly outperforms GHN-1 on all test subsets of DEEPNETS-1M with more than a 20% absolute gain in certain cases, e.g. 36.8% vs 13.7% on the BN-FREE networks (Table 3). Exploiting the structure of computational graphs is a critical property of GHNs with the accuracy dropping from 66.9% to 42.2% on ID (and even more on OOD) architectures when we replace the Gated GNN of GHN-2 with an MLP. Compared to iterative optimization methods, GHN-2 predicts parameters achieving an accuracy similar to 2500 and 5000 iterations of SGD on CIFAR-10 and Image Net respectively. In contrast, GHN-1 performs similarly to only 500 and 2000 (not shown in Table 4) iterations respectively. Comparing SGD to Adam, the latter performs worse in general except for the Vi T architectures similar to [72, 20]. To report speeds on Image Net in Table 4, we use a dedicated machine with a single NVIDIA V100-32GB and Intel Xeon CPU E5-1620 v4@ 3.50GHz. So for SGD these numbers can be reduced by using faster computing infrastructure and more optimal hyperparameters [73]. Using our setup, Table 3: CIFAR-10 results of predicted parameters for unseen ID and OOD architectures of DEEPNETS-1M. Mean ( standard error of the mean) accuracies are reported (random chance 10%). The number of parameter updates. METHOD #upd ID-TEST OOD-TEST avg max WIDE DEEP DENSE BN-FREE RESNET/VIT MLP 1 42.2 0.6 60.2 22.3 0.9 37.9 1.2 44.8 1.1 23.9 0.7 17.7/10.0 GHN-1 1 51.4 0.4 59.9 43.1 1.7 48.3 0.8 51.8 0.9 13.7 0.3 19.2/18.2 GHN-2 1 66.9 0.3 77.1 64.0 1.1 60.5 1.2 65.8 0.7 36.8 1.5 58.6/11.4 Iterative optimizers (all architectures are ID in this case) SGD (1 epoch) 0.5 103 46.1 0.4 66.5 47.2 1.1 34.2 1.1 45.3 0.7 18.0 1.1 61.8/34.5 SGD (5 epochs) 2.5 103 69.2 0.4 82.4 71.2 0.3 56.7 1.6 67.8 0.9 29.0 2.0 78.2/52.5 SGD (50 epochs) 25 103 88.5 0.3 93.1 88.9 1.2 84.5 1.2 87.3 0.8 45.6 3.6 93.5/75.7 Adam (50 epochs) 25 103 84.0 0.8 89.5 82.0 1.6 76.2 2.6 84.8 0.4 38.8 4.8 91.5/79.4 Table 4: Image Net results on DEEPNETS-1M. Mean ( standard error of the mean) top-5 accuracies are reported (random chance 0.5%). Estimated on Res Net-50 with batch size 128. METHOD #upd GPU sec. CPU sec. ID-TEST OOD-TEST avg avg avg max WIDE DEEP DENSE BN-FREE RESNET/VIT GHN-1 1 0.3 0.5 17.2 0.4 32.1 15.8 0.9 15.9 0.8 15.1 0.7 0.5 0.0 6.9/0.9 GHN-2 1 0.3 0.7 27.2 0.6 48.3 19.4 1.4 24.7 1.4 26.4 1.2 7.2 0.6 5.3/4.4 Iterative optimizers (all architectures are ID in this case) SGD (1 step) 1 0.4 6.0 0.5 0.0 0.7 0.5 0.0 0.5 0.0 0.5 0.0 0.5 0.0 0.5/0.5 SGD (5000 steps) 5k 2 103 3 104 25.6 0.3 50.7 26.2 1.4 13.2 1.1 25.4 1.1 4.8 0.8 34.8/24.3 SGD (10000 steps) 10k 4 103 6 104 37.7 0.6 62.0 38.7 1.6 22.1 1.4 36.3 1.2 8.0 1.2 49.0/33.4 SGD (100 epochs) 1000k 6 105 6 107 92.9/72.2 6Large architectures with bottleneck layers such as Res Net-50 do not appear during training. 7Architectures such as Vi T do not include BN and, except for the first layer, convolutions the two most frequent operations in the training set. SGD requires on average 104 more time on a GPU (105 on a CPU) to obtain parameters that yield performance similar to GHN-2. As a concrete example, Alex Net [74] requires around 50 GPU hours (on our setup) to achieve a 81.8% top-5 accuracy, while on some architectures we achieve 48.0% in just 0.3 GPU seconds. Table 5: Ablating GHN-2 on CIFAR-10. An average rank of the model is computed across all ID and OOD test architectures. MODEL ID-TEST OOD-TEST AVG. RANK GHN-2 66.9 0.3 56.8 0.8 1.9 1000 training architectures 65.1 0.5 52.5 1.0 2.6 No normalization ( 4.1) 62.6 0.6 47.1 1.2 3.9 No virtual edges ( 4.2) 61.5 0.4 53.9 0.6 4.1 No meta-batch (bm = 1, 4.3) 54.3 0.3 47.5 0.6 5.5 bm = 1, train 8 longer 62.4 0.5 51.9 1.0 3.7 No Gated GNN (MLP) 42.2 0.6 32.2 0.7 7.4 GHN-1 51.4 0.4 39.2 0.9 6.8 Test accuracy Test bm=1 (ID) bm=8 (ID) bm=1 (OOD) bm=8 (OOD) 1 10 102 103 104 105 106 # Training architectures Train accuracy Figure 3: GHN-2 with meta batch bm = 8 versus bm = 1 for different numbers of training architectures on CIFAR-10. Ablations (Table 5) show that all three components proposed in 4 are important. Normalization is particularly important for OOD generalization with the largest drops on the WIDE and BN-FREE networks (see C.2.1). Using meta-batching (bm = 8) is also essential and helps stabilize training and accelerate convergence (see B.2). We also confirm that the performance gap between bm = 1 and bm = 8 is not primarily due to the observation of more architectures, since the ablated GHN-2 with bm = 1 trained eight times longer is still inferior. The gap between bm = 8 and bm = 1 becomes pronounced with at least 1k training architectures (Fig. 3). When training with fewer architectures (e.g. 100), the GHN with meta-batching starts to overfit to the training architectures. Given our challenging setup with unseen evaluation architectures, it is surprising that using 1k training architectures already gives strong results. However, OOD generalization degrades in this case compared to using all 1M architectures, especially on the BN-FREE networks (see B.2). When training GHNs on just a few architectures, the training accuracy soars to the level of training them with SGD. With more architectures, it generally decreases indicating classic overfitting and underfitting cases. 