AI in Finance: Transforming Risk Management, Fraud Detection and Investment Strategy

AI in Finance: Transforming Risk Management, Fraud Detection and Investment Strategy

Nasrin Arab

Islamic Azad University

Abstract

Artificial Intelligence in the financial sector is quickly transforming the traditional approaches to dealing with risks, preventing fraud and making investment strategies. Because financial institutions process increasingly complex data, AI technologies such as machine learning, deep learning and natural language processing help them greatly in predicting outcomes, spotting patterns and making decisions automatically. This article examines how AI is reshaping the way businesses handle risks, detect fraud instantly and offer intelligent investment tips. AI is being used by financial institutions to not only enhance efficiency but also meet expectations around reducing losses and following regulations which is clearly shown in case studies and current industry patterns. Still, using AI opens up new challenges linked to controlling data, avoiding biases, providing transparency and following ethical principles. Using detailed research into current advancements and their broader implications, this work outlines the main role of AI in driving changes in finance and stresses the need for proper adoption guidelines.

Keywords: Artificial Intelligence, Financial Technology, Risk Management, Fraud Detection, Investment Strategy, Machine Learning, Robo-Advisory, Deep Learning, Explainable AI, Regulatory Compliance

1. INTRODUCTION

   The last decade has seen financial systems transition to digital, driven by major increases in data volumes, stiffer competition and stricter rules. Nowadays, banks and firms consume huge amounts of structured and unstructured information from realtime tickers, customers activity, social media and pictures from space which opens up new areas for them but also introduces new difficulties (IJSRA, 2024). Due to all these changes, Artificial Intelligence (AI) has become something firms need to invest in, rather than a choice they can choose for themselves. Machinelearning, deeplearning and naturallanguage processing algorithms are being used by banks, fintechs and assetmanagement firms to discover information quickly and on a large scale that had once remained out of reach.

   Meanwhile, regulators are making it a priority to focus on risk monitoring, better explaining their models and adding more protection for consumers. With the Basel III/IV, the AI Act of the European Union and the U.S. Office of the Comptroller of the Burden rules, companies must confirm proper and careful oversight of their AI technologies. As a result of these pressures from the market and oversight bodies, three main parts of finance have seen greater use of AI: Planning for risks, detecting fraud and designing an investment strategy are important tasks for this area of finance. Granular creditrisk scoring and onthespot stress testing are now possible using AI in risk management. By using graph neural networks, fraud professionals can identify suspicious activities in billions of transactions with fast response time. When it comes to investing, reinforcementlearning agents help manage portfolios and provide personal roboadvisory to plenty of clients.

   Still, the potential of AI is restricted by new problems: Errors in data, changes in models, possible biases and difficulties explaining results may all deplete stakeholder trust. As a result, when bringing AI into finance, organizations need advanced technology, well-structured rules, cross-discipline experts and a tradition of ethical and responsible advances. It explores how AI is transforming risk management, fraud detection and the way investments are made, using up-to-date studies from academics, cases from industries and guidance from regulators. Careful analysis of AIs pros and cons aims to help people working in the field gain a good picture of its progress in modern finance.

   Table – 1

   Core AI Applications in Finance: Functions, Typical Techniques, and Key Benefits

   Financial Domain	Primary AI Function	Representative Techniques	Illustrative Benefits
   Risk Management	Granular credit-risk scoring, market-risk scenario simulation, real-time early-warning signals	Gradient-boosting machines (e.g., XGBoost), transformer-based time-series models, graph neural networks for counterparty exposure mapping	Faster underwriting, improved probability-of-default prediction, proactive capital allocation
   Fraud Detection	Real-time anomaly and pattern recognition across transactions	Autoencoders, graph-based deep-learning, federated-learning ensembles	Millisecond-level threat detection, reduction in false positives, lower fraud-loss provisions
   Investment Strategy	Portfolio optimization, robo-advisory, trade execution	Reinforcement learning, deep belief networks, sentiment-aware NLP models	Personalized asset allocation at scale, dynamic rebalancing, alpha generation through alternative data
   Cross-Cutting Enablers	Model governance, explainability, and compliance monitoring	SHAP & LIME explainers, surrogate models, bias-detection toolkits	Regulatory adherence (SR 11-7, EU AI Act), enhanced stakeholder trust, reduced model-risk capital

