Lev Selector - WhatsApp messages (ML & AI), 2020
=================================================
2020-12-16
FeatureTools - open source python framework
for automated feature engineering.
- https://www.youtube.com/watch?v=Q5U9rEKHIsk
- https://www.featuretools.com/
- https://github.com/alteryx/featuretools
- https://innovation.alteryx.com/open-sourcing-featuretools/ (2017)
=================================================
2020-12-15
Open-source streaming visualizations for AI applications
Look at graphs and numbers changing in real time
on this page
- https://www.h2o.ai/products/h2o-wave/
=================================================
2020-12-04
Lex Fridman - very good explanation of AlphaFold 2
- https://www.youtube.com/watch?v=W7wJDJ56c88
=================================================
2020-11-27
Linsu Tech Tips - I can safely retire now (deep fake).
- https://youtu.be/G0z50Am4Uw4
=================================================
2020-11-25
Here are top libraries by stars, number of contributors,
and type.
===============
Top Python Libraries for Data Science, Data Visualization
& Machine Learning
- https://www.kdnuggets.com/2020/11/top-python-libraries-data-science-data-visualization-machine-learning.html
===============
Top Libraries for deep learning, natural language processing,
and computer vision:
- https://www.kdnuggets.com/2020/11/top-python-libraries-deep-learning-natural-language-processing-computer-vision.html
=================================================
2020-11-17
Attending NeurIPS conference used to be very expensive.
You needed to take a week off, travel, pay for hotel
and conference. And seating was limited.
But this year it is all virtual - and very affordable.
Check it out:
December 6 - 12 (Sunday through Saturday), 2020
Only $100 for everything!
- https://nips.cc/
=================================================
2020-11-01
Wow!
New work from Caltech & Purdue.
Fourier neural operator (combining
the power of Deep Neural Network
and Fourier transform) consistently
outperforms all existing deep learning
methods for parametric PDEs
(Partial Differential Equations).
- https://twitter.com/techreview/status/1322595041899171841
- https://www.technologyreview.com/2020/10/30/1011435/ai-fourier-neural-network-cracks-navier-stokes-and-partial-differential-equations/
- https://arxiv.org/pdf/2010.08895.pdf
"Why does this matter?
Because it’s far easier to approximate
a Fourier function in Fourier space
than to wrangle with PDEs in Euclidean space,
which greatly simplifies the neural network’s job.
Cue major accuracy and efficiency gains:
in addition to its huge speed advantage
over traditional methods, their technique
achieves a 30% lower error rate when
solving Navier-Stokes than previous
deep-learning methods."
=================================================
2020-10-30
nbdev - a toolset to do all your python
development in Jupyter notebooks.
Here is a lecture where Jeremy Howard
explains why he is using notebooks,
and how to use them effectively.
This one lecture gives you tons of value,
advice and good practicies.
Jeremy is amazing:
- https://www.youtube.com/watch?v=9Q6sLbz37gk
=================================================
2020-10-26
Universal translators from different companies:
- Facebook M2M-100 - translating 100 languages, open source
- Microsoft T-ULRv2 - translating 94 languages, top results
- Google MT5 - translating more than 100 languages, open source
- https://venturebeat.com/2020/10/26/google-open-sources-mt5-a-multilingual-model-trained-on-over-101-languages/
=================================================
2020-10-24
RISC-V is the future of computing - Chris Lattner and Lex Friedman
- https://youtu.be/lXdx0X2WHfY
=================================================
2020-10-22
Generate books online and PDF (ready for publishing):
using fastdoc project:
- https://github.com/fastai/fastdoc
- https://asciidoctor.org/
using Jupyter Book project:
- https://jupyterbook.org/intro.html
- https://jupyterbook.org/advanced/pdf.html
=================================================
2020-10-21
In 2005, neuroscientists at CalTech and UCLA discovered
a single neuron in a patient’s brain that appeared to
respond only to the actress Halle Berry: photos,
caricatures, even the spelling of her name.
- https://www.scientificamerican.com/article/one-face-one-neuron/
"This neuron is responding to the abstract concept
of Halle Berry rather than to any particular visual
feature. ...
