Daniel Bourke
For the last three years, the State of AI Report has been published as a snapshot of what’s happened in the field of artificial intelligence over the past 12 months. This is my review/walkthrough of the 2020 version.
A big shout out to Nathan and Ian for their huuuuuuge efforts putting this together!
See the full State of AI Report 2020 here: https://www.stateof.ai
Connect:
Web – https://www.mrdbourke.com
Email updates – https://www.mrdbourke.com/newsletter
Timestamps:
0:00 – Intro and hello
1:31 – AI report review start
2:12 – AI definitions
3:00 – SECTION 1: Research
3:37 – Transformers taking over NLP
5:50 – Universities starting AI-based degrees
6:35 – Only 15% of papers published share their code/data
8:20 – PyTorch outpacing TensorFlow in research papers
11:12 – Bigger models, datasets and compute budgets drive performance
15:40 – Increased performance costing more for incremental improvement
18:13 – Deep learning is getting more efficient
21:11 – AI for conversations is getting better
23:02 – Machine translation for code (Python to C++)
25:32 – Many algorithms starting to beat human baseline for NLP on GLUE test
27:04 – Using Transformers for computer vision
30:11 – AI performs incredibly well on mammography tasks across two regions (US and UK)
31:30 – Causal inference in ML
34:30 – ML for synthesizing new molecules
36:26 – AI starts to read DNA-encoded molecules
39:39 – AI generates tennis matches between any tennis players you want
40:46 – Transformers being used for object detection
41:46 – AI which learns from its dreams
44:26 – Really efficient on-device computer vision models
45:04 – Evolving machine learning algorithms from scratch (AutoML Zero)
46:20 – Federated learning is now booming
47:16 – Privacy-preserving ML
48:35 – Using Gaussian Processes for estimating model uncertainty
50:11 – SECTION 2: Talent
50:12 – Many top companies stealing AI professors
52:02 – Abu Dhabi opens the world’s first AI university
54:03 – Many Chinese AI PhD’s depart China for other countries
54:52 – US based companies and institutions dominate NeurIPS and ICML (ML conferences)
56:13 – Three times more AI job postings then views for AI roles
56:58 – TensorFlow and Keras have more job postings on LinkedIn than PyTorch
57:18 – SECTION 3: Industry
59:06 – INTERMISSION
59:44 – AI predicting metabolic response to food
1:00:45 – FDA acknowledge lack of policy for AI-driven systems
1:01:54 – Less than 1% of AI-based medical imaging studies are high-quality
1:02:48 – First reimbursement approval for deep learning based medical imagining
1:05:16 – Self-driving mileage in California still well above autonomous driving mileage
1:10:36 – Supervised ML improvements seem to follow an S-curve
1:12:13 – A new kind of approach to self-driving cars
1:17:17 – New AI-first chips (competition for NVIDIA)
1:19:41 – The rise of MLOps
1:21:52 – Computer vision for auto insurance claims
1:23:45 – Using NLP to detect money laundering and terrorist schemes on the web
1:25:17 – Robots in stock factories are picking millions of items per month
1:26:23 – HuggingFace’s open-source NLP work is driving NLP’s explosion
1:28:09 – SECTION 4: Politics
1:29:59 – Creating a search engine for faces
1:33:29 – GPT3 outputs bias predictions like GPT2
1:33:49 – US military adopting deep reinforcement learning techniques
1:35:52 – Fighter pilots vs AI pilots
1:38:28 – Google’s People and AI guidebook talks about fairness, interpretability, privacy
1:40:46 – China fronts big cash for chip manufacturing
1:45:00 – A call to tackle climate change, food waste, generating new battery technologies and more with ML
1:45:45 – SECTION 5: Predictions
1:49:22 – A special guest appears
#artificialintelligence #ai #machinelearning