人工智能与运动科技的相关研究与应用.pptx

1、 1. Limitations of Artificial Intelligence2. Real-World AI Applications of Sport Technology3. Case 1: Visualization and Personalization of Classes4. Case 2: 3D Modeling5. Case 3: Mathematical Training Classes6. Case 4: Content Understanding7. Discussions 1. Everyone is crazy about AI.2. It is valuab

2、le as long as you cansee its pros and cons.3. AlphaGo is with complicatedthinking but simple actions.4. Open domains and closed domains. ( ) ) 1. A massive amount of entertainment without leaving the sofa(Tiktok, Iqiyi, Bilibili)2. Checking in with loved ones without really checking in(Wechat, Weibo

3、)3. Restaurant food delivery (Meituan, Ele.me)4. Online shopping (Taobao, JD)5. No more perusing the bookstore (Dangdang, Amazon)6. News at your fingertips (Toutiao)7. No more getting lost or asking for directions (Baidu, Gaode)8. Gaming in the virtual world (Arena of Valor)https:/ A. Tech for fitne

4、ssB. Tech for healthC. Smart hardware & GargetsKeep Moving! Class Visualization can solve:GodsView1. Class visualization2. Path of Training3. Clustering of classes4. Pattern of users5. User friendlyKnowledgeGraph ofClassesVisualizationRelation Problem 1: Give a training plan like a professionalfitne

5、ss coach.1. For each user group: suitRules, suitDayRules,workoutRules = workout candidate2. Initializing connection probability as equal3. Dynamically adjust the transition probabilitybased on user feedbacks4. Relatively stable system for each groupPersonalized Classes Problem 2: How can we provide

6、users a timely feedback in fitness training?rot 0rot 90Squatrot 90rot 0Walkrot 0Taijirot 90Squat use res50 1. Know better about yourself2. Monitoring your body change3. Avarta of fitness4. Model of Keep Ups5. Entertaining model Effectively measureYou body fat!1. Know your bodytype better.2. More acc

7、uratemeasurement and .If we are asked to identify the relationship between the given two pairs: and then the first relationship can be best attributed as has-type, whereas the second relationship can beattributed as instance-of.So, we can redefine the two pairs as. WALL-E _has_genre ? For the imagef

8、eature extraction,existing works usehand-craft features, off-the-shelf features, orjointly-trained featureswith the model. Deepfeatures have showedeffectiveness inResNet152 layers2015LBP2002SIFT2004HOG2011AlexNet8 layers2012representing thesemantic information ofimages.VGG19 layers2014Network-In-Net

9、work2014GoogleNet22 layers2014- & & Consider the semantic topic or47keyword information. 91950 Ignore the correlations betweentopics or keywords. Similar semantics can be expressed by synonyms or relevant topics.Bag-of-FeaturesBoWTF-IDFTopic ModelLDA Consider the textural sequentialinformation Ignor

10、e the situations that similarsemantics can be expressed indifferent ways with varioussequences.RNN (LSTM/GRU)Word Embedding The structure of theproposed model is adual-path neuralnetwork: i.e., textGraph ConvolutionalNetwork (text GCN)(top) and imageNeural Network(image NN) (bottom). (a) The origina

11、l text andthree kinds of textualrelationships: (b)distributed semanticrelationship in theembed- ding space, (c)word co-occurrencerelationship and (d)general knowledgerelationship defined by aknowledge graph.The semanticillustration of ourproposedframework basedon GCN and CNN.Some samples of text queryresults using four of our modelson the CMPlaces dataset.In the future work, wecan extend this model to othercross-modal areas like auto-matic image captioning and