Improving Deep Lesion Detection Using 3D Contextual and Spatial Attention
https://deepai.org/publication/improving-deep-lesion-detection-using-3d-contextual-and-spatial-attention by Qingyi Tao et al.
#ComputerVision #PatternRecognition

✴️ @AI_Python_EN

REINFORCE (Section 6 of Williams, 1992) *is* an ES, and is very closely related to OpenAI-ES, NES, CMA-ES.

Simple Random Architecture Search with hand-engineered components already gets ~ SOTA for CIFAR10, PTB

https://github.com/hardmaru/gecco-tutorial-2019 …
✴️ @AI_Python_EN

Cheat sheet: TensorFlow, an open source software library for machine learning
https://www.techrepublic.com/article/tensorflow-googles-open-source-software-library-for-machine-learning-the-smart-persons-guide/ …
#machinelearning

✴️ @AI_Python_EN

Jupyter Notebook Documentation

✴️ @AI_Python_EN

Time-series analysis (TSA) is used in many fields, including finance, economics, meteorology and marketing.

Despite this, I occasionally hear the criticism that it doesn't actually work because these models assume normally distributed data and cannot handle non-linearities.

Academic researchers have studied time-series analysis for many decades, so this is not an obscure topic. A gigantic amount of research on TSA has been conducted and is public domain.

Time-series data are seldom linear, so the first objection does not make sense. Secondly, normality need not be assumed - this is a myth.

For those who'd like to learn more about TSA, two excellent introductions are:

- Time Series Analysis and Its Applications (Shumway and Stoffer)
- Time Series Analysis (Wei)

There are many others on specialized topics, such as:

- Multiple Time-Series Analysis (Lütkepohl)
- Time Series Analysis by State Space Methods (Durbin and Koopman)
- Time Series Modelling with Unobserved Components (Pelagatti)
- Hidden Markov Models for Time Series (Zucchini)
- GARCH Models (Francq and Zakoïan)
- Handbook of Volatility Models and Their Applications (Bauwens et al.)
- Nonparametric Econometrics (Li and Racine)

✴️ @AI_Python_EN

Google Machine Learning Dictionary

https://developers.google.com/machine-learning/glossary/

✴️ @AI_Python_EN

Developing the control of an exoskeleton with #Matlab & #Simulink to help patients relearn how to walk. The Statistics & Machine Learning, Curve Fitting Toolboxes were also used.

Students and Mechatronic engineers may want to check out the links in Comment-1 for a Youtube short tutorial on an exoskeleton designed with Simulink plus some research papers that I came across on the topic. There are many, but I simply list 3. See Comment-2 for the link to Lokomat's website (which is mentioned in the MathWorks blog article) to see their exoskeleton products.
https://blogs.mathworks.com/headlines/2017/06/28/robotic-exoskeleton-helps-patients-relearn-how-to-walk/

✴️ @AI_Python_EN

Really awesome #deeplearning #rnn paper 👏🏻 explaining an increase in predicted risk for clinical alerts.

Explanations in a temporal setting where a stateful dynamical model produces a sequence of risk estimates given an input at each time step.

The goal here is to alert a clinician when a patient’s risk of deterioration rises. The clinician then has to decide whether to intervene and adjust the treatment.

Given a potentially long sequence of new events since she last saw the patient, a concise explanation helps her to quickly triage the alert.

Authors developed methods to lift static attribution techniques to the dynamical setting, where they identified and addressed challenges specific to dynamics.

They then experimentally assess the utility of different explanations of clinical alerts through expert evaluation.

Here’s full paper: https://lnkd.in/duMQYyW
#healthcare #diagnostics #clinical
#prediction
✴️ @AI_Python_EN

Model evaluation techniques in one picture,
✴️ @AI_Python_EN

10 ML Books that You Need To Consider

✴️ @AI_Python_EN

Facebook & Carnegie Mellon build first AI that beats pros in 6-player Hold’em Poker. They tested Pluribus against professional poker players, including two winners of the World Series of Poker Main Event. Pluribus won decisively.

It uses Monte Carlo Counterfactual Regret Minimization algorithm which updates the traverser’s strategy by assessing the value of real and hypothetical moves.
In Pluribus, this traversal is actually done in a depth-first manner for optimization purposes.

This algorithm/bot is priceless!!
Link: https://lnkd.in/eHzfdsm
#artificialintelligence

✴️ @AI_Python_EN

Deep learning based super resolution, without using a GAN

https://towardsdatascience.com/deep-learning-based-super-resolution-without-using-a-gan-11c9bb5b6cd5

✴️ @AI_Python_EN

Size-free generalization boundsfor convolutional neural networks
https://arxiv.org/pdf/1905.12600.pdf

✴️ @AI_Python_EN

Dopamine for RL Researchers.
https://github.com/google/dopamine

✴️ @AI_Python_EN

R-Transformer: Recurrent Neural Network Enhanced Transformer "empirical results show that R-Transformer outperforms the state-of-the-art methods by a large margin in most of the tasks"
https://arxiv.org/abs/1907.05572
https://github.com/DSE-MSU/R-transformer

✴️ @AI_Python_EN

Machine learners commonly used in predictive analytics all have at least some roots in statistics.

