CS-E4000 - Seminar in Computer Science D, 11.01.2021-30.04.2021

ASA: Use of GSMA consumer RSP protocol in the IoT domain and its security implication

Tutor: Abu Shohel Ahmed

• https://www.gsma.com/esim/wp-content/uploads/2020/06/SGP.22-v2.2.2.pdf

AP: Formal Verification of Cryptographic Protocols

Tutor: Aleksi Peltonen

Formal verification is a group of techniques based on applied mathematics. These methods can be divided into two categories: (1) deductive and (2) model-based verification. Deductive verification involves inferring the correctness of a system specification with axioms and proof rules. Model-based verification, on the other hand, involves using a model checker to create a state model of the system and performing exhaustive state exploration to prove or disprove properties of the protocol. When an error or goal state is reached, the model checker typically provides a trace leading to it from the initial state. Formal verification methods are often used to prove the reliability of commonly used protocols, such as TLS 1.3, and they have been used by companies such as Amazon and Facebook to eliminate bugs in large-scale services. In this topic the student will learn about verification of cryptographic protocols with a state-of-the-art verification tool and demonstrate how it can be used to analyse a cryptographic protocol. The choice of tool and protocol will be agreed upon with the supervisor.

References: 

• http://tamarin-prover.github.io/ 
• https://prosecco.gforge.inria.fr/personal/bblanche/proverif/
•  https://www.mcrl2.org/web/user_manual/index.html

AA: Causal reasoning in reinforcement learning

Tutor: Alexander Aushev

Can reinforcement learning agents learn performing and interpreting the experiments in the environment? Motivated by how human brains explore causal structures in the environment while doing a task, the field of causal reasoning in reinforcement learning tries to adapt this ability for agents. The necessity in causal reasoning naturally arises in biology, operational research, communications and, more generally, in all fields where the environment can be represented as a system of interconnected nodes. Recent developments in the field showed how inference of causal structures result in a more accurate and interpretable performance. For this topic you will review papers related to causal reasoning in reinforcement learning and focus on challenges, applications and state-of-the-art techniques of this field.

References:

• Discovering latent causes in reinforcement learning: https://thesnipermind.com/images/Studies-PDF-Format/GershmanNormanNiv15.pdf
• Gershman, Samuel J. "Reinforcement learning and causal models." The Oxford handbook of causal reasoning. Oxford University Press, 2017. 295 https://books.google.fi/books?hl=en&lr=&id=2qt0DgAAQBAJ&oi=fnd&pg=PA295&dq=causal+inference+in+reinforcement+learning&ots=azhyblbLVV&sig=hM5tQeK0fH8-yrOXpFesGSXRVc0&redir_esc=y#v=onepage&q=causal%20inference%20in%20reinforcement%20learning&f=false

AN: Auto time series forecasting for the regression problems

Tutor: Alexander Nikitin

Modern machine learning tools become more straightforward for users without any background knowledge. AutoML techniques make it possible to automatically construct the model (choose the model type, architecture, hyperparameters). The goal of this project is to review existing methods for time series regression problems and propose a way to automatically construct solutions for any particular dataset. Prerequisite: Basic understanding of machine learning and mathematics, good knowledge of Python (or Julia, or R, or MatLab).

References:

• https://ai.googleblog.com/2020/12/using-automl-for-time-series-forecasting.html
• https://www.usenix.org/conference/opml20/presentation/huang 
•  https://towardsdatascience.com/time-series-forecasting-neuralprophet-vs-automl-fa4dfb2c3a9e

BL1: Adversarial machine learning defenses

Turtor: Blerta Lindqvist

Neural network classifiers are susceptible to attacks that cause misclassification. Many of the proposed defenses have been disputed, leaving only few standing.

