当前位置:网站首页>Say goodbye to the AI era of hand looms
Say goodbye to the AI era of hand looms
2022-08-09 13:09:00 【QbitAl】
詹士 编辑 发自 凹非寺
量子位 | 公众号 QbitAI
18世纪60年代,When the steam engine was invented,The vast majority of people don't realize“工业革命”的到来;
1946年ENIAC问世时,People don't realize,Computer in half a century later,Become a social operation support、Innovation of science and technology infrastructure;
即便是2005年,People also can't imagine without cash,Only use a mobile phone can meet the needs of food and clothing live line almost all;Without the need for a separate camera,Can be anywhere at any time record of my life.
When the birth of disruptive technology,People are always underestimate it brings to the social and economic impact,It is only in the dozens of years of history in order,The value of it can be defined.
就像比尔・盖茨所说:We always overestimate how can do it in a year,And underestimate in five or ten years can do.
如今,It's the protagonist's turn“人工智能”.
2022年,AIIndustrial chain preliminary perfect,Industry demand constantly emerging,Practitioners hold high「AI进入千行百业」Banner to expand the market,Some pioneer is successfulIPO.
降本增效、数字化转型、数字经济……have become the presentAIIndustrialization of related hot words.Lead the fourth industrial revolution and productivity iteration,AIThe value is multilateral consensus.
但10年后、50年后的AI图景,No one can clearly describe.AIDevelopment will be as we predicted and define today——
In our current view,Give out accurate answer.
面对AI的星辰大海,We just leave the surface of the earth.
In industrial maelstrom of change,It might be hard to seeAI的终局,但AI算法的“超大规模”和“精细化”,At least is the path to the final.
Very large scale and refinement of trend,AIThe Industrial Revolution is urgently needed
The so-called hyperscale,即AI算法无处不在——
The number of algorithms will be likeAPP数量一样,呈爆炸式增长,deep into production、工作、Life of every detail,Become city management、An important asset for the enterprise development.
As millions in the smartphone app store todayAPP种类,Algorithm almost covers all kinds of people、The diverse needs of all walks of life.其中有很多,Have become a part of our daily life.
so-called refinement,即单一AIThe function of the algorithm will be more and more breakdown,The innumerable tinyAI算法,According to different industries、不同场景、The use of different equipment demand,Combined into the complexity of highAI应用.
在这样的趋势下,很快,Each one we used toAI应用,Hundreds of thousands of kinds of algorithms are likely to be made of complex agent.
Refinement may causeAI的场景化,换言之,Is needed according to own actual situation and customer demand characteristics,Targeted training is suitable for the different vertical and segmentation scenarioAI算法.
However, as the scene constantly explore,Number of customized requirements from the customer will be rapid surge,这会给AIBe born to bring more challenges,For every need special custom not only increase the cost,Will also slow down the landing cycle.
Only see a community management domain,Algorithm needs include garbage overflow、高空抛物、口罩检测、Vehicle parking violation identification、The electric car enters the elevator、Residents fell、Such as large elevator tiring finely demand.制造、Energy is similar.
And the same model in different scenarios,Its applicability is not the same.
A flame of fire smoke recognition as an example,on the community street,Some cigarette certainly don't need to report to the police,Put it in the workplace,Spark as big as a welding also need not remind,But at the gas station,Demand and become a spark can't let go.
Although these scenarios to algorithm needs a great deal of fragmentation of the long tail,But it's still intelligent community management indispensable in.
However, this kind of long tail scene has a distinguishing feature of,That is sample data scarce,Can be used for training the high quality of data set is luxury.
因此,在开发过程中,Most of the time field data required original training,And after the algorithm on-line continuous iterative.Only experienced algorithm engineer under the limited amount of data,Trained a algorithm accuracy is good.
而在“超大规模”和“精细化”趋势下,AITerminal deployment adaptation algorithm,更是AILanding a big hidden pain points again.
Algorithm to use effect is good,It is necessary to chip adaptation process.
The work unfolds,To target different chips,Write a different tool chain development kit,Also the terminal chip performance quantitative adjustment,As much as possible to improve the utilization rate of chip.
目前,Most of the marketAIEnterprise only do fitNVIDIA、Qualcomm and other mainstream brand and the research chip,If the user used chip is beyond the scope of adaptation,Requires additional cost at least2~3Months time individually adapted,即便如此,Chip utilization may still be less than10%,造成极大资源浪费.