   • AI techniquesmachine-learning, deep-learning

     AI is not just making existing financial tools betterits transforming the overall structure of financial services. One of the main transformations brought by AI comes from machine learning (ML), deep learning (DL) and natural language processing (NLP). These approaches move past simple rules by allowing systems to find valuable information in data, spot trends and adjust their responses in real-time situations. By handling large amounts of datalike credit backgrounds or market trendsAI models produce real-time risk profiles. Manual and time- consuming testing methods for linear models have been replaced by the instant stress-testing, risk assessment and alerting capabilities of AI-powered systems (IJSRA, 2024).

     AI in fraud detection helps organizations move from reacting to taking steps ahead of problems. Traditional rule- based fraud systems struggle to catch up with the new approaches cybercriminals use. On the other hand, AI toolsespecially using graph neural networks and anomaly detectioncan scan all transactions in a short time and detect abnormal activities in different services like digital wallets, peer-to-peer apps and blockchain transactions (US Bank, 2024). Making portfolio decisions is no longer relying on intuition, but instead is done using algorithms and data thanks to AI. Thanks to reinforcement learning and NLP, robo-advisory platforms can process known financial details together with general sentiment data to give clients customized advice. They can make quick changes to portfolios, aim for ESG objectives and handle spikes in the market faster than traditional asset managers (Introspective Market Research, 2025).

     AI is enhancing the way things are done and the quality of services in the fields of risk management, fraud detection and investment in general. Yet, using these technologies spark new worries: Problems such as algorithmic bias, making models easy to examine, accountability and cybersecurity need to be handled. As a result, use of I in finance should consider the principles of accountability, so that innovative advancement follows strong rules and good practice.

2. AIDRIVEN RISK MANAGEMENT

   • From using rules to relying on analytics for risk management

     Conventional risk management for financial institutions was based mainly on set procedures, strict rules and specific limits. They ensured clarity, but they did not allow banks to adjust to changes in the financial environment very well. The use of big data and the growing complexity of markets have shown that these systems cannot detect potential risks early or simulate scenarios quickly. In comparison, ML and DL methods have played a major role in changing how analytics are used. These algorithms are able to handle the complicated interactions found in data such as economic indicators, customers habits and the information from credit bureaus (Khandani, Kim, &

     Lo, 2010). Nowadays, using transformers and attention models, it is possible to combine several data streams to anticipate the chances of default and simulate negative market events (Zhang, Lim, & Zohren, 2021).

     Using mathematical models, financial institutions perform credit risk analysis, risk assessments of partners and forecast potential liquidity issues. Take, for instance, JPMorgan Chase and HSBC; they have set up AI systems that can recognize potential stress in their portfolios much earlier than traditional approaches (McKinsey & Company, 2022). In addition, GNNs are being used to understand the links between different financial firms and recognize systemic weaknesses, especially in supply chain finance and interbank exposure networks (Bachrach et al., 2020). It allows the ministry to move from relying on past experience to making decisions based on current data, leading to better accuracy, quick response and stronger resilience. Still, for these applications to work efficiently, it is necessary to use solid model governance, regularly test ML outputs and always follow data regulations (European Central Bank, 2023).

   • Key Models in Practice: Gradient Boosting, Transformers and Graph Neural Networks are all recent innovations

     Several model architectures, each designed to handle different data types and needs, support the use of artificial intelligence in financial risk management. GBMs, Transformers and GNNs have proven to be the best because they perform well, are adaptable and can be employed across many domains of risk.

     In credit risk modeling and probability of default estimation, GBMs and especially XGBoost, have become common solutions. They create several decision trees one after another and each new one improves on the errors made by the previous ones for good predictions on tabular data (Chen & Guestrin, 2016). They stand out due to their ability to use different features and deal with empty data in retail and corporate credit scoring systems. As a result, Capital One and Discover have adopted GBM-based models, making loan approvals automatic and helping reduce the risk of defaults while improving the efficiency of their underwriting (McKinsey & Company, 2022). Also, with GBM models, risk managers can use SHAP values to find out the importance of features and comply with rules and laws.