References:

• https://nicholas.carlini.com/writing/2019/all-adversarial-example-papers.html

CY: Uncertainty Quantification in Neural ODE Models

Tutor: Cagatay Yildiz

Recently proposed neural ordinary differential equation (NODE) models have been revolutionary in continuous time modeling. The idea is to approximate unknown time differentials, or drift functions, via neural networks, which leads to computing the (time) integral of a neural network. This method allows learning any continuous-time phenomena, such as walking/running sequences and many physical systems involving differential equations. One aspect of continuous-time modelling that is not well-investigated is uncertainty quantification. In ODE2VAE paper (see below), we used Bayesian neural networks (BNN) to handle uncertainty over the unknown time differential function. Possible other alternatives could be implicit BNNs, deep ensembles, functional BNNs, etc. In this project, we would like to investigate whether these methods work in practice and which is more advantageous and why.

References:

• Neural ordinary differential equations: https://arxiv.org/pdf/1806.07366.pdf 
•  ODE2VAE: Deep generative second order ODEs with Bayesian neural networks: https://arxiv.org/pdf/1905.10994.pdf 
•  Deep Ensembles: https://papers.nips.cc/paper/2017/file/9ef2ed4b7fd2c810847ffa5fa85bce38-Paper.pdf

GI: Foveated Video Quality Metrics

Tutor: Gazi Illahi

Foveated video encoding (FVE) refers to video encoding where spatial quality of each video frame corresponds to visual acuity of the Human Visual System (HVS) which is non-uniform. This type of encoding requires the encoder to know (detected or predicted) gaze location of the viewer on the video frame. The motivation for FVE is the theoretical ability to reduce size of a video frame without reducing the perceptual quality of the video frame. The best way to measure the quality of an encoded video is through subjective tests, however, subjective studies are costly and time consuming. It is common to use objective computational metrics to measure video quality. Such methods include simple error metrics like mean square error PSNR and more complex methods which take into account the HVS, like SSIM, VMAF etc. However, there are few universal objective quality metrics for Foveally encoded video. The student's task would be to do a literature review of all objective video quality metrics specifically designed to measure quality of Foveally encoded video.

• 1.Lee, Sanghoon, Marios S. Pattichis, and Alan C. Bovik. "Foveated video quality assessment." IEEE Transactions on Multimedia 4, no. 1 (2002): 129-132. 
• 2. Jin, Yize, et al. "Study of 2D foveated video quality in virtual reality." Applications of Digital Image Processing XLIII. Vol. 11510. International Society for Optics and Photonics, 2020. 
• 3. S. Rimac-Drlje, G. Martinović and B. Zovko-Cihlar, "Foveation-based content Adaptive Structural Similarity index," 2011 18th International Conference on Systems, Signals and Image Processing, Sarajevo, 2011, pp. 1-4.

HD: TinyML as-a-Service - bringing Machine Learning inference to the deepest IoT edge

Tutor: Hiroshi Doyu

TinyML, as a concept, concerns the running of ML inference on Ultra Low-Power (ULP ~1mW) microcontrollers found on IoT devices. Yet today, various challenges still limit the effective execution of TinyML in the embedded IoT world. As both a concept and community, it is still under development. Here at Ericsson, the focus of our TinyML as-a-Service (TinyMLaaS) activity is to democratize TinyML, enabling manufacturers to start their AI businesses using TinyML, which runs on 32 bit microcontrollers. Our goal is to make the execution of ML tasks possible and easy in a specific class of devices. These devices are characterized by very constrained hardware and software resources such as sensor and actuator nodes based on these microcontrollers. The core componet of TinyMLaaS is Machine Learning Compiler (ML compiler)

References:

• https://youtu.be/m2sHB4DOfMg https://youtu.be/yqO6bl8rBEY https://www.mindmeister.com/1637137991?t=QfRudlGYBy
• https://osseu19.sched.com/event/TLCJ https://static.sched.com/hosted_files/osseu19/f9/elc2019-tinymlaas.pdf
• https://www.ericsson.com/en/blog/2019/12/tinyml-as-a-service-iot-edge https://www.ericsson.com/en/blog/2019/12/tinyml-as-a-service
• https://www.ericsson.com/en/blog/2020/2/how-can-we-democratize-machine-learning-iot-devices

