## Nachhaltige Digitalisierung durch ressourcenschonen-des Machine Learning <!-- Put the link to this slide here so people can follow slide: https://hackmd.io/@ak180rxXTr6byV0g1638KA/B1CcegCes --> ### 21. September 2022 #### Markus Ankenbrand <img src="https://www.uni-wuerzburg.de/fileadmin/_processed_/4/b/csm_ESF-ZDEX_Logo_neu_transparent_ec9b426f56.png" height="100px" style="background-color: white"/> ##### ESF-ZDEX Serie: Zukunftsthema - Industrie 4.0 und Nachhaltigkeit --- ## Definition **Machine Learning**: > the use and development of computer systems that are able to learn and adapt without following explicit instructions, by using algorithms and statistical models to analyse and draw inferences from patterns in data. -- Oxford Languages --- ## AI vs Machine Learning vs Deep Learning <img src="https://www.smart-digital.de/wp-content/uploads/2020/03/200218_AI_ML_DL_Venn_Diagram.png" width="70%"/> <small>Quelle: https://www.smart-digital.de/en/beyond-the-hype-ai-machine-learning-and-deep-learning-explained/</small> --- ## Neue Möglichkeiten durch AI/ML/DL --- ### Computer Vision  z.B. Automatische Analyse von Herz-MRT --- ### Textgenerierung <img src="https://i.imgur.com/vXxmnHs.png"/> <small>Quelle: https://beta.openai.com/playground</small> --- ### Bildgenerierung > a gentleman squirrel in a 19th century portrait <img src="https://i.imgur.com/1EDckil.png" width="25%"/> <small>Quelle: https://replicate.com/stability-ai/stable-diffusion</small> --- ### Spiele - besser als die Großmeister  <small>Quelle: https://levelup.gitconnected.com/alphago-beats-the-worlds-best-go-player-1d4ab1428bac</small> AlphaGo, AlphaZero, AlphaStar, ... --- ## Neue Möglichkeiten durch AI/ML/DL - Bildanalyse - Textgernerierung - Bildgenerierung - Spiele - Proteinstrukturen (AlphaFold) - ... --- ## Die Schattenseiten von AI/ML/DL --- ## Die Schattenseiten von AI/ML/DL - Fairness (z.B. Bias Amplification) - Missbrauch (z.B. Massenüberwachung durch Gesichtserkennung) - Enerie Verbrauch --- ### Fairness - Bias Amplification COMPAS (Correctional Offender Management Profiling for Alternative Sanctions) <img src="https://miro.medium.com/max/720/1*UH1SAj_EU0WQXbIEz0v0UA.png" width="40%"/> Doppelt so viele falsch-positive, abhängig von der Hautfarbe <small>Quelle: https://towardsdatascience.com/real-life-examples-of-discriminating-artificial-intelligence-cae395a90070, Bild: von Bill Oxford auf Unsplash</small> --- ### Energiebedarf: Wie schlimm ist es?  <small>Quelle: https://www.economist.com/technology-quarterly/2020/06/11/the-cost-of-training-machines-is-becoming-a-problem</small> --- ### Energiebedarf: Wie schlimm ist es? > [...] researchers have calculated that using A100s would have taken 1,024 GPUs, 34 days and $4.6million to train the model. While energy usage has not been disclosed, it’s estimated that GPT-3 consumed 936 MWh. <small>Quelle: https://numenta.com/blog/2022/05/24/ai-is-harming-our-planet</small> --- ## Was können wir tun? - Klassische ML Methoden statt Deep Learning - Transfer Learning - Shallow/Sparse Networks - Spezialisierte Hardware - Effiziente Rechenzentren --- ### Klassisches Machine Learning Eine kurze Tour klassischer Klassifikationsalgorithmen --- #### Logistische Regression  <small>Quelle: https://towardsdatascience.com/top-machine-learning-algorithms-for-classification-2197870ff501</small> --- #### K-Nearest Neighbor  <small>Quelle: https://towardsdatascience.com/top-machine-learning-algorithms-for-classification-2197870ff501</small> --- #### Entscheidungsbaum  <small>Quelle: https://towardsdatascience.com/top-machine-learning-algorithms-for-classification-2197870ff501</small> --- #### Random Forest  <small>Quelle: https://towardsdatascience.com/top-machine-learning-algorithms-for-classification-2197870ff501</small> --- #### Support Vector Machine  <small>Quelle: https://towardsdatascience.com/top-machine-learning-algorithms-for-classification-2197870ff501</small> --- #### Naïve Bayes  <small>Quelle: https://towardsdatascience.com/top-machine-learning-algorithms-for-classification-2197870ff501</small> --- #### Positive Nebeneffekte Neben Energie/CO₂-Ersparnis auch: - schnelleres Training - schnellere Anwendung - einfacheres Deployment (günstigere Hardware) - einfacher an neue Daten anpassbar - einfacher Fachwissen einzubeziehen - oft bessere Interpretierbarkeit --- ### Transfer Lernen  <small>Quelle: https://medium.datadriveninvestor.com/introducing-transfer-learning-as-your-next-engine-to-drive-future-innovations-5e81a15bb567</small> --- ### Shallow/Sparse Networks  <small>Quelle: https://developer.nvidia.com/blog/accelerating-inference-with-sparsity-using-ampere-and-tensorrt/</small> --- ### Spezialisierte Hardware  <small>Quelle: https://en.wikipedia.org/wiki/Tensor_Processing_Unit</small> --- ### Effiziente Rechenzentren  <small>Quelle: https://www.hpcwire.com/2021/05/20/google-launches-tpu-v4-ai-chips/</small> --- ## CO₂-Verbrauch transparent machen  [green-algorithms.org](https://green-algorithms.org) --- ### Danke! Fragen?