


Paper address: https://www.nature.com/articles/s41562-022-01394-8
However, before talking about this project, we need to Let me give an example first to make it easier for everyone to understand.
If I stand in front of you with a pen, and then I hide the pen behind my back, will you not be able to see the pen?
But the pen must still exist, right?
Such a simple truth, not only you understand it, but even a two-month-old baby understands it.
But the reason behind it is very intriguing. Scientists are curious, why do people naturally understand this principle?
The story of DeepMind starts from this simple curiosity.
Do you think babies don’t understand anything?
We call "the pen cannot be seen after being put down but the pen is still there" as one of the thousands of common sense in physics, and DeepMind scientists want to compare common sense in physics between AI and babies.
Luis Piloto of Princeton University and his colleagues developed a deep learning AI system that can understand some common-sense laws of the physical world.
In this way, future computer models can better imitate human thinking and solve problems with a model that has the same cognition as a baby.
In general, any AI model starts with a blank piece of paper, and then uses a variety of examples to train the model. From the input data and examples, the model generates knowledge.
However, scientists point out that this is not the case for babies.
Babies do not learn things from scratch, but they are born with some prejudgments about objective things.
Let’s take the hidden pen above as an example. Babies innately know that even if the pen is hidden, it will still be there.
This is the underlying logic of the next experiment. That is: babies have some core assumptions when they are born, and these assumptions will allow them to develop in the right direction as they grow up, and their knowledge will become better and better as time goes by and experience increases. Come more refined.
This gave the Piloto team inspiration.
Piloto wondered, would a deep learning artificial intelligence model that imitates infant behavior patterns perform better than an artificial intelligence model that starts with a blank sheet of paper and relies solely on experience learning?
The researchers further compared the two different models.
The first thing they did was the traditional method (referred to as a blank sheet of paper). They gave the AI model some visual animations of objects and let the AI learn, such as a block sliding down a slope, or a ball bouncing against a wall.
The AI model detected movement patterns in these animations, and the researchers then set out to test whether the model could predict the movement of other objects.
On the other hand, the artificial intelligence model that imitates babies has some "principles" at the beginning, and the source of these "principles" is some innate assumptions of babies about the movement and interaction between objects.
To give a simple example, babies know that two objects cannot pass through each other, an object cannot rise out of thin air, etc.
AI that imitates infant cognition——"Plato"
In fact, the common sense of physics that infants know innately goes beyond the two points mentioned above. The full version is the following five points:
1. Continuity: Objects do not travel from one place to another, but have a certain continuous path in time and space.
2. Object persistence: Objects will not disappear when out of sight.
3. Solidity: Objects do not penetrate each other.
4. Immutability: The properties of the object (such as shape) will not change.
5. Directional inertia: The path of object motion is consistent with the principle of inertia.
Based on these five points of knowledge, if you perform a magic trick or something for babies, and then something goes against their preset knowledge, they will know that you are doing it. They are clever and know that counter-intuitive phenomena are not the way things are.
Although, babies are still not as knowledgeable as older children. Babies will observe counter-common sense phenomena for a long time, and then compare them with their own preset cognitions, and finally conclude that someone is doing tricks. in conclusion.
Speaking of this, I can’t help but think of a very popular video. My parents hid behind the sheets, shook the sheets up and down a few times, and quickly hid in the room behind them while hiding behind the sheets. When the baby doesn't see his parents after the sheets disappear, he will stand there and think for a while, wondering where his parents have gone.
There is another interesting point here. That is, babies will express "surprise" after seeing counterintuitive phenomena. This sounds obvious, but the researchers also replicated this unique performance in AI.
With these foundations in mind, let’s look at the experimental results.
The AI model designed by Piloto is called PLATO (Physics Learning through Auto-encoding and Tracking Objects), which is "Plato".
PLATO was trained on almost 30 hours of videos showing how objects perform some simple movements, and then trained the model to predict the movements of these objects in different situations.
What’s interesting is that the model finally learned the five common sense points of physics mentioned above.
When there is something counter-intuitive in the video you are watching, PLATO can also show a certain degree of surprise like a baby.
Piloto and his colleagues found that the AI model using the traditional training method (a blank piece of paper) performed well, but it was not as good as the unknown and was shocked by the comparison. PLATO, an AI model that imitates babies, performed much better.
Because of the blessing of preset cognition, the latter model can more accurately predict the movement of an object, apply the preset cognition to new object movement animations, and train the model The size of the data set used will also be smaller.
The Piloto team concluded that although learning and experience accumulation are important, it is not everything.
Their research points directly to a classic question - what is innate in humans and what is learned.
The next step is to apply this kind of human cognition to AI research.
Piloto has shown us the excellent results of the new method.
However, Piloto emphasized that PLATO is not designed to be a baby behavior model. We are just borrowing some methods of baby cognition to feed artificial intelligence.
PLATO’s simulation system: feedforward perception module (left) and loop dynamic predictor module (right)
Jeff, a computer scientist at the University of British Columbia in Vancouver Clune also said that combining AI with the learning methods of human infants is an important direction.
At this moment, Clune is working with other researchers to develop their own algorithmic methods for understanding the physical world.
Author introduction
Luis Piloto is the first author of the paper and the corresponding author.
He received a bachelor's degree in computer science from Rutgers University in 2012, and then went to Princeton University to study, and received master's degrees in neuroscience in 2017 and 2021 respectively. Ph.D.
In 2016, he officially joined DeepMind and became a research scientist.
The above is the detailed content of Make AI think like a baby! DeepMind 'Plato” model published in Nature sub-journal. For more information, please follow other related articles on the PHP Chinese website!