5.2 Property Prediction Representing computational graphs of neural architectures is a challenging problem [75 79]. We verify if GHNs are capable of doing that out-of-the-box in the property prediction experiments. We also experiment with architecture comparison in C.2.4. Our hypothesis is that by better solving our parameter prediction task, GHNs should also better solve graph representation tasks. Experimental setup. We predict the properties of architectures given their graph embeddings obtained by averaging node features8. We consider four such properties (see C.2.3 for details): Accuracy on the clean (original) validation set of images; Accuracy on a corrupted set (obtained by adding the Gaussian noise to images following [53]); Inference speed (latency or GPU seconds per a batch of images); Convergence speed (the number of SGD iterations to achieve a certain training accuracy). Estimating these properties accurately can have direct practical benefits. Clean and corrupted accuracies can be used to search for the best performing architectures (e.g. for the NAS task); inference speed can be used to choose the fastest network, so by estimating these properties we can trade-off accurate, robust and fast networks [12]. Convergence speed can be used to find networks that are easier to optimize. These properties correlate poorly with each other and between CIFAR-10 and Image Net ( C.2.3), so they require the model to capture different regularities of graphs. While specialized methods to estimate some of these properties exist, often as a NAS task [80 82, 30, 75], our GHNs provide a generic representation that can be easily used for many such properties. For each property, we train a simple regression model using graph embeddings and ground truth property values. We use 500 validation architectures of DEEPNETS-1M for training the regression model and tuning its hyperparameters (see C.2.3 for details). We then use 500 testing architectures of DEEPNETS-1M to measure Kendall s Tau rank correlation between the predicted and ground truth property values similar to [80]. 8A fixed size graph embedding for the architecture a can be computed by averaging the output node features: ha = 1 |V | P v V h T v , where ha Rd and d is the dimensionality of node features. Additional baseline. We compare to the Neural Predictor (Neu Pred) [80]. Neu Pred is based on directed graph convolution and is developed for accuracy prediction achieving strong NAS results. We train a separate such Neu Pred for each property from scratch following their hyperparameters. Clean image Noisy image Convergence Kendall s Tau correlation Neu Pred GHN (MLP) GHN-1 GHN-2 GHN-1 GHN-2 Figure 4: Property prediction of neural networks in terms of correlation (higher is better). Error bars denote the standard deviation across 5 runs. Results. GHN-2 consistently outperforms the GHN-1 and MLP baselines as well as Neu Pred (Fig. 4). In C.2.3, we also provide results verifying if higher correlations translate to downstream gains. For example, on CIFAR-10 by choosing the most accurate architecture according to the regression model and training it from scratch following [19, 24], we obtained a 97.26%( 0.09) accuracy, which is competitive with leading NAS approaches, e.g. [19, 24, 32, 67 69]. In contrast, the network chosen by the regression model trained on the GHN-1 embeddings achieves 95.90%( 0.08). 5.3 Fine-tuning Predicted Parameters Neural networks trained on Image Net and other large datasets have proven useful in diverse visual tasks in the transfer learning setup [83 87, 20]. Therefore, we explore how predicting parameters on Image Net with GHNs compares to pretraining them on Image Net with SGD in such a setup. We consider low-data tasks as they often benefit more from transfer learning [86, 87]. Experimental setup. We perform two transfer-learning experiments. The first experiment is finetuning the predicted parameters on 1,000 training samples (100 labels per class) of CIFAR-10. We fine-tune Res Net-50, Visual Transformer (Vi T) and a 14-cell architecture based on the DARTS best cell [19]. The hyperparameters of fine-tuning (initial learning rate and weight decay) are tuned on 200 validation samples held-out of the 1,000 training samples. The number of epochs is fixed to 50 as in 5.1 for simplicity. In the second experiment, we fine-tune the predicted parameters on the object detection task. We closely follow the experimental protocol and hyperparameters from [88] and train the networks on the Penn-Fudan dataset [89]. The dataset contains only 170 images and the task is to detect pedestrians. Therefore this task is also well suited for transfer learning. Following [88], we replace the backbone of a Faster R-CNN with one of the three architectures. To perform transfer learning with GHNs, in both experiments we predict the parameters of a given architecture using GHNs trained on Image Net. We then replace the Image Net classification layer with the target task-specific layers and fine-tune the entire network on the target task. We compare the results of GHNs to He s initialization [57] and the initialization based on pretraining the parameters on Image Net with SGD. Table 6: CIFAR-10 test set accuracies and Penn-Fudan object detection average precision (at Io U=0.50) after fine-tuning the networks using SGD initialized with different methods. Average results and standard deviations for 3 runs with different random seeds are shown. For each architecture, similar GHN-2-based and Image Net-based results are bolded.*Estimated on Res Net-50. INITIALIZATION METHOD GPU sec. to init.* 100-SHOT CIFAR-10 PENN-FUDAN OBJECT DETECTION RESNET-50 VIT DARTS RESNET-50 VIT DARTS He s [57] 0.003 41.0 0.4 33.2 0.3 45.4 0.4 0.197 0.042 0.144 0.010 0.486 0.035 GHN-1 (trained on Image Net) 0.6 46.6 0.0 23.3 0.1 49.2 0.1 0.433 0.013 0.0 0.0 0.468 0.024 GHN-2 (trained on Image Net) 0.7 56.4 0.1 41.4 0.6 60.7 0.3 0.560 0.019 0.436 0.032 0.785 0.032 Image Net (1k pretraining steps) 6 102 45.4 0.3 44.3 0.1 62.4 0.3 0.302 0.022 0.182 0.046 0.814 0.033 Image Net (2.5k pretraining steps) 1.5 103 55.4 0.2 50.4 0.3 70.4 0.2 0.571 0.056 0.322 0.073 0.823 0.022 Image Net (5 pretraining epochs) 3 104 84.6 0.2 70.2 0.5 83.9 0.1 0.723 0.045 0.391 0.024 0.827 0.053 Image Net (final epoch) 6 105 89.2 0.2 74.5 0.2 85.6 0.2 0.876 0.011 0.468 0.023 0.881 0.023 Results. The CIFAR-10 image classification results of fine-tuning the parameters predicted by our GHN-2 are 10 percentage points better (in absolute terms) than fine-tuning the parameters predicted by GHN-1 or training the parameters initialized using He s method (Table 6). Similarly, the object detection results of GHN-2-based initialization are consistently better than both GHN-1 and He s initializations. The GHN-2 results are a factor of 1.5-3 improvement over He s for all the three architectures. Overall, the two experiments clearly demonstrate the practical value of predicting parameters using our GHN-2. Using GHN-1 for initialization provides relatively small gains or hurts convergence (for Vi T). Compared to pretraining on Image Net with SGD, initialization using GHN-2 leads to performance sim- ilar to 1k-2.5k steps of pretraining on Image Net depending on the architecture in the case of CIFAR-10. In the case of Penn-Fudan, GHN-2 s performance is similar to 1k steps of pretraining with SGD. In both experiments, pretraining on Image Net for just 5 epochs provides strong transfer learning performance and the final Image Net checkpoints are only slightly better, which aligns with previous works [85]. Therefore, further improvements in the parameter prediction models appear promising. 6 Related Work Our proposed parameter prediction task, objective in Equation 2 and improved GHN are related to a wide range of machine learning frameworks, in particular meta-learning and neural architecture search (NAS). Meta-learning is a general framework [16, 90] that includes meta-optimizers and meta-models, among others. Related NAS works include differentiable [19] and one-shot methods [12]. See additional related work in D. Meta-optimizers. Meta-optimizers [17, 18, 71, 91, 92] define a problem similar to our task, but where HD is an RNN-based model predicting the gradients w, mimicking the behavior of iterative optimizers. Therefore, the objective of meta-optimizers may be phrased as learning to optimize as opposed to our learning to predict parameters. Such meta-optimizers can have their own hyperparameters that need to be tuned for a given architecture a and need to be run expensively (on the GPU) for many iterations following Equation 1. Meta-models. Meta-models include methods based on MAML [93], Proto Nets [94] and auxiliary nets predicting task-specific parameters [95 98]. These methods are tied to a particular architecture and need