CS-E4000 - Seminar in Computer Science D, 07.09.2020-19.12.2020
Kurssiasetusten perusteella kurssi on päättynyt 19.12.2020 Etsi kursseja: CS-E4000
Available topics
AA: Inverse reinforcement learning for user modelling
Tutor: Alex Aushev
In most machine learning applications, the data is either too expensive or hard to collect; such is the case of the human-AI interaction. For example, the human-user may rate dozens of movies before he/she gets tired. But what if we had a simulator that would imitate the human-user's decision process? This way we would get as much data as we want, and the accuracy of our AI algorithm would be only limited by the precision of our user representation. Building such a representation is called user modelling, and recreating the user’s decision process from data is called inverse reinforcement learning. In this topic the student will review recent developments and trends in the inverse reinforcement learning and its application for user modelling.
References:
- A Survey of Inverse Reinforcement Learning: Challenges, Methods and Progress: https://arxiv.org/pdf/1806.06877.pdf
- Application examples:
- [1] USER MODELING FOR TASK ORIENTED DIALOGUES: https://arxiv.org/pdf/1811.04369.pdf
- [2] Exploring Dynamics and Semantics of User Interests forUser Modeling on Twitter for Link Recommendations: http://www.johnbreslin.com/files/publications/20160913_sema2016.pdf
AB1: Usability and Security Tradeoffs in QR Code usage
Tutor: Amel Bourdoucen
Quick Response Codes (QR Codes) are 2-dimensional barcodes that encode bits of information represented as black square dots on a white square grid. Recent integrations into users’ devices allowed for QR codes to be directly scanned using the device’s camera. This was a game-changing strategy which had a heavy impact on QR code usage. Businesses incorporated QR codes into their marketing strategies as well as universities, hospitals, touristic destinations and so on. Because QR Codes are a way to transmit information, some security concerns are expected with the increase of popularity of its usage. In addition to that, could imposing more security regulations create inefficiency in usability and become a burden to the user?
References:
- Hayashi, Eiji & Pendleton, Bryan & Ozenc, Fatih & Hong, Jason. (2012). WebTicket: account management using printable tokens. https://doi.org/10.1145/2207676.2208545
- Krombholz, Katharina & Fruehwirt, Peter & Kieseberg, Peter & Kapsalis, Ioannis & Huber, Markus & Weippl, Edgar. (2014). QR Code Security: A Survey of Attacks and Challenges for Usable Security. 79-90. https://doi.org/10.1007/978-3-319-07620-1_8
AB2: Usability analysis of devices in an IoT ecosystem
Tutor: Amel Bourdoucen
The Internet of Things (IoT) ecosystem has included technologies that aid users, businesses and governments thus creating more opportunities to gain value. IoT ecosystems can be equipped with sensors, washing machines, fridges and other smart devices. In this topic, we focus on the user’s acceptance of IoT in a smart home context. Students will gain insight into the main security risks in an IoT ecosystem, ways to mitigate them and further analysis into the user experience to be able to use a secure efficient system. No prior background in IoT is required.
References:
- Latvala, S., Sethi, M. & Aura, T. Evaluation of Out-of-Band Channels for IoT Security. SN COMPUT. SCI. 1, 18 (2020). https://doi.org/10.1007/s42979-019-0018-8
AD1: LiDAR processing with Deep Neural Networks
Tutor: Anton Debner
LiDAR point cloud data is a set of 3D coordinates, created by, for example, a rotating LiDAR scanner on top of an autonomous vehicle. A point cloud can provide a rough 3D representation of the surrounding environment, or perhaps an accurate model of a single object. In addition to the XYZ coordinates, the point cloud may contain additional information such as the intensity of the reflected signal (depends on, e.g., incident angle, surface material) or RGB color values (fusion of RGB camera and LiDAR data).
How is LiDAR data processed? What is the role of additional information, such as color and signal intensity? What are recent advances in LiDAR processing and sensor fusion? In addition to autonomous vehicles, what kind of modern applications are there for LiDARs?
This study is mainly a literature review, but it is also possible to test ML models on publicly available LiDAR data based on the literature review.
References:
- SemanticKITTI dataset: http://semantic-kitti.org/
- PointNet (2017): https://arxiv.org/abs/1711.09869
- SalsaNext (2020): https://arxiv.org/abs/2003.03653v3
- Deep Learning for 3D Point Clouds: A Survey: http://dx.doi.org/10.1109/TPAMI.2020.3005434
AD2: Deep Reinforcement Learning and Games
Tutor: Anton Debner
Deep reinforcement learning (DRL) has been advancing quickly along other neural network research. For example, in 2015 DeepMind’s AlphaGo beat professional players in Go. In 2019 DeepMind’s AlphaStar defeated top tier players in StarCraft II, a significantly more complex game due to its real-time nature and vast action space.
There are also several publicly available environments and frameworks with the aim of allowing anyone to try and research reinforcement learning methods. These include robotic arms, multi-legged creatures, classic 2D Atari games, early 3D games (Doom) and modern competitive games. While games are entertaining tools to research on, they are often seen as a platform for creating generalizable techniques to solve problems from other areas as well.
