PAC-Bayesian bounds are known to be tight and informative when studying the\ngeneralization ability of randomized classifiers. However, they require a loose\nand costly derandomization step when applied to some families of deterministic\nmodels such as neural networks. As an alternative to this step, we introduce\nnew PAC-Bayesian generalization bounds that have the originality to provide\ndisintegrated bounds, i.e., they give guarantees over one single hypothesis\ninstead of the usual averaged analysis. Our bounds are easily optimizable and\ncan be used to design learning algorithms. We illustrate this behavior on\nneural networks, and we show a significant practical improvement over the\nstate-of-the-art framework.\n