to be trained from scratch if it is changed. Several recent methods attempt to relax the choice of architecture in meta-learning. T-NAS [99] combines MAML with DARTS [19] to learn both the optimal architecture and its parameters for a given task. However, the best network, a, needs to be trained using MAML from scratch. Meta-NAS [100] takes a step further and only requires fine-tuning of a on a given task. However, the a is obtained from a single meta-architecture and so its choice is limited, preventing parameter prediction for arbitrary a. CATCH [101] follows a similar idea, but uses reinforcement learning to quickly search for the best a on the specific task. Overall meta-learning mainly aims at generalization across tasks, often motivated by the few-shot learning problem. In contrast, our parameter prediction problem assumes a single task (here an image dataset), but aims at generalization across architectures a with the ability to predict parameters in a single forward pass. One-shot NAS. One-shot NAS aims to learn a single supernet [102, 12, 103] that can be used to estimate the performance of smaller nets (subnets) obtained by some kind of pruning the supernet, followed by training the best chosen a from scratch with SGD. Recent models, in particular Big NAS [12] and Once For All (OFA) [102], eliminate the need to train subnets. However, the fundamental limitation of one-shot NAS is poor scaling with the number of possible computational operations [24]. This limits the diversity of architectures for which parameters can be obtained. For example, all subnets in OFA are based on Mobile Net-v3 [33], which does not allow to solve our more general parameter prediction task. To mitigate this, SMASH [104] proposed to predict some of the parameters using hypernetworks [14] by encoding architectures as a 3D tensor. Graph Hyper Networks (GHNs) [24] further generalized this approach to arbitrary computational graphs (DAGs), which allowed them to improve NAS results. GHNs focused on obtaining reliable subnetwork rankings for NAS and did not aim to predict large-scale performant parameters. We show that the vanilla GHNs perform poorly on our parameter prediction task mainly due to the inappropriate scale of predicted parameters, lack of long-range interactions in the graphs, gradient noise and slow convergence when optimizing Equation 2. Conventionally to NAS, GHNs were also trained in a quite constrained architecture space [105]. We expand the architecture space adopting GHNs for a more general problem. 7 Conclusion We propose a novel framework and benchmark to learn and evaluate neural parameter prediction models. Our model (GHN-2) is able to predict parameters for very diverse and large-scale architectures in a single forward pass in a fraction of a second. The networks with predicted parameters yield surprisingly high image classification accuracy given the extremely challenging nature of our parameter prediction task. However, the accuracy is still far from networks trained with handcrafted optimization methods. Bridging the gap is a promising future direction. As a beneficial side-effect, GHN-2 learns a strong representation of neural architectures as evidenced by our property prediction evaluation. Finally, parameters predicted using GHN-2 trained on Image Net benefit transfer learning in the low-data regime. This motivates further research towards solving our task. Acknowledgments BK is thankful to Facebook AI Research for funding the initial phase of this research during his internship and to NSERC and the Ontario Graduate Scholarship used to fund the other phases of this research. GWT and BK also acknowledge support from CIFAR and the Canada Foundation for Innovation. Resources used in preparing this research were provided, in part, by the Province of Ontario, the Government of Canada through CIFAR, and companies sponsoring the Vector Institute: http://www.vectorinstitute.ai/#partners. We are thankful to Magdalena Sobol for editorial help. We are thankful to the Vector AI Engineering team (Gerald Shen, Maria Koshkina and Deval Pandya) for code review. We are also thankful to the reviewers for their constructive feedback. [1] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211 252, 2015. [2] Sebastian Ruder. An overview of gradient descent optimization algorithms. ar Xiv preprint ar Xiv:1609.04747, 2016. [3] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. ar Xiv preprint ar Xiv:1412.6980, 2014. [4] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In European conference on computer vision, pages 646 661. Springer, 2016. [5] Andrew Brock, Theodore Lim, James M Ritchie, and Nick Weston. Freezeout: Accelerate training