The task is to perform a literature review on recent papers, focusing either on how the game-related research generalizes to other non-gaming related domains or how DRL can be used in game development (e.g., game AI, procedural content generation, game design).
References:
- AlphaGo: https://doi.org/10.1038/nature16961
- AlphaStar: https://doi.org/10.1038/s41586-019-1724-z
- OpenAI Gym: https://gym.openai.com/
- https://arxiv.org/pdf/1909.07528.pdf
- https://arxiv.org/pdf/1912.10944.pdf
AJ1: Interpretable latent space for Variational Auto-Encoders
Tutor: Anirudh Jain
Learning interpretable factors within the latent space of a variational auto-encoder(VAE) is a promising direction for deep generative models. Disentangled latent space allow us to learn factors for meaningful features of the data. For example, learning factors for thickness, digit label etc for the MNIST dataset. Interpretability is also useful for generation of new data by letting us explore the latent space efficiently in meaningful directions. In this project, you will perform a literature survey on existing techniques for interpretable latent space for VAEs and if allocated time allows, implement one of the SOTA methods.
Prerequisites: Basic knowledge of deep learning and mathematics (course work or self-taught), familiarity with Python and any deep learning framework of choice
References:
- http://proceedings.mlr.press/v97/mathieu19a.html
AJ2: Normalizing flows for graph structured data generation
Tutor: Anirudh Jain
Graph structured data provides useful representation for a lot of domains such as social networks, bioinformatics or robotics. Generative models allow us to learn a latent space from the training data which can be explored for data generation such as for drug discovery with molecular graphs. Graph structure imposes additional challenges due to arbitrary shape and size of graphs and discrete structure. Normalizing flows present a promising direction to map a base distribution to a discrete graph distribution. In this project, you will review the SOTA methods for flow based generative methods and adapt them for graph generation.
Prerequisites: Basic knowledge of deep learning and mathematics (course work or self-taught), familiarity with Python and any deep learning framework of choice
References:
- https://arxiv.org/abs/2001.09382
AJ3: Ethical Explainable Federated Machine Learning for High-Precision Management of Pandemics with Big Data over Networks
Tutor: Alexander Jung
Develop federated machine learning methods that allow to compute personalized recommendations for behaviour during pandemic (e.g. stay home, or avoid gym, or avoid school).
References:
http://federated-ml.org/tutorials/icassp2020/part2.pdf
AJ4: Statistical Analysis of Clustered Federated Learning
Tutor: Alexander Jung
Use our recent theory on networked linear models to obtain conditions that ensure accuracy of clustered federated learning.
References:
http://federated-ml.org/tutorials/icassp2020/part2.pdf
AM1: A review of methodological and coding approaches for smoking-related Instagram data
Tutor: Aqdas Malik
Researchers are increasingly using social media data for public health surveillance. In recent years, novel smoking products such as electronic-cigarettes, water pipes, shisha etc. have gained wide popularity and social acceptance among highly vulnerable adolescents and young adults’ cohort. Likewise, youth are quite active on different social media platforms, in particular Instagram, where they share and view various forms of imagery associated with tobacco usage and marketing. The current project aims to synthesize the current literature that utilizes Instagram data for classifying and categorizing smoking-related pictures/videos/other data. More specifically, what are the type of data collection and coding methods used by these studies? Furthermore, what are the different categories that have been coded/classified as the outcome of these studies?
References:
Ketonen, V., & Malik, A. (2020). Characterizing vaping posts on Instagram by using unsupervised machine learning. International Journal of Medical Informatics, 141, 104223.
- Majmundar, A., Kirkpatrick, M., Cruz, T. B., Unger, J. B., & Allem, J. P. (2020). Characterising KandyPens-related posts to Instagram: implications for nicotine and cannabis use. Tobacco control, 29(4), 472-474.
- Lee, A. S., Hart, J. L., Sears, C. G., Walker, K. L., Siu, A., & Smith, C. (2017). A picture is worth a thousand words: electronic cigarette content on Instagram and Pinterest. Tobacco prevention & cessation, 3.
AM2: Characterizing ADHD conversations on Twitter
Tutor: Aqdas Malik
Social media such as Twitter facilitate users to express health-related opinions and share pertinent information with others. Exploring and understanding large-scale commentary on different health topics provide a unique opportunity to characterize and analyze this data that can provision public health experts, clinical practitioners, as well as social scientists. ADHD (Attention deficit hyperactivity disorder) is one of the commonly diagnosed mental disorder that affects children and teenagers and can further linger into adulthood. The current project aims to collect and analyze Twitter data on #ADHD by applying different techniques such a topic modelling to characterize relevant conversations. Deeper analysis of ADHD-related dialogues on Twitter will not only help us understand the public opinion and sentiment, but can also effectively support in crafting suitable interventions by relevant entities.