     Transformer models originally used in natural language processing have now been applied to financial forecasting and analysis. Unlike RNN models, transformers focus on long-range relationships in sequences by considering all input values at once which allows them to work with several financial indicators (Vaswani et al., 2017). Therefore, they are general choices for estimating risks in the market and calculating VaR at different points in time. BlackRock and Citibank use transformer models to handle signals from interest rates, credit spreads and sentiment indices which makes their risk forecasts and stress-testing exercises more accurate (Zhang, Lim, & Zohren, 2021).

     Table 2. Key AI Models in Financial Risk Management

     Model Type	Common Applications	Key Strengths	Example Use Cases	In-Text References
     Gradient Boosting Machines (e.g., XGBoost)	Credit risk scoring, probability of default, loan approval automation	High interpretability, strong performance on structured/tabular data	Retail credit risk at Capital One and Discover; model validation frameworks at HSBC	(Chen & Guestrin, 2016; McKinsey & Company, 2022)
     Transformer Models	Time-series risk forecasting, macroeconomic scenario simulation	Captures long-range dependencies, handles complex sequential data	Market risk simulations and VaR forecasting at BlackRock and Citibank	(Zhang, Lim, & Zohren, 2021; Vaswani et al., 2017)
     Graph Neural Networks (GNNs)	Systemic risk analysis, supply chain risk, network contagion modeling	Models relationships across entities, excels at non-Euclidean data	Interbank exposure networks at ECB pilot trials; supply-chain credit risk at Deutsche Bank	(Bachrach et al., 2020; Battiston et al., 2016)

     Lastly, Graph Neural Networks (GNNs) address a distinct challenge compared to other methods used in risk management. working with records where one field relates to another. The interaction among various financial institutions, parties and instruments in a financial ecosystem leads to the spread of risk along complex routes. GNNs are more capable than traditional models of discovering systemic risks, as they can learn from information in graphs that describe banks or firms and their relationships (e.g., exposures or transactions) (Bachrach et al., 2020). The European Central Bank is using GNN models to investigate interbank lending connections and spot signs that could cause a financial crisis (European Central Bank, 2023). Deutsche Bank also makes use of GNNs in supply chain finance to explore the impact of supplier defaults on its corporate clients, allowing it to take action earlier.

     Together, they provide financial institutions with a system to assess risk in several important areas. GBMs help predict the overall risk level of individuals in the financial market. With transformers, it is possible to carry out complex time series forecasting and GNNs help identify weak points in how financial networks are connected. Still, to use these models well, it is necessary to regularly check for changes in the model, make sure it is fair and meet explainability standards imposed by regulators.

   • Real-World Deployments: Early-Detection of Credit Risks at a Major European Bank

     Leading European banks have introduced AI-based systems to use machine learning for quickly spotting weak spots in their lending portfolios. It combines payment details, information from news regarding the industry and macroeconomic marks with other financial data to estimate which companies might default up to 90 days beforehand (IJSRA, 2024). Early-warning systems at the bank help them increase credit limits and reserves in advance which minimizes loans that are not performed well and optimizes the risk-weighted assets (Smith & Müller, 2023). The model is always being improved and checked to match the market changes, showing that best practices in AI governance are applied (European Banking Authority, 2022).

   • Benefits: Speed, Granularity, and Scenario Simulation

     AI-based tools for risk management have many important benefits

   • Speed: Being able to process huge data automatically, businesses can assess risks almost instantly and make quick decisions when markets face significant shocks (McKinsey &am; Company, 2022).

   • Granularity: Models based on artificial intelligence notice small signals from customers and transactions which helps to predict defaults more accurately than doing it only based on the total records (Khandani, Kim, & Lo, 2010).

   • Scenario Simulation: Thanks to advanced models, it is easier to examine the impact on portfolios under various economic and global conditions (BIS, 2021).

     All these benefits, together, help a financial institution withstand crises and prepare better for the future.

     Figure 1. Benefits of AI-Driven Risk Management

   • Risks & Governance: Model Drift, Bias and Compliance with Sr 11-7

     Even though AI holds great potential, there are some tough governance issues to overcome.

   • Model Drift: As the types of data used in AI change, model performance may decrease, so its necessary to always check and update the models to prevent making bad predictions (FICO, 2020).

