Technology, machine learning and algorithms
Technology, machine learning and algorithms
Compressing deep learning models: distillation (Ep.104)
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
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
Pandemics and the risks of collecting data (Ep. 103)
Codiv-19 is an emergency. True. Let's just not prepare for another emergency about privacy violation when this one is over.
Join our new Slack channel
This episode is supported by Proton. You can check them out at protonmail.com or protonvpn.com
Why average can get your predictions very wrong (ep. 102)
Whenever people reason about probability of events, they have the tendency to consider average values between two extremes. In this episode I explain why such a way of approximating is wrong and dangerous, with a numerical example.
We are moving our community to Slack. See you there!
Activate deep learning neurons faster with Dynamic RELU (ep. 101)
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.
Dynamic ReLU https://arxiv.org/abs/2003.10027
WARNING!! Neural networks can memorize secrets (ep. 100)
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!
The Secret Sharer: Evaluating and Testing Unintended Memorization in Neural Networkshttps://www.usenix.org/conference/usenixsecurity19/presentation/carlini
Attacks to machine learning model: inferring ownership of training data (Ep. 99)
In this episode I explain a very effective technique that allows one to infer the membership of any record at hand to the (private) training dataset used to train the target model. The effectiveness of such technique is due to the fact that it works on black-box models of which there is no access to the data used for training, nor model parameters and hyperparameters. Such a scenario is very realistic and typical of machine learning as a service APIs.
This episode is supported by pryml.io, a platform I am personally working on that enables data sharing without giving up confidentiality.
As promised below is the schema of the attack explained in the episode.
Membership Inference Attacks Against Machine Learning Models