References:
Karami, A., Dahl, A. A., Turner-McGrievy, G., Kharrazi, H., & Shaw Jr, G. (2018). Characterizing diabetes, diet, exercise, and obesity comments on Twitter. International Journal of Information Management, 38(1), 1-6.
- Fu, K. W., Liang, H., Saroha, N., Tse, Z. T. H., Ip, P., & Fung, I. C. H. (2016). How people react to Zika virus outbreaks on Twitter? A computational content analysis. American journal of infection control, 44(12), 1700-1702.
- Sinnenberg, L., DiSilvestro, C. L., Mancheno, C., Dailey, K., Tufts, C., Buttenheim, A. M., ... & Asch, D. A. (2016). Twitter as a potential data source for cardiovascular disease research. JAMA cardiology, 1(9), 1032-1036.
BGA: Robustness and Privacy in Federated Learning
Tutor: Buse Gul Atli tekgul
Federated learning is a distributed machine learning setting, where training is done on edge devices owned by clients and coordinated via a central server or a service provider [1,2]. Federated learning allows clients to store their own data locally; therefore, it does not compromise the privacy of clients’ data. Today companies like NVIDIA and Google has federated learning settings and clients (e.g., owner of cell-phone devices) both participate the training and get personalised apps without worrying about possible leakage of their sensitive datasets.
Recent work has shown that, since the model is downloaded by clients participating to training, they can implement different adversarial attacks against federated learning models [3]. There have been attacks against the integrity: Poisoning attacks [5] aim to manipulate the model parameters such that its performance degrades completely or on some specific samples. Another well-known attack type is attacks against confidentiality: Malicious clients (or a malicious server) exploit the vulnerabilities of the learning algorithm in order to get information about other client's data [4]. In this seminar topic, the student(s) are expected to do a survey of attacks and defences related to the robustness and privacy of the federated learning.
Note: Two students can divide the topic but we can discuss the progress together. One student can focus on the privacy analysis aspect and other student can work on poisoning attacks (and possible defenses) against federated learning models.
References:
- [1] Bonawitz, Keith, et al. "Towards federated learning at scale: System design." arXiv preprint arXiv:1902.01046 (2019).
- [2] McMahan, Brendan, et al. "Communication-efficient learning of deep networks from decentralized data." Artificial Intelligence and Statistics. PMLR, 2017.
- [3] Kairouz, Peter, et al. "Advances and open problems in federated learning." arXiv preprint arXiv:1912.04977 (2019).
- [4] Nasr, Milad, Reza Shokri, and Amir Houmansadr. "Comprehensive privacy analysis of deep learning: Passive and active white-box inference attacks against centralized and federated learning." 2019 IEEE Symposium on Security and Privacy (SP). IEEE, 2019.
- [5] Fang, Minghong, et al. "Local model poisoning attacks to Byzantine-robust federated learning." 29th {USENIX} Security Symposium ({USENIX} Security 20). 2020.
BL1: Attacks in adversarial machine learning
Tutor: Blerta Lindqvist
Adversarial attacks are imperceptibly perturbed samples that cause classifiers to misclassify them. What types of attacks are there? What do they have in common, how do they differ?
References:
https://arxiv.org/abs/1412.6572
https://arxiv.org/pdf/1511.04599.pdf
- https://arxiv.org/abs/1608.04644
BL2: Defenses in adversarial machine learning
Tutor: Blerta Lindqvist
Adversarial attacks are imperceptibly perturbed samples that cause classifiers to misclassify them. How can we defend against adversarial samples? What are some of the existing methods, and could we extend them?
References:
https://arxiv.org/abs/1611.01236
https://arxiv.org/abs/1705.09064
- https://arxiv.org/abs/1704.03976
- https://arxiv.org/abs/1705.07204
ER: Semantic description of policies for intelligent services in distributed environments
Tutor: Edgar Ramos
Policies targeting how a device can be used or who is allowed to do what with the device and what kind of operations are allowed or should be enforced with the data are needed to manage IoT devices that run intelligence services. A high-level description of these policies can be provided to devices and translated to concrete policies that the device can understand (for example, read, write, id requirements, etc). One approach in this direction is the use of semantics descriptions to provide such policies using for example semantic constructions with a descriptive language. The IoT applications are inherently heterogeneous and their capacity to interpreter high-level policies is limited to their domain. Therefore domain of application should be also taken care in the definition and interpretation of such policies.
References:
- An article on policy semantics language description:
- Kagal L., Finin T., Joshi A. (2003) A Policy Based Approach to Security for the Semantic Web. In: Fensel D., Sycara K., Mylopoulos J. (eds) The Semantic Web - ISWC 2003. ISWC 2003. Lecture Notes in Computer Science, vol 2870. Springer, Berlin, Heidelberg. https://doi.org/10.1007/978-3-540-39718-2_26
- Policy language from AWS :
- https://docs.aws.amazon.com/IAM/latest/UserGuide/iam-ug.pdf#reference_policies_grammar
EV1: Cloud Gaming and Local Game Streaming
Tutor: Esa Vikberg
Cloud Gaming is a growing market, with an increasing number of companies (e.g. Google, Microsoft, Sony, Nvidia) promoting their own services for streaming on-demand games to consumers. Their development is supported by the increasing prevalence of high speed low-latency networking such as fiber optic internet and 5G. The advances in game streaming also allows streaming games from a local server, either to different screens in a house (Nvidia SHIELD, Steam Link) or to gaming partners elsewhere (Steam Remote Play). There is room to focus the project on what interests you the most. You can do a review of the available products, focusing on the technical aspects of the services. It is also possible to do a measurement study on things such as latency and the effects of computing resources.
