- Uncertainty Estimation in Deep Learning using Masked Autoencoders and Test-Time Training
Sepehr Maleki
Ph.D. Student
(Supervisor:Asst.Prof.Doruk Öner) Computer Engineering Department
Bilkent University
Abstract: Deep learning models often make predictions without indicating their confidence or explaining sources of uncertainty, limiting their deployment in safety-critical applications. This thesis proposes a framework that separates two types of uncertainty: epistemic uncertainty (model uncertainty about its parameters) and aleatoric uncertainty (inherent data noise or ambiguity). We use a dual-task architecture with a shared encoder that performs both masked autoencoding and segmentation. For epistemic uncertainty, we introduce an SVD-based weight perturbation method that identifies the most important network weights using singular value decomposition and perturbs only those weights. By running multiple forward passes with different perturbations, we measure prediction variance as a direct indicator of model parameter uncertainty. For aleatoric uncertainty, we measure how long test-time training takes to converge on the reconstruction task longer convergence indicates noisier or more ambiguous input data. This approach provides spatial uncertainty maps showing where the model is uncertain and distinguishes between uncertainty caused by the model’s limited knowledge versus uncertainty inherent in the data itself. The framework enables practical applications such as identifying when expert review is needed, when data quality should be improved, and which samples would be most valuable for active learning in medical imaging, autonomous driving, and other critical domains.
DATE: April 06, Monday @ 15:30 Place: EA 502
2. FUSOR: Space-efficient Taxonomic Classification Using Hybrid Hierarchical Interleaved Binary Fuse Filters with a Stash
Tuna Okçu
Master Student
(Supervisor:Assoc.Prof.Can Alkan) Computer Engineering Department
Bilkent University
Abstract: The exponential growth of metagenomic data has transformed taxonomic classification into a primary computational bottleneck, shifting the research focus from simple processing speed to the critical need for space-efficient indexing. Current gold-standard tools often require massive memory resources, which significantly limit their deployment on portable or resource-constrained systems. Here we introduce FUSOR, a reference-based taxonomic classification method that uses our novel hybrid hierarchical interleaved binary fuse filter (HHIBFF) data structure in its index. While binary fuse filters (BFFs) are highly space-efficient, their application in an interleaved architecture has been impractical due to high construction failure rates. FUSOR overcomes this barrier by implementing a stash-based failure recovery mechanism. By providing an auxiliary space to evict failure-inducing insertions, and strictly bounding the eviction budget to a single item per filter, we make the construction of interleaved BFFs viable for the first time. Our results show that the final index we constructed is smaller than state-of-the-art tools such as Kraken2, ganon2, and Taxor, requiring less than a third of the space in certain configurations. Availability: FUSOR is available on https://github.com/BilkentCompGen/FUSOR.
DATE: April 06, Monday @ 15:50 Place: EA 502
3. SAILR: Severity-Aware Multi-Label Whole Slide Image Retrieval via Attention-Driven Late Interaction
Sude Önder
Master Student
(Supervisor:Prof.Dr.Selim Aksoy) Computer Engineering Department
Bilkent University
Abstract: Whole Slide Image (WSI) retrieval is inherently multi-label, yet existing methods fail to account for both the co-occurrence of pathological patterns and their varying clinical severity, resulting in retrieval that does not prioritize diagnostically significant matches. We propose an attention-based late interaction framework that operates on region-level representations extracted from pretrained foundation models, enabling fine-grained similarity estimation between WSIs. The model explicitly computes interactions between region embeddings and learns to aggregate them via class-specific attention, effectively assigning higher importance to diagnostically relevant correspondences. This yields a learned similarity function that prioritizes informative region pairs over uniform or fixed aggregation schemes. We further introduce a severity-aware multi-label similarity function that embeds clinical relevance into the retrieval objective, replacing binary similarity with a continuous measure of multi-label similarity, and use it to supervise retrieval via a margin-based ranking objective. Experiments on the BRACS dataset demonstrate consistent improvements over state-of-the-art methods, while providing interpretable attention-based evidence.
DATE: April 06, Monday @ 16:30 Place: EA 502
4. Hardware acceleration of hashing algorithms using AMD NPUs
Berkan Şahin
Master Student
(Supervisor:Assoc.Prof.Can Alkan) Computer Engineering Department
Bilkent University
Abstract: With the ongoing AI boom brought upon by large language models, hardware vendors are starting to include neural processing units (NPUs) in their mobile CPU and SoC designs. These co-processors are primarily designed for accelerating low-power, on-device machine learning inference; with simple and thoroughly parallelized architectures compared to CPUs. Our work focuses on investigating non machine learning workloads using these NPUs, with a focus on computational genomics applications. To this end we aim to accelerate the MurmurHash3 algorithm, which is a non-cryptographic hashing algorithm widely used for computational genomics applications, on AMD’s AI Engine NPUs. We plan to provide a hardware-accelerated hashing library that can be used for porting applications, computational genomics or otherwise, on NPUs for energy-efficient and fast computation. We use the low-level IRON library to explicitly program the data movement and distribution required for parallelized hashing of many objects, taking advantage of the NPUs hardware-level data transformation capabilities.
DATE: April 06, Monday @ 16:50 Place: EA 502