1 前言在发布DALL·E的15个月后,OpenAI在今年春天带了续作DALL·E 2,以其更加惊艳的效果和丰富的可玩性迅速占领了各大AI社区的头条。近年来,随着生成对抗网络(GAN)、变分自编码器(VAE)、扩散模型(Diffusion models)的出现,深度学习已向世人展现其强大的图像生成能力;加上GPT-3、BERT等NLP模型的成功,人类正逐步打破文本和图像的信息界限。在DALL·E 2中,只需输入简单的文本(prompt),它就可以生成多张1024*1024的高清图像。这些图像甚至

Wav2vec 2.0 [1],HuBERT [2] 和 WavLM [3] 等语音预训练模型,通过在多达上万小时的无标注语音数据(如 Libri-light )上的自监督学习,显著提升了自动语音识别(Automatic Speech Recognition, ASR),语音合成(Text-to-speech, TTS)和语音转换(Voice Conversation,VC)等语音下游任务的性能。然而这些模型都没有公开的中文版本,不便于应用在中文语音研究场景。 WenetSpeech [4] 是

“Making large models smaller”这是很多语言模型研究人员的学术追求,针对大模型昂贵的环境和训练成本,陈丹琦在智源大会青源学术年会上做了题为“Making large models smaller”的特邀报告。报告中重点提及了基于记忆增强的TRIME算法和基于粗细粒度联合剪枝和逐层蒸馏的CofiPruning算法。前者能够在不改变模型结构的基础上兼顾语言模型困惑度和检索速度方面的优势;而后者可以在保证下游任务准确度的同时实现更快的处理速度,具有更小的模型结构。陈丹琦 普

由于复杂的注意力机制和模型设计,大多数现有的视觉 Transformer(ViT)在现实的工业部署场景中不能像卷积神经网络(CNN)那样高效地执行。这就带来了一个问题:视觉神经网络能否像 CNN 一样快速推断并像 ViT 一样强大?近期一些工作试图设计 CNN-Transformer 混合架构来解决这个问题,但这些工作的整体性能远不能令人满意。基于此,来自字节跳动的研究者提出了一种能在现实工业场景中有效部署的下一代视觉 Transformer——Next-ViT。从延迟 / 准确性权衡的角度看,

3月27号,Stability AI的创始人兼首席执行官Emad Mostaque在一条推文中宣布,Stable Diffusion XL 现已可用于公开测试。以下是一些事项:“XL”不是这个新的AI模型的官方名称。一旦发布稳定性AI公司的官方公告,名称将会更改。与先前版本相比,图像质量有所提高与先前版本相比,图像生成速度大大加快。示例图像让我们看看新旧AI模型在结果上的差异。Prompt: Luxury sports car with aerodynamic curves, shot in a

人工智能就是一个「拼财力」的行业,如果没有高性能计算设备,别说开发基础模型,就连微调模型都做不到。但如果只靠拼硬件,单靠当前计算性能的发展速度,迟早有一天无法满足日益膨胀的需求,所以还需要配套的软件来协调统筹计算能力,这时候就需要用到「智能计算」技术。最近,来自之江实验室、中国工程院、国防科技大学、浙江大学等多达十二个国内外研究机构共同发表了一篇论文,首次对智能计算领域进行了全面的调研,涵盖了理论基础、智能与计算的技术融合、重要应用、挑战和未来前景。论文链接:https://spj.scien

译者 | 李睿审校 | 孙淑娟近年来, Transformer 机器学习模型已经成为深度学习和深度神经网络技术进步的主要亮点之一。它主要用于自然语言处理中的高级应用。谷歌正在使用它来增强其搜索引擎结果。OpenAI 使用 Transformer 创建了著名的 GPT-2和 GPT-3模型。自从2017年首次亮相以来,Transformer 架构不断发展并扩展到多种不同的变体,从语言任务扩展到其他领域。它们已被用于时间序列预测。它们是 DeepMind 的蛋白质结构预测模型 AlphaFold

说起2010年南非世界杯的最大网红,一定非「章鱼保罗」莫属!这只位于德国海洋生物中心的神奇章鱼,不仅成功预测了德国队全部七场比赛的结果,还顺利地选出了最终的总冠军西班牙队。不幸的是,保罗已经永远地离开了我们,但它的「遗产」却在人们预测足球比赛结果的尝试中持续存在。在艾伦图灵研究所(The Alan Turing Institute),随着2022年卡塔尔世界杯的持续进行,三位研究员Nick Barlow、Jack Roberts和Ryan Chan决定用一种AI算法预测今年的冠军归属。预测模型图


Hot AI Tools

Undresser.AI Undress
AI-powered app for creating realistic nude photos

AI Clothes Remover
Online AI tool for removing clothes from photos.

Undress AI Tool
Undress images for free

Clothoff.io
AI clothes remover

AI Hentai Generator
Generate AI Hentai for free.

Hot Article

Hot Tools

SublimeText3 Linux new version
SublimeText3 Linux latest version

WebStorm Mac version
Useful JavaScript development tools

Dreamweaver CS6
Visual web development tools

SAP NetWeaver Server Adapter for Eclipse
Integrate Eclipse with SAP NetWeaver application server.

SublimeText3 Chinese version
Chinese version, very easy to use