References:
- https://newzoo.com/insights/articles/cloud-gaming-business-market-revenues-and-ecosystem/
- https://ieeexplore.ieee.org/document/6574660
- https://ieeexplore.ieee.org/document/6818918
- https://aaltodoc.aalto.fi/handle/123456789/19492
EV2: Multi-User XR
Tutor: Esa Vikberg
Extended Reality (XR) includes the range of technologies merging virtual and real environments from augmented reality to virtual reality. What problems are there with providing it to multiple users? How can these be overcome? You can look at the whole range of applications from Pokemon Go to VR applications with multiple users in the same room.
References:
- https://ieeexplore.ieee.org/abstract/document/8567653
- https://dl.acm.org/doi/abs/10.1145/3345554
EV3: Light Fields
Tutor: Esa Vikberg
Light fields describe the amount of light flowing in every direction through every point in space. Capturing a subset of such directions and points allows for changing viewing perspectives after capturing the state of the light field. In the case of a rendered model or a photograph, this allows the viewer to look from different directions and see the perspective change accordingly. In this project, you review the state of the technology, and discuss the current and potential use cases.
References:
- https://www.youtube.com/watch?v=BXdKVisWAco
- https://www.youtube.com/watch?v=-Hc92mP3GLw
- https://www.photonics.com/Articles/Google_to_Present_End-to-End_System_for_Immersive/a65874
- https://store.steampowered.com/app/771310/Welcome_to_Light_Fields/
HD: TinyML as-a-Service - bringing ML to the very deepest 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 8, 16 and 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.
Below, we present how we can bind the “as-a-service” model to TinyML. We will provide a high-level technical overview of our concept and introduce the design requirements and building blocks which characterize this emerging paradigm.
References:
- 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
- https://sched.co/TLCJ
- https://static.sched.com/hosted_files/osseu19/f9/elc2019-tinymlaas.pdf
- https://www.mindmeister.com/1579287016?t=XIsVONhs4x
- Also:
- https://protect2.fireeye.com/v1/url?k=522f27c7-0c8fc759-522f675c-86ee86bd5107-530a6d3a834fe74e&q=1&e=70c8edd5-6263-463e-bbaa-63eb114b722c&u=https%3A%2F%2Fdocs.google.com%2Fdocument%2Fd%2F165lfe1dmgB4SYuuE4FkWUdyVDUAzsIUledMqJfu5tms
KEM: Efficient-communication in federated learning
Tutor: Khaoula El Mekkaoui
Federated learning is an emerging machine learning field that comes to utilize the abundant diversified and naturally disperse data sources, using their resources in a safe way, without moving the data from its source to a datacenter and doing the computations locally. While providing solutions for high latency, privacy, and lower power consumption, it still suffers from challenges for heterogeneity of the data, and the intrinsic expensive communications that come from it. We would like to do a literature survey in the efficient-communication federated learning algorithms up until now and implement at least one of them on a real-life dataset.
References:
- https://icml.cc/virtual/2020/poster/6733
- http://proceedings.mlr.press/v108/reisizadeh20a.html
MS1: Physiologically-augmented VR experience
Tutor: Matti Siekkinen
Virtual reality (VR) is gaining momentum, especially on mobile and wearable devices. Particularly, VR headsets are rapidly evolving. Gaze tracking is becoming a common feature for VR headsets and some devices allow to carry out EEG (Electroencephalography) or EMG (Electromyography) measurements – EEG refers to measuring the electrical activity in the brain, whereas EMG measures the electrical activity in muscles.
The student involved in this project will investigate the applicability of EEG / EMG signals for use in VR. The primary goal is to understand to which extent it is possible to predict user actions (head and eye motion) from those signals before the actual movement happens. This, in turn, could be used to compensate for motion-to-photon latency in a VR system through predictive rendering, particularly in systems where graphics is remotely rendered and latency is a critical concern. The secondary goal is to understand how these signals can be used to model VR user experience, particularly emotional responses, which would be highly useful for analytics and advertising, for instance. The work may include conducting experiments with VR headsets in collaboration with the Aalto Brain Center and the Aalto Neuroimaging Infrastructure.
EEG has been studied extensively for understanding quality of experience (QoE) with multimedia services and human behaviour such as movement intentions. However, an EEG/EMG-instrumented VR headset only has a limited contact surface with the head; therefore, understanding its impact on the quality and amount of information available for predicting user intent and QoE modeling needs to be investigated. Further open questions include an appropriate selection of anticipatory signals and the related processing overhead for latency compensation.
