The Transformer and its variants have been proven to be efficient sequence learners in many different domains. Despite their staggering success, a critical issue has been the enormous number of parameters that must be trained (ranging from 107 to 1011) along with the quadratic complexity of dot-product attention. In this work, we investigate the problem of approximating the two central components of the Transformer -- multi-head self-attention and point-wise feed-forward transformation, with reduced parameter space and computational complexity. We build upon recent developments in analyzing deep neural networks as numerical solvers of ordinary differential equations. Taking advantage of an analogy between Transformer stages and the evolution of a dynamical system of multiple interacting particles, we formulate a temporal evolution scheme, TransEvolve, to bypass costly dot-product attention over multiple stacked layers. We perform exhaustive experiments with TransEvolve on well-known encoder-decoder as well as encoder-only tasks. We observe that the degree of approximation (or inversely, the degree of parameter reduction) has different effects on the performance, depending on the task. While in the encoder-decoder regime, TransEvolvedelivers performances comparable to the original Transformer, in encoder-only tasks it consistently outperforms Transformer along with several subsequent variants.

Tree Search (TS) is crucial to some of the most influential successes in reinforcement learning. Here, we tackle two major challenges with TS that limit its usability: distribution shift and scalability. We first discover and analyze a counter-intuitive phenomenon: action selection through TS and a pre-trained value function often leads to lower performance compared to the original pre-trained agent, even when having access to the exact state and reward in future steps. We show this is due to a distribution shift to areas where value estimates are highly inaccurate and analyze this effect using Extreme Value theory. To overcome this problem, we introduce a novel off-policy correction term that accounts for the mismatch between the pre-trained value and its corresponding TS policy by penalizing under-sampled trajectories.

The detailed architecture for VanillaNet with 7-13 layers can be found in Table 1, where each convolutional layer is followed with an activation function. For the VanillaNet-13-1.5, the number of channels are multiplied with 1.5. For classification on ImageNet, we train the VanillaNets for 300 epochs utilizing the cosine learning rate decay [5]. The λis linearly decayed from 1 to 0 on epoch 0 and 100, respectively. The training details can be fould in Table 2.

At the heart of foundation models is the philosophy of "more is different", exemplified by the astonishing success in computer vision and natural language processing. However, the challenges of optimization and inherent complexity of transformer models call for a paradigm shift towards simplicity. In this study, we introduce VanillaNet, a neural network architecture that embraces elegance in design. By avoiding high depth, shortcuts, and intricate operations like selfattention, VanillaNet is refreshingly concise yet remarkably powerful. Each layer is carefully crafted to be compact and straightforward, with nonlinear activation functions pruned after training to restore the original architecture. VanillaNet overcomes the challenges of inherent complexity, making it ideal for resourceconstrained environments. Its easy-to-understand and highly simplified architecture opens new possibilities for efficient deployment. Extensive experimentation demonstrates that VanillaNet delivers performance on par with renowned deep neural networks and vision transformers, showcasing the power of minimalism in deep learning. This visionary journey of VanillaNet has significant potential to redefine the landscape and challenge the status quo of foundation model, setting a new path for elegant and effective model design.

Traditionally convolutional neural network architectures have been designed by stacking layers on top of each other to form deeper hierarchical networks. The cortex in the brain however does not just stack layers as done in standard convolution neural networks, instead different regions are organized next to each other in a large single sheet of neurons. Biological neurons self organize to form topographic maps, where neurons encoding similar stimuli group together to form logical clusters. Here we propose new self-organization principles that allow for the formation of hierarchical cortical regions (i.e.

Prior works on Post-training Quantization (PTQ) typically separate a neural network into sub-nets and quantize them sequentially. This process pays little attention to the dependency across the sub-nets, hence is less optimal. In this paper, we propose a novel Network-Wise Quantization (NWQ) approach to fully leveraging inter-layer dependency. NWQ faces a larger scale combinatorial optimization problem of discrete variables than in previous works, which raises two major challenges: over-fitting and discrete optimization problem. NWQ alleviates over fitting via a Activation Regularization (AR) technique, which better controls the activation distribution. To optimize discrete variables, NWQ introduces Annealing Softmax (ASoftmax) and Annealing Mixup (AMixup) to progressively transition quantized weights and activations from continuity to discretization, respectively. Extensive experiments demonstrates that NWQ outperforms prior state-of-the-art approaches by a large margin: 20.24% for the challenging configuration of MobileNetV2 with 2 bits on ImageNet, pushing extremely low-bit PTQ from feasibility to usability. In addition, NWQ is able to achieve competitive or better results with only 10% computation cost of previous works.

An image is worth 16x16 words: Transformers for image82 recognition at scale.