Abstract

The focus of recent research has shifted from merely increasing the Deep Neu- ral Networks (DNNs) performance in various tasks to DNNs which are more interpretable to humans. The field of eXplainable Artificial Intelligence (XAI) has observed various techniques, including saliency-based and concept-based ap- proaches. Concept-based approaches explain the model’s decisions in simple human understandable terms called Concepts. Concepts are human interpretable units of data and the thinking ground of humans. Explanations in terms of concepts enable detecting spurious correlations, inherent biases, or clever-hans. With the ad- vent of concept-based explanations, there have been various concept representation methods and automatic concept discovery algorithms. With automatic concept dis- covery, there are a variety of discovered concept evaluation metrics. Additionally, some approaches use concepts to improve the model explainability and generaliza- tion (Concept Oriented Deep Learning CODL). The concept-based approaches are fairly new, with many new representations coming up while there is very limited work on CODL. We thus provide a systematic review and taxonomy of various concept representations and their discovery algorithms in DNNs, specifically in vision. We also provide details on concept-based model improvement literature being the first to survey CODL methods

Abstract

Humans use abstract concepts instead of hard features for generalization. Recent interpretability research has focused on human-centered concept explanations of neural networks. We present Concept Distillation, a novel method and framework for concept-sensitive training to induce human-centered knowledge into the model. We use Concept Activation Vectors (CAVs) to estimate the model’s sensitivity and possible biases to a given concept. We extend CAVs to ante-hoc training from post-hoc analysis. We distill the conceptual knowledge from a pretrained knowledgeable teacher to a student model focused on a single downstream task. Our method can sensitize or desensitize the student model towards concepts. We show applications of concept-sensitive training to debias classification and to induce prior knowledge into a reconstruction problem. We also introduce the TextureMNIST dataset to evaluate the presence of complex texture biases. We show that concept-sensitive training can improve model interpretability, reduce biases, and induce prior knowledge

Abstract

This paper presents the analysis and results of a data-driven approach for fault detection and isolation in case of complete failure of a single motor on a hexacopter. The proposed approach consists of a two-stage architecture using the Rotation Forest algorithm, which can detect faults without any false alarms and achieve a true positive classification rate of 92.6% and a false positive classification rate of 0.06%. The classification results are compared to other methods such as Logistic Regression, Gaussian Naive Bayes, AdaBoost, and Random Forest. Over 120 datasets containing approximately 21,000 data points are generated in simulation - divided into two sets for training and validation of the model. Outdoor flight tests are performed to validate the classifier algorithm further. We can detect and classify the fault within 60ms of its occurrence. A dataset is published in the open-source domain and can be used for training similar models. The work presented in this paper is data-driven (or model free) since the classifier has no knowledge of the parameters of the UAV and is derived only based on the functional relationship between input and output variables.

Abstract

Music processing is a complex and intriguing phenomenon that has garnered extensive research interest in recent years. In this study, we investigated neural representations of music features by analyzing blood-oxygen-level-dependent (BOLD) signal fMRI activations of selected regions of interest during a continuous listening task. Additionally, we compared these activations to the hidden layer activations of a special class of deep neural networks (DNNs). Using representational similarity analysis, we found a strong correlation between low-level music feature encoding in the human brain regions of the Superior Temporal Gyrus (STG) and Heschl’s gyrus (HG) to the hidden layers of the DNN encoder. Our results suggest that self-supervised DNNs can serve as a reliable architecture to study the music processing domain, which is crucial considering the absence of naturalistic listening conditions in prior research Keywords: neural networks, representation similarity analysis, music cognition, music encoding, neuromusicology

Abstract

A reliable critic is central to on-policy actor-critic learning. But it becomes challenging to learn a reliable critic in a multi-agent sparse reward scenario due to two factors: 1) The joint action space grows exponentially with the number of agents 2) This, combined with the reward sparseness and environment noise, leads to large sample requirements for accurate learning. We show that regularising the critic with spectral normalization (SN) enables it to learn more robustly, even in multi-agent on-policy sparse reward scenarios. Our experiments show that the regularised critic is quickly able to learn from the sparse rewarding experience in the complex SMAC and RWARE domains. These findings highlight the importance of regularisation in the critic for stable learning.

Abstract

Recent advances in Reinforcement Learning (RL) have led to many exciting applications. These advancements have been driven by improvements in both algorithms and engineering, which have resulted in faster training of RL agents. We present marl-jax, a multi-agent reinforcement learning software package for training and evaluating social generalization of the agents. The package is designed for training a population of agents in multi-agent environments and evaluating their ability to generalize to diverse background agents. It is built on top of DeepMind's JAX ecosystem~\cite{deepmind2020jax} and leverages the RL ecosystem developed by DeepMind. Our framework marl-jax is capable of working in cooperative and competitive, simultaneous-acting environments with multiple agents. The package offers an intuitive and user-friendly command-line interface for training a population and evaluating its generalization capabilities. In conclusion, marl-jax provides a valuable resource for researchers interested in exploring social generalization in the context of MARL. The open-source code for marl-jax is available at: \href{this https URL}{this https URL}