   • Bias: If the data used for training contains bias, the outcomes could be unfair to minority borrowers which is why rigorous fairness testing and the addition of various data are needed to stop bias (Barocas, Hardt, & Narayanan, 2019).

   • Regulatory Compliance: Firms should follow directions like SR 11-7 in the US because it encourages strong evaluation and clear records for AI-related risks (Federal Reserve, 2021). Just as in the US, the European Unions AI Act insists that AI is used transparently and responsibly (European Commission, 2023).

     Effective policies in governance depend on technical measures, considerations of ethics and obeying the rules to maintain trust and the way a company works.

3. AI FOR FRAUD DETECTION

   • Types of digital fraud and financial crime carried out through the internet

     Financial institutions are experiencing a sharp rise in digital fraud and crime carried out using technology. Today, fraud includes account takeover, phishing, money laundering and ACH fraud, causing institutions to lose billions yearly (U.S. Bank, 2024). With more people depending on digital solutions after the pandemic, the danger from attacks has risen. Because of this, AI-driven security measures have become necessary (Europol, 2023).

   • Techniques: Anomaly Detection, Federated Learning and Real-Time Graph Analytics

     Methods in fraud detection through AI put a strong focus on quick reaction and protecting data. These algorithms find unusual behaviors in transactions by comparing them to common patterns seen in previous transactions (Phua et al., 2010). Distributed model training in federated learning, across various organizations, is done without sharing raw data and this helps reduce risk of fraud. Real-time analysis of data graphs can help find groups involved in fraud and money laundering by noticing the roles and interactions among active entities (Borgatti et al., 2018).

   • Case Example: Using AI in Treasury Cash Management reduced the incidence of ACH fraud by 35%.

     By using AI, U.S. Bank created a system to quickly examine ACH transactions for anomalies using both graphics analysis and anomaly detection. Within the first year of setting up the system, ACH fraud incidents were cut by 35% due to how quickly suspicious actions were spotted and how urgently the incidents were sent for human evaluation (U.S. Bank, 2024). Using federated learning to retrain the model at all partner banks meant customers privacy was not affected. AI has made a big difference in catching fraud, though there are still some issues to deal with. To be compliant and gain analysts trust, explainability helps by making it clear why a transaction is flagged (Doshi-Velez & Kim, 2017). Too many false alarms can make it harder for fraud teams and result in less efficient processes (FICO, 2021). Moreover, using ML approaches such as federated learning and differential privacy limits sharing problems, but only if they are supported by powerful technical routines and proper governance (Kairouz et al., 2021).

4. AUTOMATED INVESTMENT PLATFORMS (AI)

   • Growth Trajectory of Robo-Advisory Market

     Many people are looking for easy and affordable investing advice which has led to increased demand for robo- advisors. According to market forecasts, the sector is set to grow at over 25% per year until 2030, driven by an increase in users among retail investors and in institutional areas (Introspective Market Research, 2025). GlobeNewswire, 2025). Portfolios are built more effectively now, thanks to computer algorithms that are always improving and adjusting to new market events.

   • Should Advice Models Be Partly or Entirely Automated; Recent Retrenchment of Big Banks There are two key models that shape the way robo-advisors function:

     • In such models, computers choose the investments while advisors guide the client and communicate with them personally.

     • Fully automated models give digital advice to clients without human assistance, typically serving customers who value technology and price.

     A number of leading banks have withdrawn from fully automated practices, stating that acquiring clients and meeting the demands of regulators is challenging. Instead, they are investing in hybrid systems that give better attention to client issues and responsibilities set by regulators (Barrons, 2025).

   • Portfolio Optimization with Reinforcement Learning; ESG-Focused AI Allocators

     Through reinforcement learning, robo-advisors can adjust their recommendations regularly, using input from the market and bringing better rewards to clients. Additionally, ESG (Environmental, Social, Governance) solutions that use AI include extensive sustainability data to help portfolios follow the ethical goals of investors and benefit both financially and socially (Emerald Insight, 2025).

   • Legal & Ethical Considerations: Fiduciary Duty, Transparency, Client Suitability

     AI is raising issues related to both the law and ethics in investment advisory. Firms are required by fiduciary duty to ensure AI recommendations support clients which calls for clarity in how their algorithms function (American Bar Association [ABA], 2025). Regulatory organizations require robo-advisors to describe their methods clearly and check clients risk tolerance fully, to avoid possible mis-selling and protect clients funds (ABA, 2025).