JH2: Simulating 5G for Distributed System Communications

Tutor:  Jaakko Harjuhahto

The task is to perform a literature study to review approaches for simulating 5G connectivity between two computers, from the point-of-view of these computers. If one computer wants to send a number of bytes via 5G to another computer connected to the same base station, how long will this take depending base station configuration, signal strength, overall network load etc? The focus is on system level or end-to-end simulation: how 5G behaves from a user's perspective. An example use case for this type of use is [1].

Practical simulators: 

•  NS-3, https://www.nsnam.org/ 
•  OmNet++, https://omnetpp.org/ 
•  Vienna 5G Link Level Simulator, [2] 
• iFogSim, [3] 

 References:

•  [1] Harjuhahto, Debner, Hirvisalo. Processing LiDAR Data from a Virtual Logistics Space. Fog-Iot 2020. https://doi.org/10.4230/OASIcs.Fog-IoT.2020.4
•  [2] Pratschner, S., Tahir, B., Marijanovic, L. et al. Versatile mobile communications simulation: the Vienna 5G Link Level Simulator. https://doi.org/10.1186/s13638-018-1239-6 
• [3] Gupta, Dastjerdi, Ghosh, Buyya. iFogSim: A toolkit for modeling and simulation of resource management techniques in the Internet of Things, Edge and Fog computing environments. https://doi.org/10.1002/spe.2509

JR4: Reduction of relational calculus queries to relational algebra and SQL

Tutor: Jussi Rintanen

The goal of the work is to develop a reduction from a general class of relational calculus queries to relational algebra and its extensions as implemented in the SQL language. Relational calculus is often a preferred way of expressing complex queries, but does have widely available query methods in existing database systems. The work investigates the possibilities of reducing general relational queries to SQL. Queries with limited quantification are reducible to basic relational algebra, but more complex features, including universal quantification, seem to require features like aggregation as present in SQL.

References:

• Equivalence of relational algebra and relational calculus query languages having aggregate functions A Klug - Journal of the ACM (JACM), 1982

LA: Learning to communicate optimally

Tutor:  Laia Amoros

Reliability and security in communication systems is traditionally treated using tools from information theory and coding theory, where mathematical formulas are derived for a few rigid communication models. A novel approach to this setup is to use different machine learning techniques to optimize reliability and security without previously assuming any channel characteristics.

References: 

• Several directions can be taken depending on the student's interests. 
•  https://wisec2020.ins.jku.at/proceedings-wiseml/wiseml20-96.pdf 
•  https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8890904

NW1: ABC (Approximate Bayesian Computation) methods for Cognitive Neuroscience

 Tutor:  Nitin Williams

Recent modelling efforts in Cognitive Neuroscience have furnished insight into mechanisms generating human Neuroscience data. However, fitting these models to experimental human Neuroscience data has been hampered by the intractability of the model likelihood functions. Likelihood-Free Inference (LFI) techniques from Approximate Bayesian Computation (ABC) have shown promise in addressing these issues. However, several variants of ABC have been proposed in the literature, making it difficult to choose which one to deploy. In this project, you will review the different variants of ABC proposed in the literature and weigh their pros and cons. You will also develop recommendations on which are most suitable for fitting models in Cognitive Neuroscience in order to gain insight on mechanisms producing the data. Finally, particularly promising ABC techniques could be compared using simulations. You will be part of the Probabilistic Machine Learning group. Prerequisites The project would be ideal for someone with knowledge of Probability and Machine Learning and an interest in Cognitive Neuroscience. Python programming experience would be useful but not essential.

