Closed threeal closed 4 years ago
Prof. Kai-Lung Hua, National Taiwan University of Science and Technology, Taiwan
## Noisy Data Problem ![image](https://user-images.githubusercontent.com/10202888/99341204-0b17bd80-28bc-11eb-9457-603fb9add561.png) ## TrustMAE ![image](https://user-images.githubusercontent.com/10202888/99341250-21257e00-28bc-11eb-9f73-7a0d77b24aa6.png) ## Memory Access Patterns ![image](https://user-images.githubusercontent.com/10202888/99341339-4914e180-28bc-11eb-8410-85c651ccc0e1.png) ## Sparse Memory Addressing ![image](https://user-images.githubusercontent.com/10202888/99341418-66e24680-28bc-11eb-9b52-ffb995a24269.png) ## Trust Region Memory Updates ![image](https://user-images.githubusercontent.com/10202888/99341597-c6405680-28bc-11eb-9def-42f1f1ef8119.png) ## Ablation ![image](https://user-images.githubusercontent.com/10202888/99341632-d5270900-28bc-11eb-8e65-2ac082791884.png) ## Detection Result ![image](https://user-images.githubusercontent.com/10202888/99341651-e112cb00-28bc-11eb-9dd8-5f1d1fb8a693.png) ## Different Noise Levels ![image](https://user-images.githubusercontent.com/10202888/99341688-f4be3180-28bc-11eb-8e71-a53bdb9c42df.png) ## Selection Results ![image](https://user-images.githubusercontent.com/10202888/99341747-0bfd1f00-28bd-11eb-98d4-532c44880602.png) ## Data Availability Problem ![image](https://user-images.githubusercontent.com/10202888/99341786-1f0fef00-28bd-11eb-8952-56efd764fee7.png) ![image](https://user-images.githubusercontent.com/10202888/99341812-2b944780-28bd-11eb-981e-a36883a57bd7.png) ## Incremental Learning ![image](https://user-images.githubusercontent.com/10202888/99341851-3ea71780-28bd-11eb-9af2-824757def8d0.png) ## ??? ![image](https://user-images.githubusercontent.com/10202888/99341870-4a92d980-28bd-11eb-8f6a-6f8486ef919c.png) ## ??? ![image](https://user-images.githubusercontent.com/10202888/99341962-71e9a680-28bd-11eb-8b42-8ad0530a252b.png) ## Our IncrementalGAN ![image](https://user-images.githubusercontent.com/10202888/99341997-8168ef80-28bd-11eb-8068-c3ac0194cab8.png) ## Size of Domain Embeding ![image](https://user-images.githubusercontent.com/10202888/99342038-99407380-28bd-11eb-9d70-5fd88d108692.png) ## Translator Network ![image](https://user-images.githubusercontent.com/10202888/99342100-b07f6100-28bd-11eb-8858-7955caed56c4.png) ## ??? ![image](https://user-images.githubusercontent.com/10202888/99342244-06540900-28be-11eb-982c-129ee3c0cf2c.png) ## User Study ![image](https://user-images.githubusercontent.com/10202888/99342257-110e9e00-28be-11eb-8148-a9519555a445.png)
## GPT-3, Most Excellent AI in Human History ![image](https://user-images.githubusercontent.com/10202888/99343570-bf1b4780-28c0-11eb-885e-3d686724eda6.png) ## Way of Learning ![image](https://user-images.githubusercontent.com/10202888/99343692-fb4ea800-28c0-11eb-90d9-b5af4be5718d.png) ## OpenAI GPT-3: Beginner Tutorials ![image](https://user-images.githubusercontent.com/10202888/99343735-10c3d200-28c1-11eb-85ef-7982cd563460.png) ## GPT-3 Can Do the Following ![image](https://user-images.githubusercontent.com/10202888/99343768-20dbb180-28c1-11eb-8d8b-a5f5a3b6671d.png) ## People worry about AI's Malfunctioning ![image](https://user-images.githubusercontent.com/10202888/99343875-5d0f1200-28c1-11eb-85a7-434cc0620e9d.png) - hegel general and secretary problem . - https://en.wikipedia.org/wiki/Georg_Wilhelm_Friedrich_Hegel maybe?? ## Future of AI ![image](https://user-images.githubusercontent.com/10202888/99343990-9fd0ea00-28c1-11eb-8ada-67ee0aa9668d.png) - AI will be our final inovation ## An Accident Caused by an Autonomous Vehicle ![image](https://user-images.githubusercontent.com/10202888/99344062-c7c04d80-28c1-11eb-8faf-a73bf88936c7.png) - In korea we don't allow AI to drive a car ??? - Do not create autonomous car because it's dangerous, but in the USA they still do it ## The Ethical Dilemma of Driverless Cars ![image](https://user-images.githubusercontent.com/10202888/99344230-24236d00-28c2-11eb-96a6-ab405a8f2fe8.png) ## Insurance