This suggests we store memories as abstract concepts ..."
=================================================
2020-10-20
Jeremy Howard & fast.ai
- https://www.fast.ai/ - home, courses
- https://www.fast.ai/about/
- https://www.fast.ai/topics/
Book:
- https://www.amazon.com/Deep-Learning-Coders-fastai-PyTorch/dp/1492045527
Interview:
- https://www.youtube.com/watch?v=6x2D1EKFHe4
=================================================
2020-10-16
AutoML - drop-in replacement for a scikit-learn estimator:
- https://github.com/automl/auto-sklearn
=================================================
2020-10-15
Evolution 2.0 - Origin of Life,
Where did the information come from?
Genetic code vs Artificial Intelligence.
A new $10M prize for a definitive answer.
- https://evo2.org/
- https://www.herox.com/evolution2.0
- https://www.youtube.com/watch?v=oG6dTuQ5z1U
=================================================
2020-10-07
DeepPavlov is an open source framework for chatbots
and virtual assistants development.
- https://deeppavlov.ai/
=================================================
2020-10-04
LSTM is dead. Long Live Transformers!
Great short lecture by Leo Dirac
- https://www.youtube.com/watch?v=S27pHKBEp30
=================================================
2020-08-30
Neuralink: Elon Musk's brain chip summary in 14 min
https://youtu.be/CLUWDLKAF1M
- https://youtu.be/CLUWDLKAF1M
=================================================
2020-08-27
TecoGAN - Super Resolution
TecoGAN (Temporally Coherent GANs) for improving resolution of video and images
TecoGAN can upscale the video quality from 720p to 4K !
- https://www.youtube.com/watch?v=MwCgvYtOLS0
=================================================
2020-08-01
GPT-3 vs Human Brain
GPT-3 has 175 Billion parameters/synapses.
Human brain has 100trillion synapses.
- https://youtu.be/kpiY_LemaTc
=================================================
2020-06-14
Two open source projects everyone should know about:
===================
ONNX Runtime (Microsoft + Facebook) is a
cross-platform inferencing and training accelerator:
- https://github.com/microsoft/onnxruntime
Usage: export your trained model (TensorFlow or PyTorch)
into ONNX format, and run efficiently.
- https://cloudblogs.microsoft.com/opensource/2019/05/22/onnx-runtime-machine-learning-inferencing-0-4-release/
===================
Huggingface Transformers
- https://huggingface.co/transformers/
- https://github.com/huggingface/transformers
Here is a very good lecture introducing main idea of transformers:
- https://www.youtube.com/watch?v=S27pHKBEp30
=================================================
2020-06-01
Very nice summary of a recent lecture - Self-Supervised Learning
- https://bdtechtalks.com/2020/03/23/yann-lecun-self-supervised-learning/amp/
=================================================
2020-05-28
About NLP (Natural Language Processing).
Look at Hugging Face, a startup company in Brooklyn, New York,
which develops AI chatbot software (NLP libraries).
- https://github.com/huggingface
- https://huggingface.co/transformers/
Note that modern NLP software usually requires lots
of hardware to run, because systems tend to be huge
- billions of parameters.
OpenAI - last year built the NLP model "GPT-2" with
up to 1.6 Billion parameters.
Recently they received a custom system from Microsoft
with 10,000 GPUs and 285,000 CPUs.
It probably costs ~ $200 Mln to build (or more).
Google Brain offers new chatbot Meena - transformer
architecture using 2.6 Billion parameters.
- https://ai.googleblog.com/2020/01/towards-conversational-agent-that-can.html
Facebook built Generative BST, a transformer-based model
comprising up to 9.4 billion parameters.
Read about chatbot "Blender" here:
- https://ai.facebook.com/blog/state-of-the-art-open-source-chatbot/
Recent record - 17 Billion parameters from Microsoft:
- https://www.microsoft.com/en-us/research/blog/turing-nlg-a-17-billion-parameter-language-model-by-microsoft/
There are also libraries trying to achieve good results with minimum hardware.