Abstract

Diurnal variation of atmospheric CO2 and its driving factors are studied using high-precision in situ mixing ratio measurements and stable carbon isotopic CO2 (δ 13C-CO2) for the first time over Shadnagar, a suburban site in India, from November 2018 to October 2019. The annual mean mixing ratios of atmospheric CO2 and δ 13 C-CO2 are observed to be 414.76 ± 4.26 ppm and −11.19 ‰ ± 1.63 ‰. Maximum diurnal variability of atmospheric CO 2 was observed in summer monsoon (17.30 ± 9.29 ppm), and the minimum was noticed in winter (7.19 ± 0.11 ppm), indicating strong seasonality in diurnal variability of amplitude. To characterize the atmospheric CO 2 sources, the Miller-Tans model was implemented separately using the CO2 and its δ 13 C-CO2 isotopic data during day and night. However, a strong source signature of δ 13C-CO2 was observed during the night-time of summer monsoon with a slope of −37.42‰ ± 0.73‰, obtained using a reduced major axis regression. The identified source indicates possible emissions from fossil fuel combustion. Further, a Lagrangian back trajectory analysis was performed to study the influence of transport on the observed CO2 mixing ratios. The model trajectory confirms the influence of the Indian summer monsoon on the variations of atmospheric CO 2 mixing ratios. Our study shows that fossil fuel emissions are one of the major contributors of CO 2 in the study location (about ∼2 to 4 ppm) with biogenic fluxes adding a clear seasonality ranging from −8 to 4 ppm.

Abstract

The study reports diurnal and seasonal variations of atmospheric CO2 across multiple locations in India using the dedicated ground-based observations, satellite observations and model simulations. Data from seven sites collected during different time periods are analysed to understand the role of biospheric, fossil fuel fluxes on diurnal and seasonal variations of atmospheric CO2. The study also examines the impact of land use/land cover on the variability of atmospheric CO2 concentration. Results show that CO2 mixing ratios are highest during night and lowest in the afternoon. Ponmudi station shows homogeneous daily CO2 mixing ratios because of an active ecosystem and free of boundary layer influence. A high altitude site Ooty exhibits higher diurnal and seasonal variations, possibly due to entrapped emissions from tourism and settlements, while Nagpur’s (an urban site), CO2 mixing ratios remain moderate throughout the year due to active plantations. The observed CO2 is further compared with the simulations from atmospheric chemistry-transport model. The model is able to cap- ture the observed seasonal cycle all over the study locations.

Abstract

A large number of computational and scientific methods commonly require decomposing a sparse matrix into triangular factors as LU decomposition. A common problem faced during this decomposition is that even though the given matrix may be very sparse, the decompo- sition may lead to a denser triangular factors due to fill-in. A significant fill-in may lead to prohibitively larger computational costs and memory requirement during decomposition as well as during the solve phase. To this end, several heuristic sparse matrix reordering methods have been proposed to reduce fill-in before the decomposition. However, finding an optimal reordering algorithm that leads to minimal fill-in during such de- composition is known to be a NP-hard problem. A reinforcement learning based approach is proposed for this problem. The sparse matrix reorder- ing problem is formulated as a single player game. More specifically, Monte-Carlo tree search in combination with neural network is used as a decision making algorithm to search for the best move in our game. The proposed method, Alpha Elimination is found to produce significantly lesser non-zeros in the LU decomposition as compared to existing state- of-the-art heuristic algorithms with little to no increase in overall running time of the algorithm. The code for the project is publicly available 1. Keywords: Reinforcement Learning · Sparse Matrices · Deep Learning · LU · MCTS

Abstract

Cocaine use disorder (CUD) is described as a compulsive urge to seek and consume cocaine despite the inimical consequences. MRI studies from different modalities have shown that CUD patients exhibit structural and/or functional connectivity pathology among several brain regions. Nevertheless, both connectivities are commonly studied and analyzed separately, which may potentially obscure its relationship between them, and with the clinical pathology. Here, we compare and contrast structural and functional brain networks in CUD patients and healthy controls (HC) using multimodal fusion. The sample consisted of 63 (8 females) CUD patients and 42 (9 females) healthy controls (HC), recruited as part of the SUDMEX CONN database. For this, we computed a battery of graph-based measures from multi-shell diffusion-weighted imaging and resting state fc-fMRI to quantify local and global connectivity. Then we used multimodal canonical component analysis plus joint independent component analysis (mCCA+jICA) to compare between techniques, and evaluate group differences and its association with clinical alteration. Unimodal results showed a striatal decrease in the participation coefficient, but applied supervised data fusion revealed other regions with cocaine-related alterations in joint functional communication. When performing multimodal fusion analysis, we observed a higher centrality of the interrelationship and a lower participation coefficient in patients with CUD. In contrast to the unimodal approach, the multimodal fusion method was able to reveal latent information about brain regions involved in impairment due to cocaine abuse. The present results could help in understanding the pathology of CUD in order to develop better pre-treatment/post-treatment intervention designs. Keywords: Cocaine use disorder; Imaging; Graph theory; sMRI; fMRI; Multimodal fusion; mCCA + jICA