在传统的开发模式下,From the definition of the business problem,The data acquisition and mark,The design of the algorithm model、调参、训练、调优,To the model of chip adaptation and performance evaluation——
The chain is not only complicated、周期长,And need a lot of manual work,The whole process often often take months.The uncertainty of algorithm efficiency,More would increase the cost of work force.
这种“must be manual,才能智能”的工匠精神,In the face of future demand for the huge amounts of,就会力不从心.
Row by hand,Can't leave the surface of the earth.大家期待AIBring the fourth industrial revolution,The liberation of more labour-intensive work,但AI自身,The but again become the labor-intensive industries.A lot of research and engineers rework,Who will liberate?
此外,In-depth industrial landing,Also need to be around in the game of standardization and customization choices.
此前,算法SDK、SaaSService prevails,Many companies hope will gradually product standardization,achieve scale development.结果却发现,AIThe deeper you go into the industry,碎片化、The more non-standard requirements,General solution to the problem of all rely on a single model is a business model doesn't work.
And to do customized solutions、Turnkey integration project,And get into high cost、利润低、The predicament of not making money,成为AICompanies are not willing to pick up the coolie live.
AIThe industrialization of the huge gap between supply and demand,The constraints of a business model,To be the reconstruction of the productivity and production relations.AI自身,Also need a industrial revolution.
AutoML,Say goodbye to the hand loomAI时代
其实,Leading players of all stripes have long been aware of the one million head,并开始着手解决.
Someone more resources and in-depth industry,Research and development of the new algorithm one by one,And bending down incoming strong binding with the industry,Early set out to build a complete plan education market,Off to do a lot of integration work.
D be center on others,Make big model device,Hope to be able to solve all problems.
And a new way,不但要做到AI开发的“降本增效”,To reduce barriers to form industry,——
用AI的方式解决AI需求,Its underlying technology fromAutoML,Mainly two words:高效.
技如其名,AutoMLRefers to the stages in machine learning to reduce manual work,把“Handmade by artisans”变成“Assembly line automation”.
From the model of super structure design to join、From the training to the model of streamline compression、And chip of adaptation and deployment……At different stages using automation solution,Let the machine to replace artificial to complete adjustable parameter、Data processing and so on heavy and complicated work.
The core idea is ready to useAI训练AI.
因AutoMLWill the original underlying framework、And the reshaping of the cooperation mode,Have insiders call it:人工智能2.0阶段的标志.
作为AutoML的提出和尝鲜者,Google has been quick to relevant layout,There are also some startup,also actively developAutoML创新,成为AICan assign all sectors of the industry practitioners.
革新AIProductivity and production relations
Reflected in the actual industry application,AutoML有多高效?
Shenzhen start-up company“共达地”The two product manager automation training platform based on the company,只用2~3周时间,Would have a quick training completed more than100个算法,涵盖了目标检测追踪、图像分类、语义分割、姿态检测、3D检测等五个大类视觉算法,覆盖80+个碎片化应用场景以及70+款AI芯片.
Simple conversion,Originally developed algorithms to the deployment needs at least half a year,In an automated way now,半天就能完成,The efficiency of exponential growth.
但将AutoMLFull belt to the commercial market,Also need to be from the perspective of customer value,Help customers at a lower cost to quickly get started,Create productivity innovation;同时,Linkage industry upstream and downstream,Optimize the industrial system of supply and demand,The reshaping of the relations between the production.
首先,Is the innovation of productivity.
虽说AutoML号称「自动」,但对非AITechnology product, product manager、Data analysts and other groups,仍属于「搞不明白」headache tool.
It is different from the tech giant in only internal technical personnel work efficiency,Totally under developmentPipelineOn the whole chain automation renovation,让不懂AIBusiness people can useAI,大幅降低AI的使用门槛,做企业背后的无限的AI生产力.
从图中可以看出,Amount to whole process implementation to the0Covenant-lite USESAutoMLTrain what you wantAI,Users only need a simple click on the button,就可以根据自身需求,Self-service upload the training data,Platform can independently complete model design、Training and tuning,In a short period of time can be trained a high qualityAI视觉算法.
目前,The platform has covered industry90%Common tasks above,categories covered:检测、分割、分类、人体、3Detc. Algorithms.Data acquisition marks can also be to amount to partner with,The product manager and business experts need to responsible for defining requirements,Can quickly complete the ground,实现“Definition is what you get”.
