Advances in breast cancer screening with deep neural networks

Recent advances in deep learning for natural images has prompted a surge of interest in applying similar techniques to medical images. Most of the initial attempts focused on replacing the input of a deep convolutional neural network with a medical image, which does not take into consideration the fundamental differences between these two types of images. Specifically, fine details are necessary for detection in medical images, unlike in natural images where coarse structures matter. This difference makes it inadequate to use the existing network architectures developed for natural images, because they work on an heavily downsampled image to reduce the memory requirements. This hides details necessary to make accurate predictions. Additionally, a single exam in medical imaging often comes with a set of views which must be fused in order to reach a correct conclusion. In our work, we propose to use a multi-view deep convolutional neural network that handles a set of high-resolution medical images. We evaluate it on large-scale mammography-based breast cancer screening (BI-RADS prediction) using 886 thousand images. We focus on investigating the impact of training set size and image size on the prediction accuracy. Our results highlight that performance increases with the size of training set, and that the best performance can only be achieved using the original resolution. This suggests that medical imaging research using deep learning must utilize as much data as possible with the least amount of potentially harmful preprocessing.

How to Optimize Market Strategy using Game Theory and How to Recognize Vehicle Types using Deep Learning

BCG Gamma is an Advanced Analytics division of The Boston Consulting Group, consisting of researchers specializing in Mathematical Modelling, Machine Learning, Operations Research, Big Data Analytics and Software Engineering. In this presentation we will walk you through two of our recent projects: one is an optimization problem solved using game-theoretic approach, the other is a classification problem solved using deep learning. In the first case, we optimized our client's market strategy, taking into account how other players would react to our client's moves. In the second case, we built a system identifying vehicle types using CCTV feeds from a highway operator in order to charge the right toll and notifying of anomalous behavior on the roads. In both cases, we were able to generate multi-million savings per year for our clients.

Charming kernels, colorful Jacobians and Hadamard-minitaurs

Deep mathematical ideas is what drives innovation in machine learning even though it is often underestimated in the era of massive computations. In this talk we show few mathematical ideas that can be applied in many important and unrelated at first glance machine learning problems.We will talk about speeding up algorithms that approximate certain similarity measures used on a regular basis in machine learning via random walks in the space of orthogonal matrices. We show how these can be also used to improve accuracy of several machine learning models, among them some recent RNN-based architectures that already beat state-of-the-art LSTMs. We explain how to "backpropagate through robots" with compressed sensing, Hadamard matrices and strongly-polynomial LP-programming. We will teach robots how to walk and show you that what you were taught in school might be actually wrong - there exist free-lunch theorems and after this lecture you will apply them in practice.

User identification based on keystroke dynamics (workshop)

Behavioral data can be easily accessed and used for augmenting user verification. In this workshop we will introduce and focus on a particular type of behavioral data, and see how we can use machine learning techniques to verify users using such data. We will go through a standard pipeline of: feature extraction, building a model for user verification, and model validation. Participants will be encouraged to play with this pipeline, and to come up with better features and models, but keeping the validation strategy fixed.

The purpose of this workshop is to familiarize the audience with several machine learning libraries, with the process of creating a standard pipeline, and with one of Nethone’s upcoming research projects. We will be using Python, and libraries like: scikit-learn, xgboost, and keras.

The workshop will be hands-on, and will require active participants to have an environment prepared beforehand (either a virtualenv, or a docker image). Instructions for preparing the environment will be available at https://github.com/nethone/keystroke-workshop-env a day before the workshop.

Understanding how Deep Network learn

The reason why deep neural networks can be trained using simple gradient descent continues to defy our understanding. In this talk we will cover the basics as well as some of the most recent research on optimization of deep networks. We will especially focus on the connection of optimization to topics in physics and information theory. I will finish by discussing the most important open problems, as well as some speculation about how training deep networks might look in the near future.

Hierarchical Reinforcement Learning

Reinforcement learning is one of most important parts of machine learning. It is inspired by behavioural biology and economies trying to solve maximisation problems. While the field have witnessed spectacular successes (Atari games, AlphaGo, Dota) it still suffers from many problems, e.g. very poor sample efficiency. In my talk I will present basics of RL and our results of "Hierarchical Reinforcement Learning with Parameters” (CoRL 2017). We experimented in a robotic setup with a manager being able to compose a relatively simple skills (like moving to a point an grasping) to solve more complicated tasks.