Some, such as regression, PCA and K-means, have been considered statistical methods since they were first developed.

Newer methods such as random forests and XGBoost will often outperform methods such as linear and binary logistic regression.

However, this depends on how the regression methods are used - statisticians often criticize data scientists for using methods like these incorrectly or for not using newer variations better suited to a particular data set.

In general, it takes more work to get statistical methods right, but this effort may pay off in the long term. Statistical methods can also be automated once the general parameters are known.

When it comes to helping us understand our data and the process(es) which generated them, as well as for what if simulations, statistical methods are better suited.

I think there is also confusion about tasks. Data are not data, and there are some tasks for which newer approaches such as Deep Learning (DL) are better suited. NLP and speech recognition are two examples.

However, because methods such as DL are better suited to "exotic" tasks than statistics, this does not mean they are better for any task (e.g., survey analytics).

✴️ @AI_Python_EN

Data scientists typically use principal components analysis (PCA) as a data reduction tool. PCA and a related method, factor analysis (FA), can do much more than this, however.

Either method can also help us understand how the variables in our data interrelate. Rotation of components and factors is especially helpful in this regard.

Unfortunately, many data scientists never use rotation and many marketing researchers habitually use varimax orthogonal rotation.

Here are some other rotation methods, many with oblique variations which allow the components/factors to be correlated:

- GEOMIN
- QUARTIMIN
- QUARTIMAX
- EQUAMAX
- PARSIMAX
- PROMAX
- OBLIMIN
- OBLIMAX
- PROCRUSTEAN
- BI-FACTOR
- BENTLER
- CRAWFORD-FERGUSON

The eigenvalue >= 1 guideline for selecting the number of components or factors is not always best, either.

There are also numerous factor extraction methods that are seldom used even when they would be more appropriate.

Moreover, we are not bound to assume our data are continuous, and I often use variations of factor analysis designed for ordinal or nominal data.

There is also no need to assume one factor solution fits all - factor mixture modeling has rescued me on more than one occasion.

✴️ @AI_Python_EN

"Multi-Label Learning", MLL (https://lnkd.in/ePhb3Fy) is a classification problem where many labels can be assigned to each instance. The more general multiple outputs/targets learning subsumes MLL for categorical targets. Wei Tong, et al, proposed SafeML (safe multi-label) model in [1] for prediction of weakly labeled data, that is when relevant labels of examples are partially known or missing which means the MLL method does not hurt performance when using weakly labeled data. The #Matlab code is in [2].

Other related posts on:
a) #MultiLabelLearning posts are here: https://lnkd.in/g_FhHDq
b) #MultiTargetLearning are here: https://lnkd.in/gxxns3a

Abstract:
Here, the MLL with weakly labeled data, i.e, labels of training examples are incomplete, which commonly occurs in real applications, e.g, image classification, document categorization is studied. This setting includes, e.g, (i) semi-supervised multi-label learning where completely labeled examples are partially known; (ii) weak label learning where relevant labels of examples are partially known; (iii) extended weak label learning where relevant & irrelevant labels of examples are partially known.

[1]"Learning safe multilabel prediction for weakly labeled data"-pdf
https://lnkd.in/g73gWJU

[2]Code
https://lnkd.in/gk4uvcH

✴️ @AI_Python_EN

Build Graph Nets in Tensorflow https://arxiv.org/abs/1806.01261

Graph Nets is DeepMind's library for building graph networks in Tensorflow and Sonnet.

Contact graph-nets@google.com for comments and questions.

What are graph networks?
A graph network takes a graph as input and returns a graph as output. The input graph has edge- (E ), node- (V ), and global-level (u) attributes. The output graph has the same structure, but updated attributes. Graph networks are part of the broader family of "graph neural networks" (Scarselli et al., 2009).

https://lnkd.in/dpG9e7g

✴️ @AI_Python_EN

Relating sentence representations in deep neural networks with those encoded by the brain

In their investigation, the researchers considered several neural network architectures for word representation, including two recently proposed models called ELMo and BERT.

They compared how these networks process particular sentences with data collected from human subjects using magnetoencephalography (MEG), a functional neuroimaging technique for mapping brain activity, as they read the same sentences.

To begin with, they decided to use sentences with a simple syntax and basic semantics, such as "the bone was eaten by the dog."

Read more below: or watch this Github link:
https://lnkd.in/dvbiJAz
#deeplearning

✴️ @AI_Python_EN

Classifying Legendary Pokemon Birds 🐦🐦🐦

👉👉👉 Try it yourself:
https://lnkd.in/eYhKNAh 👈👈👈

After only the second fastai "Practical Deep Learning for Coders" class I was able to complete an end-to-end deep learning project! 🤖🤖🤖

The main goal is to classify an image as either one of the Legendary Pokemon Birds - Articuno, Moltres or Zapdos - or an alternative class which includes everything else. Needless to say that sometimes my model gets confused about the alternative class since not so many diverse images were feed into it...

Source code:
https://lnkd.in/eRfkBx8
Forked from:
https://lnkd.in/e_k4nqN

#ai #ml #dl #deeplearning #cnn #python

✴️ @AI_Python_EN