《Joone完全指南》序

I would like to present the objectives that I had in mind when I started to write the first lines of code for Joone.

在我开始书写Joone的第一行代码时我想要提出这个曾一度存在心中的目标。

My dream was (and still is) to create the necessary framework to enable the implementation of a new approach to the use of neural networks. I felt this necessity because the biggest (and unresolved until now) difficulty is to find the fittest network for a given problem, without falling into local minima, thus allowing one to discover the best neural network architecture for the problem.

我的梦想是（而且还是）创造能开辟神经网络应用新途径的合适的框架。我觉得这是必要的，因为最大的（和尚未解决的）困难是针对给定的问题寻找绝对最优网络，不是陷入局部最小值，而是允许对给定问题足以探索最好神经网络架构的一个框架。

Okay – you'll say – this is what we can do simply by training some randomly initialized neural network with a supervised or unsupervised algorithm. Yes, that is true, but this is just scholastic theory, because training only one neural network, especially for the hard problems found in real life situations, is rarely enough to permit the discovery of optimal solutions. In addition, finding the best neural network can be a daunting task simply because we need to determine numerous parameters for any given network. Parameters such as the number of the layers, number of neurons for each layer, the transfer function, the value of the learning rate and the momentum may all require extensive manipulation while searching for problem solutions. This difficulty often leads to many frustrating failures.

好吧-你会说-这是我们使用有监督或无监督算法简单地通过训练去随机地初始化一些神经网络能做到的事情。是的，这没错，但是这仅是学术理论，因为仅凭训练一个神经网络，根本不足以对最优方案的发现提供有力证明，尤其对于现实生活场景中的难题。此外，寻找最佳神经网络可以说是一项艰巨的任务，单就任意给定网络我们就需要确定数量可观的因素（变量）。参数，诸如层的数目，每个层的神经元的数目，传递方程，学习速率和动量的值可能都需要大量的操作来寻找问题的解决方案。这种困难往往会导致许多令人沮丧的失败。

That being said my basic idea is to provide an environment which will facilitate the training of many neural networks in parallel, initialised with different weights different parameters and different architectures, enabling investigators an opportunity to find the best neural network simply by selecting the fittest neural network after the training processes.

我所要讲的最基本想法是提供一个环境，这将有利于平行的众多神经网络的训练，以不同权重不同变量和不同架构初始化，使研究者通过训练过程选择最优神经网络找到最好的神经网络。

In addition these processes could continue retraining the selected neural networks until some final parameter is reached (e.g. a low RMSE value). Similar to distillation processes the best architecture would be distilled by Joone, not by the user!

此外这些过程可以持续再训练已选定的神经网络直到满足某些终极因素（如低RMSE值）。类似于蒸馏过程，最好的架构将被Joone蒸馏出来，而非由用户确定！

Many programs in existence today permit the selection of the fittest neural network by the application of genetic algorithms. I want to go beyond this. My goal is to build a flexible environment programmable by the end user, thereby permitting the implementation of any currently existing or newly discovered global optimisation algorithm. This is why Joone has its own distributed training environment and why it is based on a cloneable engine.

目前存在的许多程序允许应用遗传算法进行优胜劣汰神经网络的选择。我的目标是建立一个灵活的终端用户可编程环境，从而允许任何已有的或新建的全局优化算法的实施。这就是为什么Joone有自己的分布式训练环境和为什么它是基于可克隆的引擎。

Moreover, my dreams do not terminate with a flexible environment but extend to providing the ability for Joone end users to not only use but also distribute trained neural networks to others for their use. For example, I'm imagining an assurance company that continuously trains many neural networks on customer risk evaluations^[1] (perhaps using the results of historical cases), distributing the best ‘distilled' (or genetically evolved) neural network to its sales force, so that they can use optimized neural networks on their mobile devices.

再者，我的梦想不会止于一个灵活的环境，而是延伸到不仅为最终Joone用户提供使用能力而且还可以布署已训练好的神经网络为他人所用。例如，我可以想象一家保险公司，基于客户风险评估（或者使用历史案例存档）持续训练众多神经网络，把最好的”蒸馏”出来的（或遗传进化）神经网络发布给它的销售团队，以便他们可以在移动设备上使用优化过的神经网络。

This is why neural networks built with Joone are serializable, remotely transportable and easily runnable using simple, small and generalized programs using any wired or wireless protocol. It also means that my dream can become a more solid reality thanks to the advent of handheld devices like mobile phones and PDAs which contain Java virtual machines. Joone is ready to run on them.

这就是为什么Joone神经网络是可序列化的，是可远程传输的和使用简便随处运行的，是使用任何网络协议的小型和广义程序。这也意味着我的梦想得益于包含Java虚拟机的手持设备可以成为一个更确定的现实。Joone将运行在手持设备上。

I sincerely hope you will find our work interesting and useful and I thank you for giving Joone a try.

我真诚地希望你会发现我们的工作有趣和有用，感谢给Joone一个机会。

Paolo Marrone

and the Joone team

^[1]
The ethics (and the law in many countries) forbids one from making racial, sexual, religious (and others) discriminations. Consequently, a decisional system based on such personal characteristics ought not to be built.

声明: 除非转自他站（如有侵权，请联系处理）外，本文采用 BY-NC-SA 协议进行授权 | 智乐兔
转载请注明：转自《《Joone完全指南》序》
本文地址：https://www.zhiletu.com/archives-4814.html
关注公众号：