References:
- https://looxidlabs.com/looxidvr/
- https://emteq.net/
- Shu Shi and Cheng-Hsin Hsu. 2015. A Survey of Interactive Remote Rendering Systems. ACM Comput. Surv. 47, 4, Article 57 (May 2015), 29 pages. (https://doi.org/10.1145/2719921)
- A. Moldovan, I. Ghergulescu, S. Weibelzahl and C. H. Muntean, "User-centered EEG-based multimedia quality assessment," 2013 IEEE International Symposium on Broadband Multimedia Systems and Broadcasting (BMSB), London, 2013, pp. 1-8. (https://doi.org/10.1109/BMSB.2013.6621743)
- Bai, Ou, et al. "Exploration of computational methods for classification of movement intention during human voluntary movement from single trial EEG." Clinical Neurophysiology 118.12 (2007): 2637-2655 (https://doi.org/10.1016/j.clinph.2007.08.025)
- Christoph Tremmel, Christian Herff, Tetsuya Sato, Krzysztof Rechowicz, Yusuke Yamani, and Dean J. Krusienski, "Estimating Cognitive Workload in an Interactive Virtual Reality Environment Using EEG", Frontiers in Human Neuroscience 13:401, 2019 (http://doi.org/10.3389/fnhum.2019.00401)
- Jan-Philipp Tauscher, Fabian Wolf Schottky, Steve Grogorick, Paul Maximilian Bittner, Maryam Mustafa, and Marcus Magnor. "Immersive EEG: evaluating electroencephalography in virtual reality." In 2019 IEEE Conference on Virtual Reality and 3D User Interfaces (VR), pp. 1794-1800. IEEE, 2019 (http://doi.org/10.1109/VR.2019.8797858)
MS2: Distributed and parallel rendering for gaming and XR
Tutor: Matti Siekkinen
The topic explores distributing the graphics workload in remote rendering systems for (mobile) gaming (e.g., Stadia, Geforce Now) and XR.
Standalone 5G systems will bring three computing tiers, local device, edge, and cloud allowing scaling the computing within a tier but also across tiers. Tight synchronization constraints, different GPU computing capacities, bandwidth, and latencies between tiers as well as a typically best effort network between computing instances makes the problem difficult.
Support for multi-GPU rendering already exists today in off-the-shelf products [1,2] and modern graphics APIs also have explicit support [3,4]. In addition, academic research has proposed various solutions [5,6]. Remote rendering systems with explicit latency compensation methods may open up the avenue for also for new solutions [7].
The task of the student is to study the different alternatives and produce a taxonomy of the solutions and discuss their suitability for remote rendering systems.
References:
- [1] AMD CrossfireTM Technology. https://www.amd.com/en/technologies/crossfire Library Catalog: www.amd.com.
- [2] SLI | GeForce. https://www.geforce.co.uk/hardware/technology/sli
- [3] 2017. Explicit Multi-GPU with DirectX 12 – Control, Freedom, New Possibilities. https://developer.nvidia.com/explicit-multi-gpu-programming-directx-12 Library Catalog: developer.nvidia.com.
- [4] Khronos Group Releases Vulkan 1.1. https://www.khronos.org/news/press/khronos-group-releases-vulkan-1-1
- [5] Youngsok Kim, Jae-Eon Jo, Hanhwi Jang, Minsoo Rhu, Hanjun Kim, and Jangwoo Kim. 2017. GPUpd: a fast and scalable multi-GPU architecture using cooperative projection and distribution. In Proceedings of the 50th Annual IEEE/ACM International Symposium on Microarchitecture. 574–586.
- [6] S. Eilemann, M. Makhinya and R. Pajarola, "Equalizer: A Scalable Parallel Rendering Framework," in IEEE Transactions on Visualization and Computer Graphics, vol. 15, no. 3, pp. 436-452, May-June 2009, doi: 10.1109/TVCG.2008.104.
- [7] Shu Shi and Cheng-Hsin Hsu. 2015. A Survey of Interactive Remote Rendering Systems. ACM Comput. Surv. 47, 4, Article 57 (July 2015), 29 pages. DOI:https://doi.org/10.1145/2719921
PP: A Scalable Serving System for a Deep Neural Network
Tutor: Phuong Pham
Serving a Deep Neural Network (DNN) efficiently on a cluster of Graphical Processing Unit (GPU) is an important problem. For instance, a cloud-scale video analysis service for thousands of users always requires a fast response of the analyzed DNN. The challenge is to distribute a large incoming workload on a cluster of GPUs with high utilization and acceptable latency. There are many projects targeting this problem such as Clipper, Tensorflow Serving, Nexus, etc.
In this work, students will study these well-known systems and give a comprehensive comparison. It is encouraged that students practice and introduce a demo for these real-world solutions. It would be great if students could introduce potential solutions for improving any of these systems.
