Our brains are biological neural networks. With big data, big compute, and machine learning algorithms, we can create artificial neural networks that are incredibly close to the real deal.
我们的大脑是生物神经网络。 借助大数据，大计算和机器学习算法，我们可以创建非常接近真实交易的人工神经网络。
For instance, the revolutionary GPT-3 model can write articles that fool 88% of readers into thinking it’s human, it can write code, poetry, song lyrics, memes, and the list goes on.
例如，革命性的GPT-3模型可以撰写使88％的读者误以为是人类的文章，可以编写代码，诗歌，歌曲歌词，模因，而且清单还在继续。
However, we still haven’t reached the “lightswitch moment” that turns on superintelligence — nor is it in sight.
但是，我们仍然还没有达到开启超级智能的“光开关时刻”-也不在眼前。
神经科学可以改善AI (Neuroscience Can Improve AI)
This is the premise that BrainOS — a novel brain-alike AutoML framework — is based on. The reason is simple: Our brains are the most powerful processors known, so we should at least test the principles by which our brains (seemingly) operate for the operation of AI.
这是BrainOS (一种新颖的类似于大脑的AutoML框架)所基于的前提。 原因很简单：我们的大脑是已知的最强大的处理器，因此我们至少应该测试大脑(貌似)为AI运作所依据的原理。
As an aside on efficiency, let’s compare the brain and AI. After all, the brain is our benchmark for AI’s accuracy, so let’s take a cursory look at efficiency as well.
除了效率之外，让我们比较一下大脑和AI。 毕竟，大脑是我们AI准确性的基准，因此让我们也粗略地看一下效率。
The brain uses 300 calories /day. It’d take 355 years (3.1m hours) to train GPT-3 on a Tesla V100, which individually draws 300 W. Multiplying Watts/hour times the total number of training hours yields 933m W. This is equal to 802 billion calories.
大脑每天消耗300卡路里 。 在Tesla V100上训练GPT-3大约需要 355年(310万小时)，这可单独消耗300 W功率。乘以瓦特/小时乘以总训练小时数可得出9.33 亿 W。这等于8020 亿卡路里 。
NVIDIA estimates that 80–90% of the energy costs of AI lie in inference — using the model after training. Calculating just GPU costs, GPT-3 then costs up to 7.2 trillion calories.
NVIDIA 估计 ，AI的能源成本中有80–90％来自推理-在训练后使用该模型。 仅计算GPU成本，GPT-3的成本就高达7.2万亿卡路里。
Ouch.
哎哟。
BrainOS isn’t being used in production just yet (we’ll get into a proposal later), but there are current attempts out there to increase efficiency.
BrainOS尚未在生产中使用(我们将在以后提出提案)，但是目前有尝试提高效率。
For one, Spiking Neural Networks are a brain-like way to increase efficiency, but they don’t offer the same benefits as AutoML that BrainOS would bring.
一方面，Spiking Neural Networks是一种类似于大脑的提高效率的方法，但它们所提供的好处不如BrainOS带来的AutoML。
While BrainOS is still a research project, there are many functional AutoML tools out there, like Apteo. For more on AutoML, check out this guide:
虽然BrainOS仍是一个研究项目，但那里有许多功能齐全的AutoML工具，例如Apteo 。 有关AutoML的更多信息，请查看此指南：
BrainOS的工作方式 (How BrainOS Works)
BrainOS automatically selects the appropriate ML model to use via these attributes:
BrainOS通过以下属性自动选择合适的ML模型：
- The data it’s given 给出的数据
- Prior experience 以往的经验
- World knowledge 世界知识
More than just being loosely connected with our ideas of how the brain works, BrainOS seeks to model neuronal behavior.
BrainOS不仅与我们关于大脑工作原理的思想松散地联系在一起，还试图为神经元行为建模。
“The system’s architecture and operation is inspired by the behavior of neuronal cells.”
“系统的架构和操作受到神经细胞行为的启发。”
建筑 (Architecture)
The high-level architecture is pretty simple: Input data (from any source) is combined with the problem context and objective. Given this, either a new model will be created or an existing model selected for training, after which the model is deployed.
高层架构非常简单：将输入数据(来自任何来源)与问题上下文和目标结合在一起。 鉴于此，将创建新模型或选择现有模型进行训练，然后部署模型。
详细组件 (Detailed Components)
Problem formalization is the entry point to the system, comprising the “Input Data” box above. Next, the critic (or qualifier) component enhances the input data by adding antedate datasets and also achieves an intermediate by applying qualifications.
问题形式化是系统的入口点，包括上面的“输入数据”框。 接下来，批注者(或限定词)组件通过添加以前的数据集来增强输入数据，并且还可以通过应用限定词来达到中间效果。
The history database, inspired by the brain’s adaptive learning features, combines history (or experience in encountered data sets) and world knowledge (or stored knowledge as well as abstract research).
受大脑适应性学习功能启发的历史数据库将历史(或遇到的数据集的经验)和世界知识(或存储的知识以及抽象研究)结合在一起。
The planner component simply plans the execution sequence of algorithms, or the flow of the system. The parallel executor is the task scheduler that determines how to efficiently execute threads.
规划器组件仅规划算法的执行顺序或系统流程。 并行执行程序是确定如何有效执行线程的任务计划程序。
The module scheduler receives threads sent by the aforementioned parallel executor, planning a schedule for execution.
模块调度器接收由上述并行执行器发送的线程，以计划执行调度。
The selector is BrainOS’s key component, which picks out the right model by conducting many steps in parallel: Searching BrainOS’s history, searching the research dataset, building a tool from scratch, and performing an ensemble learning by combining several models. The best-fit model is then selected.
选择器是BrainOS的关键组件，它通过并行执行许多步骤来选择正确的模型：搜索BrainOS的历史记录，搜索研究数据集，从头开始构建工具，以及通过组合多个模型进行整体学习。 然后选择最适合的模型。
深度认知神经网络(DCNN)—实现BrainOS (Deep Cognitive Neural Network (DCNN) — Implementing BrainOS)
DCNNs are one proposal to implement BrainOS in the real world.
DCNN是在现实世界中实现BrainOS的一项建议。
DCNNs exhibit perception and reasoning, unlike typical NNs, and enable big data analysis in near real-time on small devices (e.g. smartphones). Further, they’re extremely energy efficient — up to 300x more than similar deep neural nets.
与典型的NN不同，DCNN具有感知能力和推理能力，并且可以在小型设备(例如智能手机)上近乎实时地进行大数据分析。 此外，它们具有极高的能源效率-比类似的深度神经网络高300倍。
There isn’t any DCNN-based AutoML on the market yet, but given the sky-high computational, data, and energy requirements of traditional deep nets, I’m sure we’ll see them soon.
市场上还没有基于DCNN的AutoML，但是鉴于传统深网的计算，数据和能源需求很高，我相信我们很快就会看到它们。
翻译自: https://towardsdatascience.com/brainos-the-most-brain-like-ai-61b334c7658b
扫一扫
3224
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