and Marketing Dilemma of Driverless Vehicle ![image](https://user-images.githubusercontent.com/10202888/99344326-5f25a080-28c2-11eb-9250-abe97b78e53a.png) ## Adversarial Problem ![image](https://user-images.githubusercontent.com/10202888/99344450-b3308500-28c2-11eb-980d-db32d351c2eb.png) - Masalah dari CNN, sedikit penambahan pada noise dapat mengakibatkan hasil yang salah pada prediksi, walaupun secara tampilan gambar tidak tampak berbeda jauh. - Semua CNN punya masalah ini. ## Uni-image ![image](https://user-images.githubusercontent.com/10202888/99344687-3651db00-28c3-11eb-9fb8-ae639e7e0e5e.png) ## Pixel Redrawn For a Robust Adversarial Defense ![image](https://user-images.githubusercontent.com/10202888/99344716-4669ba80-28c3-11eb-8c38-b7663af09121.png) ## Conversation AI ![image](https://user-images.githubusercontent.com/10202888/99344761-67caa680-28c3-11eb-91c5-f254cd2050a7.png) ## Microsoft's Chatbot Tay Tweets ![image](https://user-images.githubusercontent.com/10202888/99345013-e293c180-28c3-11eb-8607-da924abeec37.png) ## Siri Call Police ![image](https://user-images.githubusercontent.com/10202888/99345036-f0e1dd80-28c3-11eb-8533-70b19ff484ab.png) - siri salah mengartikan "siri call stella" menjadi "siri call police" ## Reference ![image](https://user-images.githubusercontent.com/10202888/99345141-369ea600-28c4-11eb-861a-b7d37282fe91.png) - we have to worry about AI malfunctioning - people could said that 10 years, or some years later there could be AI that could be smarter than us - many people now trying to create standard for AI system, though it's hard ## AI Standardization Area ![image](https://user-images.githubusercontent.com/10202888/99345342-a3b23b80-28c4-11eb-8e8e-ce132a124ee3.png) ## Standardization Body ![image](https://user-images.githubusercontent.com/10202888/99345442-e6741380-28c4-11eb-8121-b41583cd7994.png) ## ??? ![image](https://user-images.githubusercontent.com/10202888/99345501-09062c80-28c5-11eb-9bac-79abe00e9a77.png) ## ??? ![image](https://user-images.githubusercontent.com/10202888/99345527-191e0c00-28c5-11eb-8682-fceb5e122027.png) ## ??? ![image](https://user-images.githubusercontent.com/10202888/99345555-25a26480-28c5-11eb-8013-5c79c414156e.png) ## OECD Background ![image](https://user-images.githubusercontent.com/10202888/99345634-4ff42200-28c5-11eb-85b7-3bf612ccda94.png) ## Asilomar AI Principles ![image](https://user-images.githubusercontent.com/10202888/99345675-65694c00-28c5-11eb-8ba0-3eba300acccc.png) ## ??? ![image](https://user-images.githubusercontent.com/10202888/99345690-6f8b4a80-28c5-11eb-84f3-11df4f363eff.png) ## Longer Term Issues ![image](https://user-images.githubusercontent.com/10202888/99345718-7d40d000-28c5-11eb-8a17-9904b165003b.png) ## ??? ![image](https://user-images.githubusercontent.com/10202888/99345765-9c3f6200-28c5-11eb-9f7d-4066257b3d3a.png) ## ??? ![image](https://user-images.githubusercontent.com/10202888/99345821-c2650200-28c5-11eb-9167-8ddfb94148fd.png) - kita enggak bisa mengatakan kalau CNN memiliki akurasi bagus, karena kita tidak tahu, tidak ada yang tahu apa yang akan terjadi dengan sistem seiring berjalannya waktu ## ??? ![image](https://user-images.githubusercontent.com/10202888/99345907-fb04db80-28c5-11eb-81c2-c74ee6c6e15a.png) ## ??? ![image](https://user-images.githubusercontent.com/10202888/99345965-1cfe5e00-28c6-11eb-8889-f6af39211408.png) ## ??? ![image](https://user-images.githubusercontent.com/10202888/99346004-30a9c480-28c6-11eb-8f53-979a57c97d6f.png) - AI system is not perfect ## Thanks ![image](https://user-images.githubusercontent.com/10202888/99346058-4c14cf80-28c6-11eb-8115-f4a94511ad08.png) - KAIST pernah mencoba melakukan riset AI untuk militer, tapi kemudian dikecam oleh berbagai ilmuwan di dunia, mereka mengatakan "please stop this, it could be dangerous" - Walaupun enggak bisa dipungkiri banyak negara lain yang mungkin seperti indonesia yang juga ingin melakukan hal tersebut
Alfred Boediman. PhD., Samsung Research and Development Institute Indonesia (SRIN), Indonesia