by progressively freezing layers. ar Xiv preprint ar Xiv:1706.04983, 2017. [6] Dami Choi, Alexandre Passos, Christopher J Shallue, and George E Dahl. Faster neural network training with data echoing. ar Xiv preprint ar Xiv:1907.05550, 2019. [7] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. ar Xiv preprint ar Xiv:1502.03167, 2015. [8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770 778, 2016. [9] NVIDIA. Nvidia data center deep learning product performance. URL https://developer.nvidia. com/deep-learning-performance-training-inference. [10] Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. ar Xiv preprint ar Xiv:2005.14165, 2020. [11] Emma Strubell, Ananya Ganesh, and Andrew Mc Callum. Energy and policy considerations for deep learning in nlp. ar Xiv preprint ar Xiv:1906.02243, 2019. [12] Han Cai, Chuang Gan, Tianzhe Wang, Zhekai Zhang, and Song Han. Once-for-all: Train one network and specialize it for efficient deployment, 2019. [13] Neil C Thompson, Kristjan Greenewald, Keeheon Lee, and Gabriel F Manso. The computational limits of deep learning. ar Xiv preprint ar Xiv:2007.05558, 2020. [14] David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. ar Xiv preprint ar Xiv:1609.09106, 2016. [15] Alex Krizhevsky et al. Learning multiple layers of features from tiny images. 2009. [16] Timothy Hospedales, Antreas Antoniou, Paul Micaelli, and Amos Storkey. Meta-learning in neural networks: A survey. ar Xiv preprint ar Xiv:2004.05439, 2020. [17] Marcin Andrychowicz, Misha Denil, Sergio Gomez, Matthew W Hoffman, David Pfau, Tom Schaul, Brendan Shillingford, and Nando De Freitas. Learning to learn by gradient descent by gradient descent. In Advances in neural information processing systems, pages 3981 3989, 2016. [18] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. 2016. [19] Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. ar Xiv preprint ar Xiv:1806.09055, 2018. [20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. ar Xiv preprint ar Xiv:2010.11929, 2020. [21] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. Open Review.net, 2017. URL https://openreview.net/forum?id= SJU4ay Ygl. [22] Petar Veliˇckovi c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. Graph attention networks. In International Conference on Learning Representations, 2018. URL https://openreview.net/forum?id=r JXMpik CZ. [23] Vijay Prakash Dwivedi, Chaitanya K Joshi, Thomas Laurent, Yoshua Bengio, and Xavier Bresson. Benchmarking graph neural networks. ar Xiv preprint ar Xiv:2003.00982, 2020. [24] Chris Zhang, Mengye Ren, and Raquel Urtasun. Graph hypernetworks for neural architecture search. ar Xiv preprint ar Xiv:1810.05749, 2018. [25] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. ar Xiv preprint ar Xiv:1409.1556, 2014. [26] Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1492 1500, 2017. [27] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700 4708, 2017. [28] Barret Zoph and Quoc V Le. Neural architecture search with reinforcement learning. ar Xiv preprint ar Xiv:1611.01578, 2016. [29] Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V Le. Learning transferable architectures for scalable image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8697 8710, 2018. [30] Chenxi Liu, Barret Zoph, Maxim Neumann, Jonathon Shlens, Wei Hua, Li-Jia Li, Li Fei-Fei, Alan Yuille, Jonathan Huang, and Kevin Murphy. Progressive neural architecture search. In Proceedings of the European conference on computer vision (ECCV), pages 19 34, 2018. [31] Esteban Real, Alok Aggarwal, Yanping Huang, and Quoc V Le. Regularized evolution for image classifier architecture search. In Proceedings of the aaai conference on artificial intelligence, volume 33, pages 4780 4789, 2019. [32] Xin Chen, Lingxi Xie, Jun Wu, and Qi Tian. Progressive darts: Bridging the optimization gap for nas in the wild. ar Xiv preprint ar Xiv:1912.10952, 2019. [33] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1314 1324, 2019. [34] Yujia Li, Daniel Tarlow, Marc Brockschmidt, and Richard Zemel. Gated graph sequence neural networks. ar Xiv preprint ar Xiv:1511.05493, 2015. [35] Kyunghyun Cho, Bart Van Merriënboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. ar Xiv preprint ar Xiv:1406.1078, 2014. [36] Andrew Brock, Soham De, and Samuel L Smith. Characterizing signal propagation to close the performance gap in unnormalized resnets. ar Xiv preprint ar Xiv:2101.08692, 2021. [37] Andrew Brock, Soham De, Samuel L Smith, and Karen Simonyan. High-performance large-scale image recognition without normalization. ar Xiv preprint ar Xiv:2102.06171, 2021. [38] Laurent Sifre and Stéphane Mallat. Rigid-motion scattering for texture classification. ar Xiv preprint ar Xiv:1403.1687, 2014. [39] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. ar Xiv preprint ar Xiv:1704.04861, 2017. [40] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132 7141, 2018. [41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. ar Xiv preprint ar Xiv:1706.03762, 2017. [42] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. ar Xiv preprint ar Xiv:1607.06450, 2016. [43] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83 97, 1955. [44] Aric Hagberg, Pieter Swart, and Daniel S Chult. Exploring network structure, dynamics, and function using networkx. Technical report, Los Alamos National Lab.