Who invented convolutional neural networks? (people.idsia.ch)

Convolutional neural networks (CNNs) emerged incrementally rather than from a single inventor. Key technical milestones began with Kunihiko Fukushima: he introduced rectified linear units in 1969 and the Neocognitron (a multilayer architecture with convolution-like layers and spatial downsampling) in 1979. In the late 1980s, supervised training and modern convolution terminology appeared: Alex Waibel applied time-delay (1D) convolutions with backpropagation (1987), Homma et al. coined "convolution" for NNs, and Wei Zhang et al. produced the first backprop-trained 2D CNN for character recognition (1988–89). Yann LeCun’s 1989 Bell Labs paper followed shortly after. Pooling shifted from Fukushima’s spatial averaging to max-pooling in the early 1990s, and GPU-accelerated scaling (notably Dan Ciresan’s DanNet in 2011) delivered the practical breakthroughs that led to superhuman image recognition. The significance for AI/ML is twofold: foundational algorithmic ideas (weight sharing, local receptive fields, ReLUs, pooling) were established decades ago, and modern success largely reflects engineering—scaling those ideas with backprop, better optimizers, and massively cheaper compute (GPUs). The early concentration of work in Japan reflects 1980s research funding and robotics leadership. Credit is therefore diffuse: invention was a chain of theoretical advances and empirical engineering, with today’s breakthroughs depending critically on hardware and implementation at scale.