Deep Learning

Using large deep learning models on limited hardware or edge devices is definitely prohibitive. There are methods to compress large models by orders of magnitude and maintain similar accuracy during inference.
In this episode I explain one of the first methods: knowledge distillation
 Come join us on Slack
Reference
Distilling the Knowledge in a Neural Network https://arxiv.org/abs/1503.02531
Knowledge Distillation and Student-Teacher Learning for Visual Intelligence: A Review and New Outlooks https://arxiv.org/abs/2004.05937

In this episode I briefly explain the concept behind activation functions in deep learning. One of the most widely used activation function is the rectified linear unit (ReLU). While there are several flavors of ReLU in the literature, in this episode I speak about a very interesting approach that keeps computational complexity low while improving performance quite consistently.
This episode is supported by pryml.io. At pryml we let companies share confidential data. Visit our website.
Don't forget to join us on discord channel to propose new episode or discuss the previous ones. 
References
Dynamic ReLU https://arxiv.org/abs/2003.10027

One of the best features of neural networks and machine learning models is to memorize patterns from training data and apply those to unseen observations. That's where the magic is. However, there are scenarios in which the same machine learning models learn patterns so well such that they can disclose some of the data they have been trained on. This phenomenon goes under the name of unintended memorization and it is extremely dangerous.
Think about a language generator that discloses the passwords, the credit card numbers and the social security numbers of the records it has been trained on. Or more generally, think about a synthetic data generator that can disclose the training data it is trying to protect. 
In this episode I explain why unintended memorization is a real problem in machine learning. Except for differentially private training there is no other way to mitigate such a problem in realistic conditions.At Pryml we are very aware of this. Which is why we have been developing a synthetic data generation technology that is not affected by such an issue.
 
This episode is supported by Harmonizely. Harmonizely lets you build your own unique scheduling page based on your availability so you can start scheduling meetings in just a couple minutes.Get started by connecting your online calendar and configuring your meeting preferences.Then, start sharing your scheduling page with your invitees!
 
References
The Secret Sharer: Evaluating and Testing Unintended Memorization in Neural Networkshttps://www.usenix.org/conference/usenixsecurity19/presentation/carlini

Why so much silence? Building a company! That's why :) I am building pryml, a platform that allows data scientists build their applications on data they cannot get access to. This is the first of a series of episodes in which I will speak about the technology and the challenges we are facing while we build it. 
Happy listening and stay tuned!

In the last episode of 2019 I speak with Filip Piekniewski about some of the most worth noting findings in AI and machine learning in 2019. As a matter of fact, the entire field of AI has been inflated by hype and claims that are hard to believe. A lot of the promises made a few years ago have revealed quite hard to achieve, if not impossible. Let's stay grounded and realistic on the potential of this amazing field of research, not to bring disillusion in the near future.
Join us to our Discord channel to discuss your favorite episode and propose new ones. I would like to thank all of you for supporting and inspiring us. I wish you a wonderful 2020!Francesco and the team of Data Science at Home

Some of the most powerful NLP models like BERT and GPT-2 have one thing in common: they all use the transformer architecture. Such architecture is built on top of another important concept already known to the community: self-attention.In this episode I explain what these mechanisms are, how they work and why they are so powerful.
Don't forget to subscribe to our Newsletter or join the discussion on our Discord server
 
References
Attention is all you need https://arxiv.org/abs/1706.03762
The illustrated transformer https://jalammar.github.io/illustrated-transformer
Self-attention for generative models http://web.stanford.edu/class/cs224n/slides/cs224n-2019-lecture14-transformers.pdf

Generative Adversarial Networks or GANs are very powerful tools to generate data. However, training a GAN is not easy. More specifically, GANs suffer of three major issues such as instability of the training procedure, mode collapse and vanishing gradients.
 
In this episode I not only explain the most challenging issues one would encounter while designing and training Generative Adversarial Networks. But also some methods and architectures to mitigate them. In addition I elucidate the three specific strategies that researchers are considering to improve the accuracy and the reliability of GANs.
 
