Deep Learning and Artificial Intelligence Valerio Maggio, PhD with Pytorch @leriomaggio [email protected] 

Me Pun Who? Background in CS PhD in Machine Learning for Software Engineering Research: ML/DL for BioMedicine wearing face masks since… 

The AI Revolution 

Artificial Intelligence according to Google 

Machine Learning Artificial Intelligence Machine Learning Data Science Deep Pattern Matching ⛑ 

ML and Data Science Data Science Venn diagram By Drew Conway Data Loading Preprocessing Model learning API Interface (Model Adapted from “What about tests in Machine Learning projects?” Sarah Diot-Girard - EuroSciPy 2019 (toy) Data Science Pipeline 

What is Machine Learning 

Machine Learning “Machine learning is the science (and art) of programming computers so they can learn from data” Aurélien Géron, Hands-on Machine Learning with Scikit-Learn and TensorFlow Source: bit.ly/ml-simple-definition “(ml) focuses on teaching computers how to learn without the need to be programmed for specific tasks” S. Pal & A. Gulli, Deep Learning with Keras 

Machine Learning Machine learning teaches machines how to carry out tasks by themselves. It is that simple. The complexity comes with the details Louis Pedro Coelho, Building Machine Learning Systems with Python (and that’s probably one of the reason why you’re here :) 

The (Machine) Learning is about DATA Data are one of the most important part of a ml solution Importance: Data >> Model ? Learning by examples Data Preparation is crucial! data algorithms 

BioMedicine: another data case ? Contemporary Life Science is about data recent advances in sequencing techs and instruments (e.g. “bio-images”) huge datasets generated at incredible pace from human observation to data analysis cheminformatics (drug discovery) Research Impact —> Social and Human Impact 

Why Deep Learning, btw? Subset of ML w/ very specific model: (Deep?) Neural Networks State of the art Theory ’50 / ’80 hw acceleration to train (~new) learning structure + composability (2018/19) 

What about Deep Learning 

Deep Learning A multi-layer feed-forward neural network that starts w/ an input layer fully connected, which is followed by multiple hidden layer of non-linear transformation 

More details… 

More details… ReLu | sigmoid | tanh 

More details… Repeat for each layer… 

More details… Image Classification Task 

More details… Summary: A Neural Network is: • Built from layers; each of which is: • a matrix multiplication, • then add bias • then apply non-linearity Learn values for parameters; W and b (for each layer using Back-Propagation) 

Machine Learning for dummies (a.k.a. ML explained to computer scientists) Note: I *am* a computer scientist + Matrix Multiplication Random Number Generation Machine Learning = ( ) t≅2k Deep 

Ml / Dl basics in a NutShell 

features labels (raw) data ML/DL Model Training Trained Model (unseen) data Test Predictions Supervised learning supervision 

features labels (raw) data ML/DL Model Training Trained Model (unseen) data Test Similarities/likelihood UnSupervised learning 

labels (raw) data DL Model Training Trained Model (unseen) data Test Predictions supervision features Deep Supervised learning 

Supervised Training Loop labels (raw) data Model Parameters Loss loss predictions 

Supervised Training Loop breakdown.. (raw) Data - a.k.a. Observations / Input Items about which we want to predict something. We usually will denote observation with x. Labels - a.k.a. Targets (i.e. Ground Truth) Labels corresponding to observations. These are usually the things being predicted. Following standard notations in ML/DL, we will use y to refer to these. Model f(x) = ˆy A mathematica expression or a function that takes an observation x and predicts the value of its target label. Predictions - a.k.a. Estimates: Values of the Targets generated by the model - usually referred to as ˆy Parameters - a.k.a. Weights (in DL terminology) Parameters of the Model. We will refer to them using the w. Loss Function L(y, ˆy): Function that compares how far off a prediction is from its target for observations in the training data. The loss function assigns a scalar real value called the loss. The lower the value of the loss, the better the model is predicting. The Loss is usually referred to as L Source:D. Rao et al. - Natural Language Processing with PyTorch, O’Reilly 2019 

(DL) Terms: everyone on the same page? also ref: bit.ly/nvidia-dl-glossary Epochs Batches and mini-batch learning Parameters vs HyperParameters (e.g. weights vs layers) Loss & Optimiser (e.g. Cross Entropy & SGD) Transfer learning Gradient & Backward Propagation Tensor 

Python has its say Machine Learning Deep Learning “There should be one, and preferably one, way to do it” The Zen of Python 

Multiple Frameworks? 

If someone tells you … 

Deep Learning Frameworks Static Graph Dynamic Graph X b W * + σ xTW + b (xTW + b) σ Computational Graph Models Linear (or Dense) + + y L y’ fc1 fc2 fc3 fc4 fc5 + + y1 L y’ fc2 fc3 fc4 fc5 fc1 X1 epoch 1, batch 1 + + L y’ fc2 fc3 fc4 fc5 fc1 

Deep Learning Frameworks Static Graph Dynamic Graph X b W * + σ xTW + b (xTW + b) σ Computational Graph Models Linear (or Dense) + + y L y’ fc1 fc2 fc3 fc4 fc5 + + L y’ fc2 fc3 fc4 fc5 fc1 + + y2 L y’ fc2 fc3 fc4 fc5 fc1 X2 epoch 1, batch 2 

Deep Learning Frameworks Static Graph Dynamic Graph X b W * + σ xTW + b (xTW + b) σ Backwards and Gradients Calculation Linear (or Dense) + + y L y’ fc1 fc2 fc3 fc4 fc5 + + y L y’ fc2 fc3 fc1 X fc5 fc4 Backprop Autograd Record 

Deep Learning Frameworks Static Graph Dynamic Graph X b W * + σ xTW + b (xTW + b) σ Backwards and Gradients Calculation Linear (or Dense) + + y L y’ fc1 fc2 fc3 fc4 fc5 + + y L y’ fc2 fc3 fc1 X fc5 fc4 Record Replay Backprop Autograd & 

rundown review of basic PyTorch features we will see soon Spoiler Alert 

Tensors, NumPy, Devices Numpy-like API tensor -> ndarray tensor <- ndarray CUDA support 

Neural Module subclassing Definition of layers (i.e. tensors) Definition of graph (i.e. network) 

Loss and Gradients optimiser criterion & loss backprop & update 

Dataset and DataLoader transformers Dataset DataLoader 

So…. 

Deep learning Repository: http://github.com/leriomaggio/deep-learning-pytorch mybinder Colaboratory (more on this in few minutes) with PyTorch 

Deep Learning Course Approach: Data Scientist Always prefer dev/practical aspects (tools & sw) Work on full pipeline (e.g. data preparation) Emphasis on the implementation Perspective: Researcher No off-the-shelf (so no “black-box”) solutions” References and Further Readings to know more features 

Let’s roll up our sleeves & get on with the hands-on