(LANL), Los Alamos, NM (United States), 2008. [45] Alain Barrat, Marc Barthelemy, Romualdo Pastor-Satorras, and Alessandro Vespignani. The architecture of complex weighted networks. Proceedings of the national academy of sciences, 101(11):3747 3752, 2004. [46] Jiaxuan You, Jure Leskovec, Kaiming He, and Saining Xie. Graph structure of neural networks, 2020. [47] Anna Golubeva, Behnam Neyshabur, and Guy Gur-Ari. Are wider nets better given the same number of parameters? ar Xiv preprint ar Xiv:2010.14495, 2020. [48] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. ar Xiv preprint ar Xiv:1605.07146, 2016. [49] Quynh Nguyen and Matthias Hein. The loss surface of deep and wide neural networks. In International conference on machine learning, pages 2603 2612. PMLR, 2017. [50] Rupesh Kumar Srivastava, Klaus Greff, and Jürgen Schmidhuber. Training very deep networks. ar Xiv preprint ar Xiv:1507.06228, 2015. [51] Sara Hooker. The hardware lottery, 2020. [52] Angus Galloway, Anna Golubeva, Thomas Tanay, Medhat Moussa, and Graham W Taylor. Batch normalization is a cause of adversarial vulnerability. ar Xiv preprint ar Xiv:1905.02161, 2019. [53] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. ar Xiv preprint ar Xiv:1903.12261, 2019. [54] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European conference on computer vision (ECCV), pages 3 19, 2018. [55] Siyuan Qiao, Huiyu Wang, Chenxi Liu, Wei Shen, and Alan Yuille. Micro-batch training with batchchannel normalization and weight standardization. ar Xiv preprint ar Xiv:1903.10520, 2019. [56] Hongyi Zhang, Yann N Dauphin, and Tengyu Ma. Fixup initialization: Residual learning without normalization. ar Xiv preprint ar Xiv:1901.09321, 2019. [57] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026 1034, 2015. [58] Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 249 256, 2010. [59] Oscar Chang, Lampros Flokas, and Hod Lipson. Principled weight initialization for hypernetworks. In International Conference on Learning Representations, 2019. [60] Uri Alon and Eran Yahav. On the bottleneck of graph neural networks and its practical implications. ar Xiv preprint ar Xiv:2006.05205, 2020. [61] Salah El Hihi and Yoshua Bengio. Hierarchical recurrent neural networks for long-term dependencies. In Advances in neural information processing systems, pages 493 499, 1996. [62] Meng Liu, Zhengyang Wang, and Shuiwang Ji. Non-local graph neural networks. ar Xiv preprint ar Xiv:2005.14612, 2020. [63] Hongbin Pei, Bingzhe Wei, Kevin Chen-Chuan Chang, Yu Lei, and Bo Yang. Geom-gcn: Geometric graph convolutional networks. ar Xiv preprint ar Xiv:2002.05287, 2020. [64] Jiaxuan You, Rex Ying, and Jure Leskovec. Position-aware graph neural networks. In International Conference on Machine Learning, pages 7134 7143. PMLR, 2019. [65] Yiding Yang, Xinchao Wang, Mingli Song, Junsong Yuan, and Dacheng Tao. Spagan: Shortest path graph attention network. ar Xiv preprint ar Xiv:2101.03464, 2021. [66] Pavlo M Radiuk. Impact of training set batch size on the performance of convolutional neural networks for diverse datasets. Information Technology and Management Science, 20(1):20 24, 2017. [67] Zhaohui Yang, Yunhe Wang, Xinghao Chen, Boxin Shi, Chao Xu, Chunjing Xu, Qi Tian, and Chang Xu. Cars: Continuous evolution for efficient neural architecture search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1829 1838, 2020. [68] Chaoyang He, Haishan Ye, Li Shen, and Tong Zhang. Milenas: Efficient neural architecture search via mixed-level reformulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11993 12002, 2020. [69] Guohao Li, Guocheng Qian, Itzel C Delgadillo, Matthias Muller, Ali Thabet, and Bernard Ghanem. Sgas: Sequential greedy architecture search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1620 1630, 2020. [70] Olga Wichrowska, Niru Maheswaranathan, Matthew W Hoffman, Sergio Gomez Colmenarejo, Misha Denil, Nando Freitas, and Jascha Sohl-Dickstein. Learned optimizers that scale and generalize. In International Conference on Machine Learning, pages 3751 3760. PMLR, 