Generative Adversarial Networks

Generative adversarial networks (GAN) are a powerful subclass of generative models, mostly known for being able to generate samples of photo-realistic images. In the first part of this talk I will present the main idea behind GAN and give an overview of several popular models.

In the second part I will discuss the problem of evaluating GANs and present a recent large scale study comparing as fairly as possible some of the most popular GAN algorithms (and VAE) on several datasets.

Deep Learning for Developers

In recent months, Deep Learning has become the hottest topic in the IT industry. However, its arcane jargon and its intimidating equations often discourage software developers, who wrongly think that they’re “not smart enough”. We’ll start with an explanation of how Deep Learning works. Then, through code-level demos based on Apache MXNet and Tensorflow, we’ll demonstrate how to build, train and use models based on different network architectures (MLP, CNN, LSTM, GAN). Finally, you will learn about Amazon SageMaker, a new service that lets you train and deploy models into a production-ready hosted environment.

Deep neural networks for speech and natural language processing

Deep neural networks yield state of the art performance in speech recognition and allow for end-to-end training in which of a model's components collaborate to solve the task at hand. I will present end-to-end trainable attention-based recurrent neural networks that directly directly transcribe speech features into sequences of phonemes or characters. The networks learn the alignment between the speech and its transcription and are trained directly to optimize the probability of the correct transcription. I will show the advantages and challenges, such a as language model integration, related to successful application of this family of neural networks. I will conclude the talk with a review of other applications of attention-based recurrent networks in NLP, such as parsing. And with other uses of neural networks in speech processing, such as voice conversion and style transfer.

Is Artificial Intelligence a threat to musicians? – Music generation techniques

Conference party for all participants of the conference and invited guests

Architectures for big scale machine vision applications (remote lecture)

I will present research that I worked on during my Ph.D. My primary interest lays in the basics of architectures for big scale applications. I will explain the idea behind Inception and what had we change in inception-v3 to have it the best single model on ImageNet 2015. I will present our winning submission to MS COCO detection challenge and how did we adopt the feature extractors for that use case. A few months ago, we have announced the work of automatically updating Google Maps based on Google Street View imagery, where we have used the inception-v3 for the text transcription. I will also cover our latest works on dense prediction problems i.e. instance segmentation through pix2vec pixel embeddings and search for optimal decoder architecture in dense prediction problems.

Deep learning - basics and beyond (workshop)

Deep learning has succeeded at such difficult tasks as driving cars or winning a game of Go and Dota2. It all sounds spectacular... but how do you create a state-of-the-art neural network for an even simpler task - image classification? In this workshop we will try to make every set approachable, from setting up the environment through building first models to tinkering with experiments.

During the hands-on session, you will experiment with an artificial neural network for image classification and learn practical hacks for how to tune the network for your needs, using techniques such as transfer learning and data augmentation. By the end of the workshop you will be able to create and optimize a deep learning project from scratch.

In this workshop we’ll be using PyTorch, a deep learning framework in Python. You will have a chance to understand why it is a tool of choice for machine learning researchers and data scientists. And why Andrej Karpathy’s skin has improved since he started using it!

Siamese architecture

In this talk, I will present an overview of a Siamese neural network architecture. This architecture, as well as its extension to a triplet network, is eagerly used in several machine learning applications that require distance learning, such as image retrieval and face recognition. Although the fundamental idea behind these architectures is fairly straightforward, its successful application often requires some tricks. To discuss them, I will use a few case study examples including a feature descriptor learning method for simultaneous localisation and mapping (SLAM) developed as part of a Google Tango collaboration.

Topics in Reinforcement Learning

I will talk about reinforcement learning with neural networks. As an introduction, I will discuss a spectrum of RL algorithms focusing on actor-critic methods. Then I will attempt to give an overview of questions currently pursued by the community, e.g. exploration, transfer, scaling up etc. Finally I will present two applications of multi-agent RL pursued at OpenAI - competitive robotics and Dota 2.

Details to be announced

Applications of machine learning in traffic optimization

I will be talking about possible applications of machine learning in traffic optimization (and in optimizing some other complex processes). I will describe the process of building traffic metamodels by approximating outcomes of traffic simulations using machine learning algorithms (e.g., deep neural networks) and explain how it may be used in real-time traffic signal control and transport planning (e.g., to find optimal locations and capacities of parkings and charging stations for electric vehicles). I will also tell about possible applications of machine learning in the area of connected and autonomous vehicles, which are expected to revolutionize transportation in the near future.