由于现实中,Many small and medium-sized enterprises have urgent need for fragmentation scene algorithm,The team also joint data vendors,Fast automation「算法商城」——
enable customers to0代码、Plug and play the way,快速将AIAlgorithm is applied to their business chain,实现智能化升级.
目前,The mall contains nearly a scene,适配70the remaining chips5000A variety of high-precision algorithm,For the use of direct customer.
第二,Is the reshaping of the production relations.
Empowerment through openness,将AI交付能力赋予广泛的生态合作伙伴,common practiceAI赋能百业.
目前,AIIndustry chain includes data vendors、芯片厂商、The role of infrastructure vendors several links such as,Total corporate positioning on open,因此,在通过AutoMLPlatform to reshape industry chain process,Keep the full openness of each link partners.
For chip manufacturers adaptation, for example.
Due to different chip platform is based on its chip hardware architecture features,To develop their own tool chain,在AIIn model generation and deployment link,Need to take into account the different chip platform hardware fitment and utilization.
共达地AI平台在SDKLevel will integrate different tool chain,Complete the adaptive transition model to the terminal chip,满足AutoMLTraining platform of the generated model can be a key to the issuance of the terminal equipment,并让AIAlgorithm model of full effectiveness.
基于AutoML的高效,Total finished almost all the pre adaptation of mainstream chip and box,Chip utilization rate can be up to50~60%,相较于10%Average utilization of industry,Greatly improved the work force effectiveness.
Not just quantitative changes,Industrial revolution of the singularity
From the past an algorithm need a group ofAISpecial development engineer for months,By now a at the line of business、Don't understand the algorithm development or code programming product manager,一键三连,就可以根据自身需求,Algorithm efficiently producing high quality.
AutoML带来的Not just quantitative changes,Industrial revolution of the singularity.
当前,Enterprise has been more and more circles began to applicationAutoML技术改变AI开发模式.Many enterprises is to benefit from total automation training platform and the algorithm of mall.
基于AutoML技术,Total to create a build with customers,The magnitude of the vertical for all walks of life and niche scenario-basedAI需求,The definition of common business problems,Helping our clients achieve algorithm efficient customization and distributed deployment,Quickly meet the needs of all kinds of custom,提高开发效率,Reduce the manpower and research and development cost,Win-win cooperation with the customer.
Such as the state's urban services technology company,平安智慧城市,Just let it go try to applyAutoMLTechnology for its production scene algorithm.
Through the total automation of training platform,Peace wisdom city developers need not code,Can detect parking、Covers loss of or damage to the test、烟火检测、Strong-arming detection algorithm model of fragmentation scene training,Can also be a key to end deployment.这一过程中,Fastest algorithm development all the way to the deployment time spent a few hours.
又如,In the application of intelligent traffic scene,The traffic administrative department of the city, capital of hunan province in the building of the related project,Thousand visual purge used the amount to theAutoMLAutomated training platform,定制了“Driving is not wearing a seatbelt recognition”、“Phone identification while driving”Relating to the safe driving a seriesAI视觉算法.
With little inputAI算法工程师的情况下,Two weeks is complete all kinds of complicated traffic situationsAI算法模型训练.
These cases were clearly demonstratedAutoML对于AIVery large scale and is necessary for the fine development of——
让AIComposed of several algorithm integrating applications from、Only a single task simple agent,Gradually evolved into consists of massive algorithm、Complex agent with comprehensive ability,To complete a variety of complex tasks.
如果将AICompare it to an airplane,那么初始的AIApplications like the Wright brothers made the first plane,结构简陋,只能飞行12秒.而未来的AI应用,Like millions parts passenger aircraft today,Every day can deliver people from the earth at one end to the other end of the destination.
As moral amount to the company“Common destination”,Success at the same time to help others,own success.
对此,In total, founder andCEOZhao Cong with a summary of fun:做AIDoes not have to be establishedAI团队.
对企业来说,Through the amount to automate training platform,Can let the fragmentation of the long tail of sceneAIThe algorithm is implemented quickly,After deployment can fast optimization iteration、持续升级,Thus gaining additional value.
0The development of code covenant-lite way,也将AIThe threshold of the talent to a minimum,Enable integrators、方案商、Distributors have fast onesAI能力,让AIThe development of the algorithm development is no longer a burden,But into improved competitiveness and efficiency of the weapon.
