Human-pose estimation is a technology with many applications such as healthcare, smart homes, and new methods of human-computer interaction. However, traditional RGB camera-based systems come with significant privacy risks and can perform poorly in dark rooms. A new approach to human-pose estimation, estimating through the use of mmWave radars, could solve these problems. mmWave creates a point cloud of a person, rather than a direct RGB image, and is therefore not affected by dark conditions, while simultaneously letting the subject stay anonymous. Current mmWave models are very accurate, on the order of centimetres, but generally too costly to run without a GPU.

In this paper, we create an optimised mmWave human-pose estimation model that runs more accurately without a GPU compared to a baseline model. We do this by analysing a baseline model to find which parts can be compressed without excessively losing accuracy.

Our improved model has an inference time of 41 ms with a Mean Absolute Error (MAE) of 7.72 cm on an embedded device. Compared to the baseline, this model saves 85.9% latency, at the cost of 4.8% MAE accuracy.

Through finding which parts can be compressed most effectively, we also gain insight into the relative importance of each component of the model. We also identify components that, with further research, could be improved to increase the accuracy of the model. ... Read more

Chemical Reaction Network (CRN) discovery is a time-consuming task that can be automated using program synthesis. However, the search space for realistic CRNs is large, making exhaustive search intractable. This paper investigates whether incorporating bond-breaking energy as a search heuristic can improve the efficiency of grammar-based CRN discovery. This paper proposes two heuristics: a maximum-bond-order heuristic and a more sophisticated delta-energy heuristic, which prioritise reactions with lower net energy change. Both are benchmarked against naive breadth-first search across an entire synthesis pipeline and the three individual stages, for seven Chemical Reaction Networks: water formation, methane combustion, photosynthesis, ethylene glycol formation, methyl acetate hydrolysis, an esterification reaction, and fermentation of glucose. Results show that both the Delta-Energy and Max-Bond heuristics consistently reduce the number of candidates explored and total runtime compared to BFS, with Delta-Energy generally outperforming the Max-Bond heuristic. However, neither heuristic guarantees improvement in all cases. These findings suggest that energy-guided search is a promising direction for scalable CRN discovery, with more accurate bond-energy estimation being a natural avenue for future work. ... Read more

Recovering a chemical reaction network (CRN) from concentration data can be framed as program synthesis, but the search scales poorly: reaching the ground-truth network requires enumerating a very large number of candidates. Hard constraints on the synthesizer's grammar prune candidates by chemical rules such as atom valence and mass balance, yet among the valid candidates a uniform search has no sense of which reactions are plausible. We ask where in a top-down CRN synthesizer knowledge from a reaction database most reduces the candidates explored before the target is found? We compare three integration points across five benchmarks, using the USPTO-50K and Rhea databases: (1) a database-derived building-block vocabulary, (2) a probabilistic context-free grammar (PCFG) over the grammar rules, and (3) an output reranker. The building-block vocabulary is the most positive result: it decides whether the target network is recovered at all, and a shuffle control attributes this to the database's frequency content rather than vocabulary size. Each corpus unlocks only the chemistry it covers, and rank-normalised merging recovers targets that a naive union dilutes. The PCFG adds ordering gain, while the reranker, acting only on the finished list, helps under corpus match and hurts under mismatch. Matching a corpus to the target chemistry, not enlarging it, is what turns a reaction database into useful search guidance.

... Read more

Recovery of full Chemical Reaction Networks from an incomplete (partially observable) problem is an extensive and combinatorial expensive process. This work investigates some of the possible methods to restrict the overall search space of program synthesis and obtain a better rank for the target network, through the use of reaction templates. As such, filter-based network pruning and ranking, as well as reaction size constraint are compared to the existing base synthesizer that this work builds upon, both in terms of search-space and target rank. Additionally, runtime is also considered as a second metric for these experiments. Overall, these methods have obtained some improvements in the small subset of benchmarked networks, but achieving integration of reaction templates into CRN program synthesis still remains an open challenge. ... Read more