References:
- https://stevenwhang.com/tfx_paper.pdf
- https://www.usenix.org/system/files/conference/nsdi17/nsdi17-crankshaw.pdf
- https://pdfs.semanticscholar.org/0c0f/353dbac84311ea4f1485d4a8ac0b0459be8c.pdf
SDP1: Concrete problems in AI safety
Tutor: Sebastiaan De Peuter
The rapid improvement in RL performance over the last two decades has lead to an increase in interest in AI safety. One of the question AI safety tries to answer is how we can make sure that the values of an AI will always be aligned with out own. You can read about some of the concrete problems the field of AI safety tries to tackle in the linked publications. In a recent publication a team at DeepMind has introduced gridworlds which yield unsafe behaviour in some of the RL algorithms we use every day. For this project you will choose one or more of these concrete problems to study. You will identify where current algorithms go wrong, survey proposed mitigations and -- time permitting -- implement one of these mitigations to see how well it works in practice.
References:
- https://arxiv.org/pdf/1606.06565.pdf
- https://arxiv.org/pdf/1711.09883.pdf
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.
References:
- https://tor-lattimore.com/downloads/book/book.pdf
SJ: Natural Language Generation for Mental Health Counseling
Tutor: Shaoxiong Ji
Natural language generation is one of the brightest jewels in the crown of natural language processing.
The recent OpenAI's GPT-3 has gained impressive performance in generating human-like text.
This topic focuses on the mental health domain and aims to generate responses to people expressing their mental issues online.
During some periods like the pandemic, people struggle with mental health issues, and many may not get help from mental health practitioners.
Generative methods can generate supportive responses to help mitigate people's mental health issues.
This project will study some small models (e.g., miniGPT) and knowledge-distilled models to generate higher-quality responses to mental health-related posts instead of cumbersome generative models.
Prerequisites: deep learning courses (e.g., Aalto CS-E4890 or Stanford CS-224N) and PyTorch programming
References:
- https://github.com/karpathy/minGPT
- https://arxiv.org/abs/1910.01108
SS1: Brain-Computer Interfaces and Authentication
Tutor: Sanna Suoranta
Recent development of brain imaging technologies and reduction of the device costs allow new ways to authenticate users. Brain-Computer Interface may use e.g. EEG signals to authenticate users, even though BCI is mainly used to help disabled people in their daily lives and communication. What is the state of the art now in authentication that uses psychometric data?
References:
- Sidas Saulynas,Charles Lechner &Ravi Kuber. Towards the Use of Brain–Computer Interface and Gestural Technologies as a Potential Alternative to PIN Authentication International Journal of Human–Computer Interaction Volume 34, 2018 - Issue 5 https://www.tandfonline.com/doi/full/10.1080/10447318.2017.1357905
- Ghana Al-Hudhud, Mai Adbulaziz Alzamel, Eman Alattas, Areej Alwabil. Using brain signals patterns for biometric identity verification systems. Computers in Human Behavior Volume 31, February 2014, Pages 224-229 https://www.sciencedirect.com/science/article/pii/S0747563213003464
SS2: Educational games for children
Tutor: Sanna Suoranta
Nowadays even small kids play a lot with smart phones and tablet computers. Of course, their parents would like them to play games that teach something, e.g. to read or to code. For example, University of Jyväskylä has developed Ekapeli that helps reading both normal students and students who have learning disabilities (lot of research has done on it, but it may be in Finnish). What need to be considered especially when an educational game is designed for under school-age children or those who has just started their school? Are the games different, when their goal is to teach reading, mathematical skills or coding skills?
References:
- Junnan Yu, Chenke Bai, Ricarose Roque. Considering Parents in Coding Kit Design: Understanding Parents' Perspectives and Roles. CHI '20: Proceedings of the 2020 CHI Conference on Human Factors in Computing SystemsApril 2020 Pages 1–14https://doi.org/10.1145/3313831.3376130
- Nyet Moi Siew, Jolly Geofrey, Bih Ni Lee. Students’ Algebraic Thinking and Attitudes towards Algebra: The Effects of Game-Based Learning using Dragonbox 12 + App. The Research Journal of Mathematics and Technology, Volume 5, Number 1, ISSN 2163-0380.
- Marika Peltonen. Ekapeli ja sen tulosten analysointi. Master's thesis, University of Jyväskylä, 2007. https://jyx.jyu.fi/bitstream/handle/123456789/12510/1/URN_NBN_fi_jyu-2007907.pdf
- Engaging children with educational content via Gamification. Kalpana Nand, Nilufar Baghaei, John Casey, Bashar Barmada, Farhad Mehdipour & Hai-Ning Liang Smart Learning Environments. volume 6, Article number: 6 (2019)
SS3: Games for elderly people
Tutor: Sanna Suoranta
Computer science is nowadays used in healthcare in various ways. One field under research are games that could help elderly people to maintain their functional capacity.
What need to be considered especially when a game is designed for healty elderly people or for those whose already have cognitional or physical problems?