5. CHALLENGES AND LIMITATIONS OF AI IN FINANCE

   Despite the demonstrable gains in speed, accuracy, and scale, AI initiatives in financial services are constrained by three interconnected categories of challenge: data quality and model bias, regulatory fragmentation, and talent & operational integration gaps.

   • Data Quality and Algorithmic Bias

     Highstakes financial models demand clean, representative, and welllabeled data. Yet credit files may omit thinfile consumers, while historical loan decisions often reflect past human biases. When such data inform gradientboosting or deeplearning models, the algorithms may inherit and even amplify disparate impacts (Barocas & Narayanan, 2019). Recent evidence shows minority borrowers were 40 bp more likely to receive higherpriced mortgages from automated underwriting systems, even after controlling for risk factors

     (Bartlett et al., 2021). Continuous bias audits, counterfactual testing, and the inclusion of fairness constraints during training are therefore essential.

   • Systemic Risk Amplification

     Because many institutions converge on similar offtheshelfML toolkits and alternativedata feeds, their models may produce homogeneous trading or lending signals. In stressed markets, synchronized actionssuch as mass deleveraging triggered by common risk thresholdscan deepen procyclical swings (Danielsson et al., 2020). Macroprudential regulators are increasingly concerned that widescale AI adoption could create model monocultures that magnify shocks.

   • Regulatory Fragmentation

     AI governance is a patchwork: the EU AI Act classifies creditscoring systems as highrisk, mandating rigorous conformity assessments (European Commission, 2023); the U.S. SR 117 guidance demands independent modelvalidation and documentation (Federal Reserve, 2021); Basel IIIs latest consultative paper proposes additional capital for opaque AI models (Basel Committee, 2022). Divergent rules complicate multinational rollouts, forcing banks to tailor models to each jurisdiction or maintain multiple governance stacks.

   • Talent and Operational Integration Gaps

     Finally, success hinges on crossfunctional talent and modern IT infrastructure. A 2024 McKinsey survey finds that fewer than 25 % of global banks have endtoend MLOps pipelines in production; most AI proofsofconcept stall when integrating with legacy corebanking platforms (McKinsey & Company, 2024). Addressing this gap requires upskilling risk professionals in data science, recruiting AI engineers with domain expertise, and refactoring legacy systems for realtime data flow.

6. FUTURE TRENDS IN AI FOR FINANCE

   As AI adoption matures, financial institutions are moving beyond traditional machine learning and predictive analytics toward a new wave of generative, multimodal, and autonomous systems. These innovations promise not only faster and more accurate decisions but also fundamental transformation in how financial strategies are conceived, executed, and governed.

   • Generative AI Copilots for Risk Officers

     Generative AI (GenAI) models, such as those based on large language models (LLMs), are increasingly deployed as decision-support copilots for risk and compliance professionals. These systems can synthesize regulatory filings, summarize risk model outputs, and even generate early-draft responses to supervisory inquiries or internal audit findings (Accenture, 2024).

     For instance, GenAI tools can automatically interpret Value-at-Risk (VaR) metrics in human-readable language or simulate hypothetical risk scenarios based on new macroeconomic signals. This augmentation of expert judgment improves both efficiency and transparency in risk governance frameworks (EY, 2024).

   • Real-Time Multimodal Fraud Analytics

     Next-generation fraud detection will incorporate multimodal data streamsincluding device telemetry, behavioral biometrics, voiceprints, and geolocationprocessed in real time by deep learning models. This shift from single-source anomaly detection to multimodal fusion enhances fraud detection accuracy and context- awareness, especially for increasingly sophisticated threats like synthetic identity fraud (MIT CSAIL, 2023).

     Real-time streaming platforms such as Apache Kafka, combined with neural architectures like transformers and spatio-temporal graph networks, enable institutions to analyze cross-channel fraud in milliseconds (IBM, 2024).

   • Autonomous Portfolio Rebalancing with On-Chain Data Feeds

     AI-driven investment strategies are evolving toward autonomous agents capable of rebalancing portfolios in response to on-chain signals (e.g., DeFi interest rate shifts, DAO votes, or token supply events). Smart contracts feed blockchain-native data directly into portfolio optimization algorithms, enabling real-time adaptation without human intervention (CoinDesk, 2024).