Automatically tracking the positioning and alignment of human limbs, also known as Human Pose Estimation (HPE), was traditionally pioneered by camera-based systems like the Microsoft Kinect, and remains critical across domains from interactive gaming to healthcare patient monitoring. Millimeter-wave (mmWave) radar has emerged as a compelling alternative; by utilizing electromagnetic waves to detect points on the surface of objects, it offers a more cost-effective, privacy-preserving, and robust solution than traditional cameras. However, the spatial "point-clouds" generated by mmWave radars are particularly irregular, requiring pre-processing before they can be fed into deep learning models. While these pre-processing techniques are well-documented and can easily be implemented on in high-level environments like Python, adapting and optimizing these pipelines for low-power embedded devices remains an underexplored challenge. It is currently not clear whether point-cloud pre-processing can overcome the memory and computational restrictions of low-power devices.
This thesis profiles the memory footprint and latency of executing mmWave point-cloud pre-processing on micro-controllers, specifically an STM32 Cortex-M7 with 320 KB of SRAM with the goal of real-time performance by processing each data sample in under 100 ms.
We propose and evaluate seven pipeline variants, incorporating hardware-acceleration, lightweight alternative algorithms, pipeline restructuring to eliminate computational redundancies, and a single-pass iteration strategy to minimize cache misses. Experimental results demonstrate that structural optimization compresses peak memory consumption from 90 KB to 50 KB, successfully approaching the theoretical lower bound dictated by the output buffers. Our most highly optimized configuration achieves an exceptional average latency of 8.13 ms (with a worst-case peak of 12 ms), comfortably satisfying our real-time constraints.
Further analysis revealed that the average point count per frame is the primary driver of computational performance. Ultimately, this work validates that efficient, real-time end-to-end radar processing is entirely viable on highly resource-constrained micro-controllers. ... Read more

This report studies valence constraints for molecule synthesis in a staged program-synthesis framework for chemical reaction network discovery. The baseline system already performs limited molecule validation, but it does so as a final filter after candidate SMILES strings have been generated. The change evaluated here is a valence-constrained grammar for ringless molecules that encodes atom valence directly in the grammar, restricting derivations to locally valence-consistent construction steps.

The resulting grammar reproduces the same ringless molecule sets as the legacy baseline on the audited tests and benchmarks, and all generated molecules pass the repository’s valence-validity checks. Runtime results are mixed: on the small water benchmark, the new grammar becomes faster from depth 5 onward and reaches a 2.87× speedup at depth 10, while on methane and urea it remains slower throughout the measured ringless depth series. A fixed-count methane benchmark shows that this slowdown is not mainly caused by legacy validity checking, but by the added search overhead of the larger valence-aware grammar.

The main conclusion is that the new grammar preserves the audited ringless output behaviour while shifting pruning earlier in the search, but this does not translate into consistent runtime gains. ... Read more

Artificial Intelligence is increasingly being used in everyday devices. However, most AI systems are designed to run on powerful computers or cloud servers rather than on small, low-power devices such as microcontrollers. Running AI directly on these devices can reduce energy consumption and enable systems to operate without an internet connection. AIfES (Artificial Intelligence for Embedded Systems) is a machine learning framework that allows neural networks to be trained directly on microcontrollers. However, it currently lacks support for depthwise convolution, an important operation used in efficient neural network architectures such as MobileNet. As a result, many modern computer vision models cannot be trained within the framework.

This project extends AIfES with support for depthwise convolution and integrates the new operator into the existing training pipeline. The implementation was validated using a combination of manually verified test cases, comparisons with TensorFlow, and image classification experiments on embedded hardware. The results show that the new operator functions correctly during both inference and training. Models containing the implemented layer successfully learned classification tasks and achieved behavior similar to equivalent TensorFlow models. By adding support for depthwise convolution, this work expands the range of neural network architectures that can be trained directly on microcontrollers and contributes to making on-device AI more practical and flexible. ... Read more

Embedded sensing systems relying on energy harvesting — such as electromagnetic radiation, thermoelectric energy, and kinetic energy — generally are not able to harvest sufficient power to function under normal operations for most devices, and thus operate under severe power constraints. To ensure sustainable, battery free functionality, the microcontroller (MCU) must remain at a low power deep sleep state during idle periods. It is woken up by a sensor, sending an external hardware interrupt when an environmental event occurs. However, there is a trade off between a sensor’s power consumption, detection range, accuracy, and latency. This paper presents two primary contributions: 1) An empirical evaluation of various sensor wake up systems. 2) The design and implementation of a multi stage hierarchical event detection pipeline. This pipeline consists of an ultra low power coarse sensor that activates a high accuracy, but higher power sensor, minimizing the current draw while staying reliable. ... Read more