更进一步看,The mall automation training platform and the algorithm,Total to step into the industry without a line to do total package integration project、Without having to touch the final application,But become algorithm behind the enabler,用AIProductivity and productive relations change,创造新的商业模式——
通过降低AIApplication unit cost,Step by step to help upstream and downstream enterprises to carry outAI技术赋能,Everyone to form a long and deep cooperation,Relying on them into various industries,Achieve economies of scale,完成AIThe goal of empowering all industries.
降低单位成本,Use scale effect to describeAI未来
Look back at the beginning,People will underestimate the value of disruptive innovation technology brings,或许正是因为,The technology has yet to realize large-scale effect,The application does not reduce the cost of the industry would be acceptable to the degree of.
《Prediction Machines》一书中提到,A base product prices fell sharply,Will change the whole world.
蒸汽机的出现,Did not immediately ignite industrial revolution,But in the unit cost down,After large-scale application,Just opened up by machine instead of manual labor age.
The emergence of vacuum tube computer,Did not immediately human science and technology revolution,Until the emergence of very large scale integrated circuit,With gradual development of electronic design automation,really pushedPC走入千家万户.
AI时代,The scene again again.
在AI规模化落地,And to assign to all sectors for the target under the trend of,挑战即是机遇.If not completely change the high cost of、Low efficiency of traditional development path,lack of economy,Can seriously hinder the development of artificial intelligence.
But with the efficiency of disruptive innovation,实现AIBatch version of the algorithm、大规模生产,And quickly into the industry of capillary,Help customer to exponential function improve,To build the core assets of digital transformation,才可能让AI走进各行各业.
No time to pursue bright,But in the feet on the ground to provide industry new ideas.With technology innovation waveAI的规模效应,或许这,Is total to believeAI未来.
边栏推荐
- 00后写个暑假作业,被监控成这笔样
- The latest interview summary in 20022 brought by Ali senior engineer is too fragrant
- 微信支付开发流程
- 放下手机吧:实验表明花20分钟思考和上网冲浪同样快乐
- 一甲子,正青春,CCF创建六十周年庆典在苏州举行
- 900页数学论文证明旋转的黑洞不会爆炸,丘成桐:30多年来广义相对论首次重大突破...
- ABAP 面试题:如何使用 ABAP 编程语言的 System CALL 接口,直接执行 ABAP 服务器所在操作系统的 shell 命令?
- 推荐一个免费50时长的AI算力平台
- Adalvo acquires its first branded product, Onsolis
- 专业人士使用的 11 种渗透测试工具
猜你喜欢
放下手机吧:实验表明花20分钟思考和上网冲浪同样快乐
How to upload local file trial version in binary mode in ABAP report
虚拟机安装出现的问题汇总
世界第4疯狂的科学家,在103岁生日那天去世了
h264协议
信息系统项目管理师必背核心考点(六十三)项目组合管理的主要过程&DIPP分析
Shell正则表达式,三剑客之grep命令
Reading and writing after separation, performance were up 100%
京东架构师呕心整理:jvm与性能调优有哪些核心技术知识点
鹅厂机器狗花式穿越10m梅花桩:前空翻、单桩跳、起身作揖...全程不打一个趔趄...
随机推荐
shell脚本------函数的格式,传参,变量,递归,数组
金融业“限薪令”出台/ 软银出售过半阿里持仓/ DeepMind新实验室成立... 今日更多新鲜事在此...
脱光衣服待着就能减肥,当真有这好事?
告别手摇织布机的AI时代
中科院打脸谷歌:普通电脑追上量子优越性,几小时搞定原本要一万年的计算...
标准C语言学习总结14
WPF implements a MessageBox message prompt box with a mask
Ways to prevent data fraud
#物联网征文#小熊派设备开发实战
两分钟录音就可秒变语言通!火山语音音色复刻技术如何修炼而成?
Shell之常用小工具(sort、uniq、tr、cut)
00后写个暑假作业,被监控成这笔样
Glory to the Blue Yonder, speeds up the strategic growth
我们真的需要DApp吗?App真的不能满足我们的幻想吗?
基于CAP组件实现补偿事务与幂等性保障
GRPC整体学习
MongoDB-查询中$all的用法介绍
正则表达式(规则,匹配,和实际使用)
非科班AI小哥火了:他没有ML学位,却拿到DeepMind的offer
Scala 高阶(七):集合内容汇总(上篇)