References:

• Hadida et al. (2018) "Bayesian optimization of large-scale biophysical networks" NeuroImage 174: 219-236

SDP1: Tractable Deep Density Estimation with SPNs

Tutor: Sebastiaan De Peuter

Sum-product networks (SPNs) [1] are a class of deep probabilistic models for modeling joint densities. SPNs are tree-like computational structures containing sum nodes and product nodes where values flow from children to their parents. SPNs can be interpreted as latent variable models or as very large mixtures of factorization of the density they model. What makes SPNs interesting is that they allow you to tractably calculate any marginal and conditional of the joint density they model, which can be used to for example learn to complete images. Furthermore they naturally support a mixture of discrete and continuous variables. Parameters can be determined by gradient-based methods but learning the right structure for an SPN is still an area of active research (see [2] for a recent example). The goal for this project is for you to do a survey of the SPN literature, paying special attention to what novel applications SPNs have enabled. For the practically inclined this project can also involve a novel application or a comparison of SPN inference methods. The project will be carried out in the PML research group.

References: 

• [1] https://ieeexplore.ieee.org/abstract/document/6130310
• [2] http://papers.nips.cc/paper/8864-bayesian-learning-of-sum-product-network

SDP2: Contextual Bandit Algorithms

Tutor: Sebastiaan De Peuter

Bandit problems are stateless decision problems where you are given a set of options (arms) to choose from but are not told how good each option is. In every iteration you choose an option and then receive a reward for that option. Over time you can use these rewards to estimate the value of every option. The goal in bandit problems is to maximize the sum of rewards over time, this requires trading off exploration (trying out options to observe their reward) and exploitation (using the best known option currently to get guaranteed high reward). This project will be about contextual bandits, which is a variant of the original problem where at every iteration you are given a "context", some additional information to make your choice, which also determines the reward of the different options. An example of a contextual bandit problem is an assistance problem where at every iteration you have to help a user whose type is given by the context. Different types of users require different kinds of assistance so the value of your assistance options depends on the user type. In this project you will survey contextual bandit algorithms and apply them to a real problem. The project is flexible. If you don't currently have a good background in RL or bandit algorithms you can do a literature survey, supplemented with an application if time permits. If you have more background or are a quick learner then we have a real project related to the example from above which you can tackle. The project will be carried out in the PML research group.

References:

• https://tor-lattimore.com/downloads/book/book.pdf

SJ: Benchmarking MIMIC-IV medical code prediction with NLP models

Tutor: Shaoxiong JI

Automatic medical code assignment is a routine healthcare task for medical information management and clinical decision support. The International Classification of Diseases (ICD) coding system, maintained by the World Health Organization (WHO), is widely used among various coding systems. Thus, the medical code assignment task is also called ICD coding. It uses clinical notes of discharge summaries to predict medical codes in a supervised manner with human-annotated codes, which is formulated as a multi-class multi-label text classification problem in the medical domain. This project conducts a benchmarking study on the new MIMIC-IV dataset with NLP and deep learning, specifically, neural attention models and multitask learning.

References: 

• https://www.nature.com/articles/s41597-019-0103-9 https://mimic-iv.mit.edu

TG: Adversarial examples for large pre-trained NLP models

Tutor:  Tommi Gröndahl

Large pre-trained neural networks have become the state-of-the-art in machine learning, particularly in natural language processing (NLP). The most prevalent of such pre-trained NLP models are BERT [1] and its derivatives, and GPT(-2/3) [2]. While these models show good performance in e.g. language classification and generation, they have been demonstrated to be vulnerable to adversarial attacks [3, 4, 5]. The aim of this project is to survey existing research on constructing adversarial examples for large pre-trained NLP models, and analyze some of the possible reasons for the success of such attacks.

References: 

• [1] https://arxiv.org/abs/1810.04805 
• [2] https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf 
• [3] https://arxiv.org/pdf/2004.09984.pdf
•  [4] https://arxiv.org/abs/2010.05648
•  [5] https://arxiv.org/abs/1908.07125

NW2: LFI (Likelihood-Free Inference) techniques for Cognitive Neuroscience