The most tedious issues of GANs
 
Convergence to equilibrium
 
A typical GAN is formed by at least two networks: a generator G and a discriminator D. The generator's task is to generate samples from random noise. In turn, the discriminator has to learn to distinguish fake samples from real ones. While it is theoretically possible that generators and discriminators converge to a Nash Equilibrium (at which both networks are in their optimal state), reaching such equilibrium is not easy. 
 
Vanishing gradients
 
Moreover, a very accurate discriminator would push the loss function towards lower and lower values. This in turn, might cause the gradient to vanish and the entire network to stop learning completely. 
 
Mode collapse
 
Another phenomenon that is easy to observe when dealing with GANs is mode collapse. That is the incapability of the model to generate diverse samples. This in turn, leads to generated data that are more and more similar to the previous ones. Hence, the entire generated dataset would be just concentrated around a particular statistical value. 
 
The solution
 
Researchers have taken into consideration several approaches to overcome such issues. They have been playing with architectural changes, different loss functions and game theory.
 
Listen to the full episode to know more about the most effective strategies to build GANs that are reliable and robust. Don't forget to join the conversation on our new Discord channel. See you there!
 

What happens to a neural network trained with random data?
Are massive neural networks just lookup tables or do they truly learn something? 
Today’s episode will be about memorisation and generalisation in deep learning, with Stanislaw Jastrzębski from New York University.
Stan spent two summers as a visiting student with Prof. Yoshua Bengio and has been working on 
Understanding and improving how deep network generalise
Representation Learning
Natural Language Processing
Computer Aided Drug Design
 
What makes deep learning unique?
I have asked him a few questions for which I was looking for an answer for a long time. For instance, what is deep learning bringing to the table that other methods don’t or are not capable of? Stan believe that the one thing that makes deep learning special is representation learning. All the other competing methods, be it kernel machines, or random forests, do not have this capability. Moreover, optimisation (SGD) lies at the heart of representation learning in the sense that it allows finding good representations. 
 
What really improves the training quality of a neural network?
We discussed about the accuracy of neural networks depending pretty much on how good the Stochastic Gradient Descent method is at finding minima of the loss function. What would influence such minima?Stan's answer has revealed that training set accuracy or loss value is not that interesting actually. It is relatively easy to overfit data (i.e. achieve the lowest loss possible), provided a large enough network, and a large enough computational budget. However, shape of the minima, or performance on validation sets are in a quite fascinating way influenced by optimisation. Optimisation in the beginning of the trajectory, steers such trajectory towards minima of certain properties that go much further than just training accuracy.
As always we spoke about the future of AI and the role deep learning will play.
I hope you enjoy the show!
Don't forget to join the conversation on our new Discord channel. See you there!
 
References
 
Homepage of Stanisław Jastrzębski https://kudkudak.github.io/
A Closer Look at Memorization in Deep Networks https://arxiv.org/abs/1706.05394
Three Factors Influencing Minima in SGD https://arxiv.org/abs/1711.04623
Don't Decay the Learning Rate, Increase the Batch Size https://arxiv.org/abs/1711.00489
Stiffness: A New Perspective on Generalization in Neural Networks https://arxiv.org/abs/1901.09491

In this episode I speak with Jon Krohn, author of Deeplearning Illustrated a book that makes deep learning easier to grasp. 
We also talk about some important guidelines to take into account whenever you implement a deep learning model, how to deal with bias in machine learning used to match jobs to candidates and the future of AI. 
 
 
You can purchase the book from informit.com/dsathome with code DSATHOME and get 40% off books/eBooks and 60% off video training

Join the discussion on our Discord server
In this episode I explain how a research group from the University of Lubeck dominated the curse of dimensionality for the generation of large medical images with GANs. The problem is not as trivial as it seems. Many researchers have failed in generating large images with GANs before. One interesting application of such approach is in medicine for the generation of CT and X-ray images.Enjoy the show!
 
References
Multi-scale GANs for Memory-efficient Generation of High Resolution Medical Images https://arxiv.org/abs/1907.01376