References:
- David A.Raichlen andGene E.Alexander. Adaptive Capacity: An Evolutionary Neuroscience Model Linking Exercise, Cognition, and Brain Health. Trends in Neurosciences. Volume 40, Issue 7, July 2017, Pages 408-421
- Juyoung Sunwoo, Wallace Yuen, Christof Lutteroth, Burkhard Claus Wünsche . Mobile games for elderly healthcare. CHINZ '10: Proceedings of the 11th International Conference of the NZ Chapter of the ACM Special Interest Group on Human-Computer InteractionJuly 2010 Pages 73–76https://doi.org/10.1145/1832838.1832851
- Josef Wiemeyer & Annika Kliem. Serious games in prevention and rehabilitation—a new panacea for elderly people? Eur Rev Aging Phys Act (2012) 9:41–50. DOI 10.1007/s11556-011-0093-x
SS4: Rotation invariant data science from sparse temporal point-cloud data
Tutor: Stephan Sigg
Point cloud data is generated by Lidar, rgb depth cameras or by mmWave Doppler radars. Unlike static point cloud data, many application domains feature temporal point clouds, which can describe objects in motion such as activities, gestures or moving objects.
Moving point clouds constitute non-structured (data-points do not fall on a pre-defined grid) 5D input for data analysis and machine learning tools (x,y,z,time,intensity).
The aim of this seminar topic is to suggest effective approaches to 'consume' such temporal point cloud (possible after appropriate transformation and pre-processing) with a machine learning architecture, while
1) capturing temporal dependencies and
2) achieving rotation-invariance of the model towards point-cloud data
The student shall
a) survey the related work
b) suggest an own approach to the problem
c) report the result of the suggested approach on temporal point cloud data (will be provided).
References:
- Qi, Charles Ruizhongtai, et al. ""Pointnet++: Deep hierarchical feature learning on point sets in a metric space."" Advances in neural information processing systems. 2017.
- Xiao, Zelin, et al. ""Endowing Deep 3d Models With Rotation Invariance Based On Principal Component Analysis."" 2020 IEEE International Conference on Multimedia and Expo (ICME). IEEE, 2020.
- Zhang, Zhiyuan, et al. ""Rotation invariant convolutions for 3D point clouds deep learning."" 2019 International Conference on 3D Vision (3DV). IEEE, 2019.
SS5: Anticipation of Signal drop through environmental sensing
Tutor: Stephan Sigg
In ultra-reliable low latency communication (URLLC) it is essential to proactively, rather than reactively conduct corrective action on the radio link so that the high reliability/low error rate requirement is not compromised.
In particular, in low latency and time sensitive communications, there might be no time for HARQ retransmissions.
In such cases, the radio access network (RAN) operation could be assisted by external sensors, e.g. video, lidar or Radar. Such information could, for instance, predict closing doors or obstacles that are about to block a radio link.
The RAN could then take corrective actions before any errors happen.
The student shall
1) survey the literature around this topic
2) group and structure existing solutions
References:- Charan, Gouranga, Muhammad Alrabeiah, and Ahmed Alkhateeb. "Vision-Aided Dynamic Blockage Prediction for 6G Wireless Communication Networks." arXiv preprint arXiv:2006.09902 (2020).
TC: Quantifying uncertainty in deep learning
Tutor: Tianyu Cui
How to estimate the predictive uncertainty of deep learning models, i.e., provide meaningful confidence values in addition to the label predictions, is drawing more attention. Predictive uncertainty is increasingly being used to make decisions in important applications such as medical decision support, self-driving cares, and financial forecasts. There have been various approaches proposed to quantify uncertainty in deep learning ranging from frequentist approaches to Bayesian methods. In this project, the student will review the current literature about quantifying prediction uncertainty in DL and present comparison studies on real-world datasets.
References:
- https://arxiv.org/abs/2007.13481
- https://openreview.net/forum?id=BJxI5gHKDr
- https://arxiv.org/abs/1906.02530
TM1: Deep Learning based Sparse Linear Solvers
Tutor: Thaha Mohammed
Sparse linear algebra is central to many areas of engineering, science and business. The community has done considerable work on proposing new methods for
SpMV (sparse matrix-vector multiplication) computations and iterative sparse solvers on GPUs [1,2]. Due to vast variations in matrix features, no single method performs well
across all sparse matrices. A few tools on automatic prediction of best-performing SpMV kernels have emerged recently and require many more efforts to fully utilize
their potential. The utilization of a GPU by the existing SpMV kernels is far from its full capacity. Moreover, the development and performance analysis of SpMV techniques
on GPUs have not been studied in sufficient depth. The aim of this work is to study, analyze, and develop novel Deep Learning based techniques to automatically select the best sparse storage format and associated SpMV algorithms on GPUs.
Pre-requisite: Knowledge of various Deep Learning (DL) techniques and a DL framework.