Chemical Reaction Networks (CRNs) are essential for understanding complex reactive processes, yet incomplete experimental data often leave many networks only partially known. Grammar-driven program synthesis offers an approach to completing partial CRNs, but atom-by-atom construction of molecular candidates causes a severe combinatorial explosion, and the baseline synthesiser lacks awareness of the structural context of target molecules. It is not yet known whether substructure-aware heuristics can improve the computational tractability of CRN discovery via program synthesis. To investigate this, BRICS (Breaking of Retrosynthetically Interesting Chemical Substructures) fragments were incorporated into the molecular context-free grammar, and Tanimoto similarity of Morgan2 fingerprints was used to guide reaction and network synthesis. The results show that BRICS fragmentation achieves a 13.8 percentage-point improvement in completing partial reactions on a dataset missing complex organic compounds by encoding large substructures as single grammar rules. Conversely, the enhanced grammar solves 27.4 percentage points fewer problems than the baseline on reactions missing small non-carbon species, as increased branching delays discovery of small molecules. Moreover, molecular similarity guidance does not improve performance in reaction rebalancing from SynRXN datasets, but it substantially reduces the search space in an example esterification CRN synthesis problem, requiring 2,565 fewer candidate reactions and 409 fewer candidate networks before discovering the targets. Thus, BRICS fragmentation and similarity-guided heuristics have distinct strengths. Future frameworks should split the candidate molecule pool between fragment-enhanced and atom-by-atom methods to successfully capture both large structural fragments and small, dissimilar species. ... Read more

This paper aims to investigate the effect of conflicting demonstrations on Inverse Reinforcement Learning (IRL). IRL is a method to understand the intent of an expert, by only feeding it demonstrations of that expert, which may be a promising approach for areas such as self driving vehicles, where there are a lot of demonstrations from experts. This paper aims to investigate the effect of conflicting demonstrations on IRL. Demonstrations may not always come from the same expert or the expert may prioritize different goals at times. For example, a driver may not always do grocery shopping at the same store or they may take a slightly different route on different occasions. The results showcase a negative effect from severely conflicting demonstrations on the ability of Max Entropy IRL to recover rewards, but do show some slightly optimistic results on more than two goals. ... Read more

A key issue in Reinforcement Learning (RL) research is the difficulty of defining rewards. Inverse Reinforcement Learning (IRL) is a technique that addresses this challenge by learning the rewards from expert demonstrations. In a realistic setting, expert demonstrations are collected from humans, and it is important to acknowledge that these demonstrations can deviate from rationality due to systematic biases known as cognitive biases. One group of cognitive biases, known as risk-sensitive cognitive biases, pertains to individuals' attitudes and behaviors towards risk and uncertainty. This paper investigates the extent to which IRL can learn from demonstrations that contain risk-sensitive cognitive biases such as loss aversion and risk aversion. Modelling biases using concepts from Prospect Theory and System 1 and 2 model and using Maximum Entropy IRL algorithm, this paper concludes that IRL can recreate similar solutions to experts but inferring the underlying motivations and the interactions between them is an intricate problem that requires novel approaches. ... Read more

Inverse Reinforcement Learning (IRL) aims to recover a reward function from expert demonstrations in a Markov Decision Process (MDP). The objective is to understand the underlying intentions and behaviors of experts and derive a reward function based on their reasoning, rather than their exact actions. However, expert demonstrations can be influenced by various types of noise (e.g., from random behavior) which can affect their accuracy and effectiveness in solving the MDP. This research investigates the capability of IRL to recover reward functions from noisy demonstrations. Three types of noises, namely Random Action Noise, Random Bias Noise, and Sparse Noise, are introduced and modeled. Demonstrations are generated with these noises, and the corresponding reward functions are recovered. Comparisons are made between the noisy and optimal recovered rewards using various metrics. The results indicate that IRL exhibits certain tolerance level against Random Events and Sparse Noise, while being more vulnerable to Random Bias Noise. ... Read more

