3 accepted papers
"We present the latest generation of MobileNets: MobileNetV4 (MNv4). They feature universally-efficient architecture designs for mobile devices. We introduce the Universal Inverted Bottleneck (UIB) search block, a unified and flexible structure that merges Inverted Bottleneck (IB), ConvNext, Feed Fo…
Low-precision quantization is recognized for its efficacy in neural network optimization. Our analysis reveals that non-quantized elementwise operations which are prevalent in layers such as parameterized activation functions batch normalization and quantization scaling dominate the inference cost o…
Automatic speech generation algorithms, enhanced by deep learning techniques, enable an increasingly seamless and immediate machine-to-human interaction. As a result, the latest generation of phone-calling bots sounds more convincingly human than previous generations. The application of this technol…