Look for example at Google's Reformer
Watch this video by Siraj Raval - The Reformer:
- https://www.youtube.com/watch?v=rNG_hpSyZcE
And here is some code:
- https://github.com/google/trax/tree/master/trax/models/reformer
=================================================
2020-05-28
Just three years ago, experts were forecasting
that jobs in the arts were safe from AI invasion.
But since then robot-artists earned millions in sales and auctions.
- https://news.artnet.com/exhibitions/bucharest-biennial-curated-by-artificial-intelligence-1872342
=================================================
2020-05-14
Financial Trading via Image Classification
- https://arxiv.org/abs/1907.10046v2
Idea: experienced traders develop intuition.
They can recommend to buy or not just looking at the time graph.
It would be fun to train an AI system to do the same.
In this recent paper JPMorgan AI researches have trained
ML algorithms to match the standard "buy" signals
(input - images, output - some standard signal).
It would've been more interesting if they had trained
to predict market moves instead of only training to match
standard signals. I was surprised that they haven't
included RNNs (Recurrent Neural Networks) into their
set of classification methods.
=================================================
2020-05-07
Which comes first: training a model or extracting high-quality features?
New approach - do both simultaneously:
CURL = Contrastive Unsupervised Representations for Reinforcement Learning
- https://arxiv.org/pdf/2004.04136.pdf
=================================================
2020-04-26
AutoML-Zero from Google.
This is the continuation of the work by Quoc V. Le et al from Google.
The idea is to use genetic algorithms
for optimizing Neural Network architectures.
The goal is to use these algorithms to build
ML solutions without human intervention (from zero).
- https://www.sciencealert.com/coders-mutate-ai-systems-to-make-them-evolve-faster-than-we-can-program-them
- https://arxiv.org/abs/2003.03384
=================================================
2020-04-25
I have posted three short articles calculating COVID death rates.
Estimating Coronavirus Death Rate using US numbers
- https://github.com/lselector/ml\_ai\_doc/blob/master/\_posts\_articles/20200423\_COVID\_death\_rates.txt
If US will follow Sweden and remove lockdown, we
will end up with more than 300,000 deaths.
- https://github.com/lselector/ml\_ai\_doc/blob/master/\_posts\_articles/20200424\_Sweden\_deaths.txt
Why there were so many deaths in Italy?
- https://github.com/lselector/ml\_ai\_doc/blob/master/\_posts\_articles/20200425\_Italy\_deaths.txt
=================================================
2020-04-14
This is about self-supervised learning !
Very elegant idea.
- https://www.nytimes.com/2020/04/08/technology/ai-computers-learning-supervised-unsupervised.html
=================================================
2020-04-12
Recent term "Responsible AI Framework"
usually means a combination of 5 principles:
1. Security with validation, monitoring, verification
2. Transparent, explainable, provable
3. Ethical, understandable, legal
4. Governance with AI operating models, processes
5. Test for bias in data, models, human use of algorithms
=================================================
2020-03-27
- https://www.youtube.com/watch?v=qFmaSNP6\_z4
This is a very good 5-min video by Lex Fridman, host of AI interviews.
He talks about statistics, that masks work.
He talks about switching from
"look at the idiot wearing mask"
to
"look at the idiot not wearing mask".
And also shows how to make simple masks from old T-shirt in 2 minutes.
=================================================
2020-03-18
Coronavirus Spreading Simulations
- https://www.washingtonpost.com/graphics/2020/world/corona-simulator/
=================================================
2020-03-17
White House 'call to action' for AI
- https://techcrunch.com/2020/03/16/coronavirus-machine-learning-cord-19-chan-zuckerberg-ostp/
CORD-19 data set:
- https://pages.semanticscholar.org/coronavirus-research
- https://www.kaggle.com/sudalairajkumar/novel-corona-virus-2019-dataset
Research on META:
- https://www.meta.org/feed-previews/covid-19/af84352b-285c-4b73-9254-41610da4413b
Google's website for COVID-19 screening
- https://www.projectbaseline.com/study/covid-19/
=================================================
2020-03-13
Coronavirus Outbreak Trajecties - by country
- https://www.reddit.com/r/CoronavirusUS/comments/fhawki/i\_used\_primary\_data\_from\_who\_to\_predict\_the/
Good DeepMind AI-generated predictions
- https://www.businessinsider.com/google-deepmind-ai-predictions-coronavirus-2020-3
=================================================
2020-03-10
Coronavirus Competition Results (Remdesivir)
- https://youtu.be/EVoZMRmtBkY
=================================================
2020-02-25
To survive, your business has to collect data
and use ML & AI to make better decisions.