Inverse Reinforcement Learning (IRL) is a subfield of Reinforcement Learning (RL) that focuses on recovering the reward function using expert demonstrations. In the field of IRL, Adversarial IRL (AIRL) is a promising algorithm that is postulated to recover non-linear rewards in environments with unknown dynamics. This study investigates the potential benefits of applying the Curriculum Learning (CL) strategy to the AIRL algorithm. For our experiments, we use a randomized partially observable Markov decision process in the form of a grid-world-like environment. Using only expert demonstrations obtained with an RL algorithm under the true reward function, we train AIRL in a variety of configurations and identify an effective curriculum. Our results show, that a well-constructed curriculum can enhance the performance of AIRL twofold in both key aspects: the speed of convergence and the efficiency of using expert demonstrations. We thus conclude that CL can be a useful addition to an AIRL-based solution. Full code is available online in the supplementary material https://github.com/mikhail-vlasenko/curriculum-learning-IRL. ... Read more

Purpose: This paper explores the potential of machine learning (ML) algorithms to mitigate uncertainty in early environmental assessments (ex-ante LCA), which are hindered by prospective nature and limited quantitative data availability. Methods: A systematic literature review with keyword searches on Scopus identified three ML categorization groups in ex-ante LCA: streamlined LCA, ex-ante LCA parameter projection, and ancillary models and data. Two following case studies addressed literature gaps in price forecasting for economic allocation and recycling rate projections. Results: In streamlined LCA, 16 studies linked molecular and technical parameters to project production-related emissions of organic chemicals, applied product clustering of product groups, and generated spatially explicit impact category results. The application of ex-ante LCA parameter projection, as evidenced by 14 publications, involves the use of ML to project life cycle inventory (LCI) data, project characterization factors, and integrate natural parameters with LCI data in a comprehensive modeling approach. In nine other papers the applications to ex-ante LCA remained undefined but potentially applicable. For both case studies, best results were obtained with a Recurrent Neural Networks (RNN) algorithm with long-short-term-memory (LSTM). Commodity price forecasting in the first case study achieved a projection accuracy of 0.96 (MSE), 0.98 (RSME), and 10.17% (MAPE) for copper and 88.86 (MSE), 9.43 (RMSE), and 21.23% (MAPE) for molybdenum. Probability modelling is identified as a modeling approach which incorporates uncertainty. The recycling rate forecast case study identified plastic recycling and glass recycling rates as the best suiting covariates and demonstrated multivariate modeling possibilities with 0.22 (MSE), 0.48 (RSME), and 0.38% (MAPE) in a model with 68 covariates. Discussion: A limited yet growing body of literature indicates that ML applications in ex-ante LCA represent an emerging field of science. While streamlined LCA shows promise, it faces constraints related to data precision and a static nature. In the ex-ante LCA parameter projection categorization, the sub-group of similarity clustering of LCI processes suffers from data uncertainty in LCI databases, making the approach more suitable for updates of existing technologies than for emerging ones. On the other hand, LCI generation through ex-ternal parameters represents a highly technology-specific case, showing significant promise. The projection of characterization factors and the sub-group of integrated modeling are identified as promising, but the limited number of scientific studies hinders the generalizability of these findings. Case studies on price forecasting and recycling rate projection demonstrate ML’s applicability in economic allocation and waste treatment projections. Overall, the results suggest that ML holds potential for reducing uncertainty in ex-ante LCA, laying the groundwork for focused research and contributing to a nuanced understanding of uncertainty reduction in this domain. Recommendations: The paper emphasizes the need for targeted research in the goal and scope phase and in End-of-Life (EoL) treatment forecasts, e.g. via the use of time-series multivariate modeling. Furthermore, it encourages further exploration of streamlined LCA into applications with a high degree of technical predictors, along with the extended projection of characterization factors and integrated modeling. Additionally, the use of probabilistic modeling as a tool to incorporate uncertainty into the modeling is recommended, aiming to enhance the applicability and transparency of ML applications for reducing uncertainties in ex-ante LCA. ... Read more