2017. [71] Luke Metz, Niru Maheswaranathan, C Daniel Freeman, Ben Poole, and Jascha Sohl-Dickstein. Tasks, stability, architecture, and compute: Training more effective learned optimizers, and using them to train themselves. ar Xiv preprint ar Xiv:2009.11243, 2020. [72] Jingzhao Zhang, Sai Praneeth Karimireddy, Andreas Veit, Seungyeon Kim, Sashank J Reddi, Sanjiv Kumar, and Suvrit Sra. Why adam beats sgd for attention models. ar Xiv e-prints, pages ar Xiv 1912, 2019. [73] Priya Goyal, Piotr Dollár, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: Training imagenet in 1 hour. ar Xiv preprint ar Xiv:1706.02677, 2017. [74] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pages 1097 1105, 2012. [75] Wei Li, Shaogang Gong, and Xiatian Zhu. Neural graph embedding for neural architecture search. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 4707 4714, 2020. [76] Wei Wen, Hanxiao Liu, Hai Li, Yiran Chen, Gabriel Bender, and Pieter-Jan Kindermans. Neural predictor for neural architecture search. ar Xiv preprint ar Xiv:1912.00848, 2019. [77] Haifeng Jin, Qingquan Song, and Xia Hu. Auto-keras: An efficient neural architecture search system. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 1946 1956, 2019. [78] Nils M Kriege, Fredrik D Johansson, and Christopher Morris. A survey on graph kernels. Applied Network Science, 5(1):1 42, 2020. [79] Ilya Makarov, Dmitrii Kiselev, Nikita Nikitinsky, and Lovro Subelj. Survey on graph embeddings and their applications to machine learning problems on graphs. Peer J Computer Science, 7, 2021. [80] Wei Wen, Hanxiao Liu, Yiran Chen, Hai Li, Gabriel Bender, and Pieter-Jan Kindermans. Neural predictor for neural architecture search. In European Conference on Computer Vision, pages 660 676. Springer, 2020. [81] Jovita Lukasik, David Friede, Heiner Stuckenschmidt, and Margret Keuper. Neural architecture performance prediction using graph neural networks. ar Xiv preprint ar Xiv:2010.10024, 2020. [82] Bowen Baker, Otkrist Gupta, Ramesh Raskar, and Nikhil Naik. Accelerating neural architecture search using performance prediction. ar Xiv preprint ar Xiv:1705.10823, 2017. [83] Simon Kornblith, Jonathon Shlens, and Quoc V Le. Do better imagenet models transfer better? In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2661 2671, 2019. [84] Minyoung Huh, Pulkit Agrawal, and Alexei A Efros. What makes imagenet good for transfer learning? ar Xiv preprint ar Xiv:1608.08614, 2016. [85] Behnam Neyshabur, Hanie Sedghi, and Chiyuan Zhang. What is being transferred in transfer learning? ar Xiv preprint ar Xiv:2008.11687, 2020. [86] Maithra Raghu, Chiyuan Zhang, Jon Kleinberg, and Samy Bengio. Transfusion: Understanding transfer learning for medical imaging. ar Xiv preprint ar Xiv:1902.07208, 2019. [87] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. ar Xiv preprint ar Xiv:1910.04867, 2019. [88] Py Torch. Pytorch object detection finetuning tutorial. URL https://pytorch.org/tutorials/ intermediate/torchvision_tutorial.html. [89] Liming Wang, Jianbo Shi, Gang Song, and I-fan Shen. Object detection combining recognition and segmentation. In Asian conference on computer vision, pages 189 199. Springer, 2007. [90] Jürgen Schmidhuber and AI Blog. Metalearning machines learn to learn (1987-). [91] Louis Kirsch and Jürgen Schmidhuber. Meta learning backpropagation and improving it. ar Xiv preprint ar Xiv:2012.14905, 2020. [92] Hugo Siqueira Gomes, Benjamin Léger, and Christian Gagné. Meta learning black-box population-based optimizers. ar Xiv preprint ar Xiv:2103.03526, 2021. [93] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 1126 1135. JMLR. org, 2017. [94] Jake Snell, Kevin Swersky, and Richard S Zemel. Prototypical networks for few-shot learning. ar Xiv preprint ar Xiv:1703.05175, 2017. [95] Adriana Romero, Pierre Luc Carrier, Akram Erraqabi, Tristan Sylvain, Alex Auvolat, Etienne Dejoie, Marc-André Legault, Marie-Pierre Dubé, Julie G Hussin, and Yoshua Bengio. Diet networks: thin parameters for fat genomics. ar Xiv preprint ar Xiv:1611.09340, 2016. [96] James Requeima, Jonathan Gordon, John Bronskill, Sebastian Nowozin, and Richard E Turner. Fast and flexible multi-task classification using conditional neural adaptive processes. ar Xiv preprint ar Xiv:1906.07697, 2019. [97] Huaiyu Li, Weiming Dong, Xing Mei, Chongyang Ma, Feiyue Huang, and Bao-Gang Hu. Lgm-net: Learning to generate matching networks for few-shot learning. In International conference on machine learning, pages 3825 3834. PMLR, 2019. [98] Luca Bertinetto, João F Henriques, Jack Valmadre, Philip HS Torr, and Andrea Vedaldi. Learning feed-forward one-shot learners. ar Xiv preprint ar Xiv:1606.05233, 2016. [99] Dongze Lian, Yin