References:
- [1] Salvatore Filippone, Valeria Cardellini, Davide Barbieri, and Alessandro Fanfarillo. 2017. Sparse Matrix-Vector Multiplication on GPGPUs. ACM Trans. Math. Softw. 43, 4, Article 30 (March 2017), 49 pages. DOI:https://doi.org/10.1145/3017994
- [2] Akrem Benatia, Weixing Ji, Yizhuo Wang, and Feng Shi. 2018. BestSF: A Sparse Meta-Format for Optimizing SpMV on GPU. ACM Trans. Archit. Code Optim. 15, 3, Article 29 (October 2018), 27 pages. DOI:https://doi.org/10.1145/3226228
TM2: Distributed Learning on the Edge
Tutor: Thaha Mohammed
The proliferation of resource-constrained mobile sensors and smart devices have led to the generation of a large amount of data. Due to the recent advancements in Deep Learning, AI-based services and applications have experienced have formed a significant enabler for smart cities and factories to intelligent transport systems and much more. Deep Learning (DL)/ Machine Learning (ML) models are often built from the collected data (training), to enable the detection, classification, and prediction of future events (inference). Due to the limited computing resources, these models are often offloaded to powerful computing nodes such as cloud servers. However, it is difficult to satisfy the latency, reliability, and bandwidth constraints as well as the privacy concerns while offloading data to cloud servers for training and inference of AI models. Thus in recent years, to meet this demand, AI services and tasks have been pushed to the network edge to meet the requirements. The study aims to review artificial intelligence in the context of the IoT, mainly techniques employed for the training of DL/ML models at the network edge and to identify the challenges, open problems, and future trends.
References:
- Zhou, Z., Chen, X., Li, E., Zeng, L., Luo, K., & Zhang, J. (2019). Edge Intelligence: Paving the Last Mile of Artificial Intelligence With Edge Computing. Proceedings of the IEEE, 107(8), 1738–1762. DOI: 10.1109/jproc.2019.2918951
- McMahan, B., Moore, E., Ramage, D., Hampson, S. & Arcas, B.A.y.. (2017). Communication-Efficient Learning of Deep Networks from Decentralized Data. Proceedings of the 20th International Conference on Artificial Intelligence and Statistics, in PMLR 54:1273-1282
- Teerapittayanon, S., Mcdanel, B., & Kung, H. (2017). Distributed Deep Neural Networks Over the Cloud, the Edge and End Devices. 2017 IEEE 37th International Conference on Distributed Computing Systems (ICDCS). DOI: 10.1109/icdcs.2017.226
VH: Orchestration Techniques for Fog Computing
Tutor: Vesa Hirvisalo
Fog computing [1] is an emerging paradigm for organizing computing services for Industrial Internet of Things (IIoT). Fog computing extends the concepts and practices of cloud computing toward the rapidly increasing number of connected devices. Many aspects of fog computing is currently under intense research. One of these is orchestration as it is central for cloud computing systems. Currently, Kubernetes [2] serves as an good example container-orchestration. However, the fog computing domains (e.g., [3]) have their inherent requirements and there also other basic options for virtualization than containers [4].
The task to is make an overview of orchestration techniques for Fog computing. The overview can include also experimental work, but it should base its findings on a survey of the techniques.
References:
- [1] Yousefpour & al. All One Needs to Know about Fog Computing and Related Edge Computing Paradigms - A Complete Survey. Journal of Systems Architecture. DOI:10.1016/j.sysarc.2019.02.009
- [2] Kubernetes.io
- [3] Harjuhahto, Debner, Hirvisalo. Processing LiDAR Data from a Virtual Logistics Space. Fog-Iot 2020.
- [4] Morabito, Cozzolino, Ding, Beijar, Ott. Consolidate IoT Edge Computing with Lightweight Virtualization. IEEE Network, volume 32, issue 1, 2018. DOI:10.1109/MNET.2018.1700175
WM: Capacity planning for vehicular fog computing
Tutor: Wencan Mao
5G and beyond networks provide broad prospects for smart vehicular applications. Vehicles that are equipped with fog devices and V2V communication modules can serve as local cloud to provide computing services to the nearby vehicles to reduce the latency and increase the Quality of Service. To meet the demand, fog computing resources need to be rational planned based on the computing demand of vehicular applications as well as spatio-temporal pattern of urban traffic.
References:
- 1. Y. Xiao and C. Zhu (2017). Vehicular fog computing: Vision and challenges, 2017 IEEE International Conference on Pervasive Computation and Communication Workshop, PerCom Workshop 2017, pp. 6–9, 2017.
- 2. M. Noreikis, Y. Xiao and A. Ylä-Jaäiski, "QoS-oriented capacity planning for edge computing," 2017 IEEE International Conference on Communications (ICC), Paris, 2017, pp. 1-6, doi: 10.1109/ICC.2017.7997387.
- 3. M. Noreikis, Y. Xiao and Y. Jiang, "Edge Capacity Planning for Real Time Compute-Intensive Applications," 2019 IEEE International Conference on Fog Computing (ICFC), Prague, Czech Republic, 2019, pp. 175-184, doi: 10.1109/ICFC.2019.00029.