exploring artificial intelligence
neural networks for life sciences
Felipe Delestro
contact@delestro.com
1. Introduction
2. Basics of Artificial Neural Networks
3. ANNs for embryo classification
4. How to get started
Layout
Introduction
- Artificial Neural Networks (ANNs) are computational models inspired by biological neurons in the brain
- ANNs recognize patterns in data and make predictions based on those patterns
- ANNs are typically organized into layers of neurons, with each layer performing a different type of processing
Overview
In the 1920s, McCulloch and Pitts developed the first mathematical model of a neuron, known as the McCulloch-Pitts neuron
McCulloch WS, Pitts W.
A logical calculus of the ideas immanent in nervous activity.
The bulletin of mathematical biophysics. 1943
During the 1950s Frank Rosenblatt developed the first neural network model capable of learning, known as the Perceptron
Perceptron , Mark I
National Museum of American History
Washington, D.C.
Basics of Artificial Neural Networks
- Perceptron is a mathematical model that mimics the behavior of a single neuron
- It takes inputs, applies weights, sums them up, and passes the result through an activation function to produce an output
- Perceptrons are essentially linear classifiers that make decisions by combining inputs with weights and producing output using an activation function
Perceptrons
https://towardsdatascience.com/perceptron-the-artificial-neuron-4d8c70d5cc8d
https://medium.com/@shrutijadon/survey-on-activation-functions-for-deep-learning-9689331ba092
Weights are learned through training, where the perceptron is presented with a set of labeled examples and it adjusts its weights in order to correctly classify the inputs
Perceptron cooking as a chef
Stable diffusion, OpenJourney model
https://scikit-learn.org/stable/modules/clustering.html
On its own, the perecptron is limited to linear decision boundaries.
Adding layers
- Multi-layer perceptrons (MLPs) or feedforward neural networks were developed to overcome the linearity limitation
- MLPs have multiple layers, including input, hidden, and output layers
- MLPs can learn more complex features and relationships between inputs
Visualization of a fully connected neural network, version 1
Marijn van Vliet
MLP as a detective solving a crime
Stable diffusion, OpenJourney model
Backpropagation
- Backpropagation is an algorithm for training ANNs to learn from their mistakes and improve their accuracy over time
- Backpropagation adjusts the weights of the connections in the network to minimize the difference between the actual output and the desired output
- It propagates the error backwards through the network, from output layer to input layer, and adjusts the weights based on the contribution of each connection to the error
- Backpropagation aims to find the set of weights that produces the smallest error for a given set of input-output pairs
Backpropagation as a choreographer teaching a dance
Stable diffusion, OpenJourney model
Deep learning
- Multi-layer perceptrons (MLPs) have limitations in learning complex patterns, which deep learning can solve by using many layers of neurons.
- The number of layers in deep neural networks can range from a few dozen to several hundred or more, making it possible to learn intricate features and relationships within the data.
- Layers can be specialized, and two types are very important: convolution and pooling layers
Convolution
- Convolution is a weighted sum operation used in signal and image processing to extract features from data.
- A small filter is slid over the input signal or image, and the results of element-wise multiplication are summed up to obtain a single output value.
https://commons.wikimedia.org/wiki/File:2D_Convolution_Animation.gif
https://www.youtube.com/watch?v=KuXjwB4LzSA
But what is a convolution?
Pooling
- Pooling is a technique used in neural networks to reduce the spatial dimensionality of feature maps obtained from convolutional layers.
- The pooling function (such as max pooling or average pooling) reduces the information in each region to a single value, resulting in a smaller feature map with a reduced spatial dimensionality.
- Pooling helps make the subsequent layers of the network more robust to slight variations in the input and provide translational invariance to the input data.
https://www.geeksforgeeks.org/cnn-introduction-to-pooling-layer/
Pooling as understanding a book by summarizing the chapters
Stable diffusion, OpenJourney model
convolution + pooling made possible the creation of the convolutional Neural Networks, or CNNs
They correspond to the vast majority of DL models currently in usage.
ANNs for embryo classification
José Celso
Applied mathematics
Marcelo Nogueira
Embryology
Cummulative Google Scholar results for "deep learning" and "image"
Started to work on embryo classification
Human embryo development during 5 days, data from time-lapse system
User manually perform measurements in the image
Graphical interface allows inference using the ANN
Compromises
- MLPs cannot read directly from the image, a feature extraction step was needed
- The system predicts quality of the embryo, which could be a biased label
Alexandra Boussommier
CEO ImVitro
https://doi.org/10.1093/humrep/dead023
Embryoscope
(Vitrolife)
MIRI
(ESCO)
Geri
(Genea Biomedx)
Automatic embryo cropping using YoloV5. The original model was retrained using a few hundred annotated images
https://github.com/ultralytics/yolov5
https://www.youtube.com/watch?v=cHDLvp_NPOk
YoloV5 model in action for object detection
pregnancy
no pregnancy
AUC: 0.68
video score
clinical features
pregnancy
no pregnancy
AUC: 0.73
How to get started
Knowledge background
-
Python:
A popular programming language for deep learning and is used extensively in the field for tasks such as data manipulation, preprocessing, and training models. -
Machine learning
Understanding the principles and concepts of machine learning is important for developing deep learning models as it helps to identify the appropriate models, evaluate their performance, and make modifications to improve them. -
"Classical" approaches
Knowledge of classical approaches such as linear regression, logistic regression, and decision trees can help understand the fundamentals of deep learning and develop more complex models. It can also help in selecting appropriate preprocessing techniques and handling missing data.
Recommended framework
High level approaches
BioImage model Zoo
thanks!
https://slides.com/delestro/ann-unesp
Exploring Artificial Intelligence
By Felipe Delestro
Exploring Artificial Intelligence
neural networks for life sciences
- 219