If your business does not use ML - it will lose to those who use it.
It is "live or die" situation.
Very simple.
A company needs to implement AI transormation.
Here is a great step-by-step playbook:
- https://landing.ai/ai-transformation-playbook/
=================================================
2020-02-23
Recent interview with Andrew Ng:
- https://www.youtube.com/watch?v=0jspaMLxBig
=================================================
2020-02-20
Talking about reliability of software.
It is estimated that standard "test-driven" development
allows to uncover only ~10% of bugs.
But "Fuzz testing" allows to uncover ~ 80%, so only 10%
is left to be found by customers.
Listen to this video "Fuzz or lose ..." - very inspirational.
You can skip the technicalities in this lecture,
just absorb the big picture.
- https://www.youtube.com/watch?v=k-Cv8Q3zWNQ
=================================================
2020-02-14
Citizen Data Scientist (CDS) - a term coined in 2016
by Joao Tapadinhas, Carlie Idoine from Gartner research:
- https://www.gartner.com/en/documents/3534848
The basic idea is to make advanced analytics accessible
to a wider audience by using better tools.
Companies like DataRobot use this idea as their
main marketing statement. They say that you
can not possibly hire all the Data Scientists you need,
but you can buy tools which will allow business
users to do Data Science themselves.
=================================================
2020-02-04
Cryptocurrency - China becomes cashless, Facebook Libra:
- https://www.youtube.com/watch?v=XvJnTQjdC-E
- https://www.youtube.com/watch?v=vPu4kn5GN5M
=================================================
2020-01-26
Self-explainable models.
The model should not only give a prediciton, but
also explain why it made this prediction.
- https://www.kdnuggets.com/2019/01/explainable-ai.html
- https://towardsdatascience.com/python-libraries-for-interpretable-machine-learning-c476a08ed2c7
- https://www.youtube.com/watch?v=Q8rTrmqUQsU
- https://www.youtube.com/watch?v=B-c8tIgchu0
By the way, today is my birthday - 62 years !
=================================================
2020-01-24
Dataset Search has indexed almost 25 million datasets - and growing
- https://datasetsearch.research.google.com/
=================================================
2020-01-21
This year - all about graphs.
Graph algorithms:
- https://www.kdnuggets.com/2019/09/5-graph-algorithms-data-scientists-know.html
Graph databases:
- https://en.wikipedia.org/wiki/Graph\_database
Data managements:
- https://www.kdnuggets.com/2019/10/data-scientist-data-management.html
Graph neural networks:
- https://www.youtube.com/watch?v=bA261BF0bdk
=================================================
2020-01-12
AI in 2020 - predictions from Siraj Raval
- https://www.youtube.com/watch?v=eN9Lb3vXsAw
=================================================
2019-12-04
David Makovoz, a Lead Data Scientist at TCS.com
and a good friend of mine has sent me the
“Gartner's Hype Cycle For AI, 2019”:
- https://www.forbes.com/sites/louiscolumbus/2019/09/25/whats-new-in-gartners-hype-cycle-for-ai-2019/#7fb8774c547b
Here is the main graph:
-
Speech Recognition is predicted to deliver the most
significant transformational benefits of all technologies
on the Hype Cycle.
Some other technologies to consider:
- AI Cloud Services
- AutoML
- Augmented Intelligence
- Explainable AI
- Edge AI
- Reinforcement Learning
- Quantum Computing
- AI Marketplaces
=================================================
=================================================
=================================================
=================================================
- test