## Model Use Case: I Know You ![image](https://user-images.githubusercontent.com/10202888/99351389-10cccd80-28d3-11eb-8b42-e1815eb022e4.png) ## Node Pruning ![image](https://user-images.githubusercontent.com/10202888/99351437-29d57e80-28d3-11eb-9f41-9ba54f68c975.png) - untuk melakukan proses membutuhkan komputasi yang berat - node pruning merupakan salah satu metode untuk mengurangi banyaknya komputasi yang dibutuhkan - masalah ketiga adalah energy efficiency, karena komputasi yang berat akan memakan lebih banyak daya batere ## Grounding in Nature ![image](https://user-images.githubusercontent.com/10202888/99351594-82a51700-28d3-11eb-9378-66bb4764fdd9.png) - secara natural, pruning terjadi dalam otak manusia, dimana dibandingkan ketika umur satu tahun, remaja memiliki jumlah synapses yang lebih sedikit ## Timeline ![image](https://user-images.githubusercontent.com/10202888/99351650-a5cfc680-28d3-11eb-84db-f6b93b3549ba.png) ## Deep Learning in Microcontrollers ![image](https://user-images.githubusercontent.com/10202888/99351764-ecbdbc00-28d3-11eb-9255-68d5cc305fb9.png) - saat ini ada sejuta embedded device di dunia dan akan semakin meningkat jumlahnya. ## Obsession With Model Compression ![image](https://user-images.githubusercontent.com/10202888/99352048-84bba580-28d4-11eb-91a6-47ed53907296.png) - mobile device memiliki kelemahan pada kemampuan komputasi dan energy - sehingga ketika model memiliki ukuran yang besar akan lebih susah dideploy pada aplikasi mobile. - ada dua pendekatan untuk mengatasi masalah tersebut, melakukan pruning pada model (mengurangi ukuran model), atau meningkatkan kemampuan hardware ## Fundamental On-Device ML Challenges ![image](https://user-images.githubusercontent.com/10202888/99352193-c0566f80-28d4-11eb-9f71-a22b42c311b1.png) ## Automated Specialization ![image](https://user-images.githubusercontent.com/10202888/99352325-04e20b00-28d5-11eb-83cf-374cc15a41a7.png) ![image](https://user-images.githubusercontent.com/10202888/99352443-3bb82100-28d5-11eb-8a1c-97d6f92fe472.png) ## Automated Specialization Use Case ![image](https://user-images.githubusercontent.com/10202888/99352548-686c3880-28d5-11eb-970d-d775edde3c0f.png) ## Memory and Compute Sharing ![image](https://user-images.githubusercontent.com/10202888/99352658-9cdff480-28d5-11eb-8462-50b16e6062ed.png) ![image](https://user-images.githubusercontent.com/10202888/99352716-b41ee200-28d5-11eb-8e00-af7e068d88c9.png) ## Memory and Compute Sharing Use Case ![image](https://user-images.githubusercontent.com/10202888/99352866-f7795080-28d5-11eb-9ac3-d3aa335bb7ee.png) ## On-Device AI Goes Beyond Classification ![image](https://user-images.githubusercontent.com/10202888/99353029-5212ac80-28d6-11eb-8ee8-6bb4e36923d6.png) - mobile device mulai cukup kuat untuk menjalankan machine learning, tanpa perlu adanya bantuan dari cloud ## SOA Accuracy Will Come From Efficient Models ![image](https://user-images.githubusercontent.com/10202888/99353147-88502c00-28d6-11eb-823e-8ec061e02051.png) ## ??? ![image](https://user-images.githubusercontent.com/10202888/99353514-4378c500-28d7-11eb-8693-661caf563e8b.png) ## Research Use Case ![image](https://user-images.githubusercontent.com/10202888/99353618-758a2700-28d7-11eb-8ef8-f6994b158f70.png) ## Test Result ![image](https://user-images.githubusercontent.com/10202888/99353727-aff3c400-28d7-11eb-9e91-9eefbb37fb6a.png) ## ??? ![image](https://user-images.githubusercontent.com/10202888/99353840-e7627080-28d7-11eb-9b43-06c6ebf5b1b8.png) ## ??? ![image](https://user-images.githubusercontent.com/10202888/99353857-f21d0580-28d7-11eb-9e9d-53bc892499bc.png) ## On Device Research Challenges ![image](https://user-images.githubusercontent.com/10202888/99353971-27295800-28d8-11eb-92d0-03e6401e1cd0.png) - AI telah menjadi trend setter di global market - semua industri, kecil besar sekarang mendapat benefit dari adanya AI