Zheng, Yintao Xu, Yanxiong Lu, Leyu Lin, Peilin Zhao, Junzhou Huang, and Shenghua Gao. Towards fast adaptation of neural architectures with meta learning. In ICLR. JMLR. org, 2020. [100] Thomas Elsken, Benedikt Staffler, Jan Hendrik Metzen, and Frank Hutter. Meta-learning of neural architectures for few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12365 12375, 2020. [101] Xin Chen, Yawen Duan, Zewei Chen, Hang Xu, Zihao Chen, Xiaodan Liang, Tong Zhang, and Zhenguo Li. Catch: Context-based meta reinforcement learning for transferrable architecture search. In European Conference on Computer Vision, pages 185 202. Springer, 2020. [102] Jiahui Yu, Pengchong Jin, Hanxiao Liu, Gabriel Bender, Pieter-Jan Kindermans, Mingxing Tan, Thomas Huang, Xiaodan Song, Ruoming Pang, and Quoc Le. Bignas: Scaling up neural architecture search with big single-stage models. ar Xiv preprint ar Xiv:2003.11142, 2020. [103] Xin He, Kaiyong Zhao, and Xiaowen Chu. Automl: A survey of the state-of-the-art. Knowledge-Based Systems, 212:106622, 2021. [104] Andrew Brock, Theodore Lim, James M Ritchie, and Nick Weston. Smash: one-shot model architecture search through hypernetworks. ar Xiv preprint ar Xiv:1708.05344, 2017. [105] Gabriel Bender, Pieter-Jan Kindermans, Barret Zoph, Vijay Vasudevan, and Quoc Le. Understanding and simplifying one-shot architecture search. In International Conference on Machine Learning, pages 550 559, 2018. [106] Guixiang Ma, Nesreen K Ahmed, Theodore L Willke, and S Yu Philip. Deep graph similarity learning: A survey. Data Mining and Knowledge Discovery, pages 1 38, 2021. [107] Yunsheng Bai, Hao Ding, Song Bian, Ting Chen, Yizhou Sun, and Wei Wang. Simgnn: A neural network approach to fast graph similarity computation. In Proceedings of the Twelfth ACM International Conference on Web Search and Data Mining, pages 384 392, 2019. [108] DC Dowson and BV Landau. The fréchet distance between multivariate normal distributions. Journal of multivariate analysis, 12(3):450 455, 1982. [109] Chia-Cheng Liu, Harris Chan, Kevin Luk, and AI Borealis. Auto-regressive graph generation modeling with improved evaluation methods. In 33rd Conference on Neural Information Processing Systems. Vancouver, Canada, 2019. [110] Julian Zilly, Hannes Zilly, Oliver Richter, Roger Wattenhofer, Andrea Censi, and Emilio Frazzoli. The frechet distance of training and test distribution predicts the generalization gap. 2019. [111] Rylee Thompson, Elahe Ghalebi, Terrance De Vries, and Graham W Taylor. Building lego using deep generative models of graphs. ar Xiv preprint ar Xiv:2012.11543, 2020. [112] Pinar Yanardag and SVN Vishwanathan. Deep graph kernels. In Proceedings of the 21th ACM SIGKDD international conference on knowledge discovery and data mining, pages 1365 1374, 2015. [113] Misha Denil, Babak Shakibi, Laurent Dinh, Marc Aurelio Ranzato, and Nando de Freitas. Predicting parameters in deep learning. In C J C Burges, L Bottou, M Welling, Z Ghahramani, and K Q Weinberger, editors, Advances in Neural Information Processing Systems 26, pages 2148 2156. Curran Associates, Inc., 2013. [114] Neale Ratzlaff and Li Fuxin. Hypergan: A generative model for diverse, performant neural networks. In International Conference on Machine Learning, pages 5361 5369. PMLR, 2019. [115] Iou-Jen Liu, Jian Peng, and Alexander G Schwing. Knowledge flow: Improve upon your teachers. ar Xiv preprint ar Xiv:1904.05878, 2019. [116] Yu Cheng, Duo Wang, Pan Zhou, and Tao Zhang. A survey of model compression and acceleration for deep neural networks. ar Xiv preprint ar Xiv:1710.09282, 2017. [117] Yann Dauphin and Samuel S Schoenholz. Metainit: Initializing learning by learning to initialize. 2019. [118] Chen Zhu, Renkun Ni, Zheng Xu, Kezhi Kong, W Ronny Huang, and Tom Goldstein. Gradinit: Learning to initialize neural networks for stable and efficient training. ar Xiv preprint ar Xiv:2102.08098, 2021. [119] Debasmit Das, Yash Bhalgat, and Fatih Porikli. Data-driven weight initialization with sylvester solvers. ar Xiv preprint ar Xiv:2105.10335, 2021. [120] Yue Yu, Jie Chen, Tian Gao, and Mo Yu. Dag-gnn: Dag structure learning with graph neural networks. In International Conference on Machine Learning, pages 7154 7163. PMLR, 2019. [121] Xiaojie Guo and Liang Zhao. A systematic survey on deep generative models for graph generation. ar Xiv preprint ar Xiv:2007.06686, 2020. [122] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dollár. Designing network design spaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10428 10436, 2020. [123] Jiaxuan You, Zhitao Ying, and Jure Leskovec. Design space for graph neural networks. Advances in Neural Information Processing Systems, 33, 2020.