"We're at an unprecedented point in human history where artificially intelligent machines could soon be making decisions that affect many aspects of our lives. But what if we don't know how they reached their decisions? Would it matter?"
Marianne Lehnis, Technology of Business reporter, BBC.
In this open lecture series, you will get the opportunity to understand what AI, and specifically Deep Neural Networks, can do and how they function. Some of the best researchers at Campus UiO, Gaustadbekkdalen will share their insights, and will be open for questions on these topics. Our goal is to facilitate a broader discussion between academic researchers, industry and public sector practitioners and entrepreneurs.
Prof. Ole Christian Lingjærde, UiO
A talk on the evolution of Neural Networks (and later Deep Neural Networks) with focus on what properties can and which properties cannot be modelled with these methods.
Wednesday Aug 21
Tuesday Aug 27
Dr. Anders Hansen, UiO/Cambridge
Abstract: DL has had unprecedented success and has revolutionised artificial intelligence (AI). However, despite the success, DL methods have a serious achilles heel; they are universally unstable and thus demonstrate highly non-human-like behaviour. This has serious consequences, and Science recently published a paper warning about the potentially fatal consequences. The question is: why do AI algorithms based on deep learning become unstable and perform so different to humans? Current mathematical theory on neural networks cannot explain this. We will demonstrate the reason for this discrepancy: neural networks do not learn the structures that humans learn, but completely different structures. These different (false) structures correlate well with the original structure that humans learn, hence the success, however they are completely unstable, yielding non-human performance.
Dr. Anders Hansen, UiO/Cambridge
CS transformed medical imaging through the 2017 approval by the US Food and Drug Administration (FDA) of CS techniques in Magnetic Resonance Imaging (MRI), resulting in widespread use (scanners at the UiO hospitals are now run with CS). However, the success of deep learning (DL) has sparked a vast interest in the question: can DL can outperform CS in image reconstruction? Indeed, in 2018 Nature published the paper "Image reconstruction by domain transform manifold learning" representing the tip of the iceberg of DL methods promising improved performance and "... observed superior immunity to noise…" Despite the promise, DL becomes, as in the classification problem, completely unstable also for image reconstruction (however, for completely different reasons). The phenomenon can be understood by linking CS and DL. In fact, CS may be viewed as a constructive way, without any learning, to build stable neural networks for image reconstruction. The question then becomes: why do trained neural networks based on DL become unstable yet the constructed (untrained) networks based on CS remain stable?
Thursday Aug 29
Tuesday Sept 3
Gudbrand Eggen, Data Scientist and Lab Lead at StartupLab DNN lectures.
An “invisible hand” of algorithms is influencing us ever more. Some algorithms are using our personal data while other algorithms are using the data of people similar to us. How can current and future regulation help us understand and control how we are influenced? Will such regulation let the most unscrupulous actors dominate the field? Is it always possible to give an explanation of why an algorithm made a decision which makes sense to the person affected by the decision? Can such explanations be verified? Might such explanations reveal trade secrets? Also, will we be able to predict whether actions we take will cause an algorithm to make a different decision about us in the future?
Prof. Geir Storvik, UiO
Use of neural network, both deep and shallow, require learning a large number of parameters (weights) from observed data. Due to limited amounts of data there is a danger of overfitting. Further, the amount of uncertainty in the estimated parameters can have huge effect on the prediction performance. In this seminar we will discuss the Bayesian approach to learning and fitting neural networks. Both benefits of such approaches and conceptual as well as computational challenges will be discussed.
Tuesday Sep 10
Tuesday Sep 17
Dr. Arnt-Børre Salberg, Norwegian Computing Centre
In this lecture we aim to understand and explain the behavior of a deep convolutional neural network.
By analyzing the response of each layer we gain knowledge how the CNN is working, and what triggers a given neuron. We also perform a “sensitivity analysis” in order to explain which features or pixels in the input image that is important for the corresponding network output (prediction), and which pixels are the network sensitive to.
Dr. Anders Løland., Norwegian Computing Centre
Many machine learning methods are more or less black boxes for the end user. Even those who develop the machine learning methods can struggle with explaining how methods actually work, or which variables that are most decisive for a specific prediction. The latter is a GDPR requirement for automated individual decision-making. We will give an overview of current methods for opening black boxes. Some of them are useful, while others need more research to be more useful than harmful.
Tuesday Sep 24
Tuesday Oct 8
Morten Hjorth-Jensen, Michigan State University, USA and University of Oslo
Machine Learning (ML) is one of the most exciting and dynamic areas of modern research and application. Here I will show how we can use physics inspired deep learning algorithms like Boltzmann machines to study study problems in quantum many body physics, quantum computing, and chemical and material physics
Dr. Asbjørn Berge, SINTEF
To quote Geoffrey Hinton :"The pooling operation used in convolutional neural networks is a big mistake and the fact that it works so well is a disaster." Pooling is a necessity in most large convolutional nets, but explicitly break hierarchical relationships in the data, or at least makes learning of the underlying representations very inefficient. At the same time, almost all interesting problems have clear hierarchies or geometric relationships. One approach to efficiently learn hierarchical correspondences is Capsule Networks where neurons are extended to output a tensor instead of a scalar. This has several benefits, for example the ability of each (extended) neuron to predict geometric transforms. Theory, use cases, training and extensions will be discussed.
Tuesday Oct 15
Tuesday Oct 22
Signe Moe og Signe Riemer-Sørensen,, SINTEF
How can we use machine learning to infer physical properties of a system and how can physical properties be built into DNNs? What are the challenges for industrial application of DNNs? We will discuss the concepts of hybrid models and the importance of domain knowledge when implementing DNNs.
In this lecture we aim to take multiple, simultaneous shots at DNNs to understand what goes on inside the boxes. We will aim to get researchers from different methodological angles; statistics, computational science, information theory to spar in real time and try to make sense of when and how DNNs work and why.
Tuesday Oct 29
Tuesday Nov 5
Every new area of research comes with its own unique sociology, heroes and villains and tribe dynamics. In this talk we dive into the 'human side' of the DNN evolution. We try to establish a historical timeline of claims and delivered results, and home in on the beliefs on further evolution of DNNs and AI.
The Lecture Series is free of charge. All you have to do is sign up as a participant, and you will receive updates on the schedule.