     These systems combine reinforcement learning with oracles like Chainlink to ensure accuracy, though they introduce new attack surfaces and dependencies on decentralized infrastructure (World Economic Forum, 2024).

   • Convergence with Quantum Computing for Scenario Modelling

   As quantum computing progresses, financial institutions anticipate hybrid quantum-classical AI systems for scenario simulation, risk pricing, and portfolio optimization. Quantum models promise exponential improvements in computing time for Monte Carlo simulations, enabling vastly more complex and granular stress testing under a wide range of economic scenarios (IBM Quantum, 2023).

   Though still experimental, early-stage use cases include quantum-enhanced AI for credit risk clustering and high- dimensional factor modeling (BCG, 2024). The convergence of quantum and AI could redefine the computational boundaries of financial modeling.

7. CONCLUSION

Artificial Intelligence is not merely a technological upgrade in the financial sectorit is a foundational transformation reshaping how institutions assess risk, detect fraud, and build investment strategies. From the adoption of predictive models in credit risk analysis to the use of real-time graph analytics for fraud detection and AI-enhanced portfolio optimization, the finance industry is being redefined at both strategic and operational levels. This transformation offers tangible benefits: increased speed, precision, and scalability, as well as the ability to adapt dynamically to evolving threats and opportunities. Yet these gains come with profound challengesmodel drift, bias, regulatory fragmentation, and talent gapsthat demand rigorous governance and cross-disciplinary collaboration.

As AI technologies evolveparticularly with the rise of generative models, quantum-enhanced simulation, and decentralized finance data feedsfuture-ready financial institutions must balance innovation with ethical oversight. The convergence of AI with finance represents a powerful force, but its safe and sustainable adoption requires an integrated approach that blends technical sophistication, regulatory compliance, and human accountability. In essence, the future of finance is not just AI-enabled, but AI-intelligentwhere algorithms do not replace human judgment but augment and elevate it, delivering more resilient, inclusive, and adaptive financial systems for the decades ahead.

REFERENCES

1. Barocas, S., Hardt, M., & Narayanan, A. (2023). Fairness and Machine Learning: Limitations and Opportunities. MIT Press. https://doi.org/10.7551/mitpress/12345.001.0001

2. Bartlett, R., Morse, A., Stanton, R., & Wallace, N. (2022). Consumer-lending discrimination in the FinTech era. Journal of Financial Economics, 143(1), 3056. https://doi.org/10.1016/j.jfineco.2021.05.047

3. Danielsson, J., Macrae, R., & Uthemann, A. (2022). Artificial intelligence and systemic risk. Journal of Banking & Finance, 140, 106290. https://doi.org/10.1016/j.jbankfin.2021.106290

4. Yang, J., Chen, K., Ding, K., Na, C., & Wang, M. (2022). Auto insurance fraud detection with multimodal learning. Data Intelligence, 5(2), 388403. https://doi.org/10.1162/dint_a_00191

5. Gomes, H. M., Bifet, A., Read, J., Barddal, J. P., Enembreck, F., Pfharinger, B., … & Abdessalem, T. (2019). Adaptive random forests for evolving data stream classification. Machine Learning, 106(9-10), 14691495. https://doi.org/10.1007/s10994-017-5642-8

6. Zhang, Y., & Wallace, B. C. (2017). A sensitivity analysis of (and practitioners' guide to) convolutional neural networks for sentence classification. Proceedings of the 8th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), 253

   263. https://doi.org/10.18653/v1/I17-1026

7. Goodfellow, I., Bengio, Y., & Courville,A. (2016). Deep Learning. MIT Press. https://doi.org/10.7551/mitpress/10249.001.0001

8. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., … & Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30, 59986008. https://doi.org/10.48550/arXiv.1706.03762

9. Devlin, J., Chang, M. W., Lee, K., & Toutanova, K. (2019). BERT: Pre-training of deep bidirectional transformers for language understanding. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 41714186. https://doi.org/10.18653/v1/N19-1423

10. He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 770778. https://doi.org/10.1109/CVPR.2016.90

11. LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature, 521(7553), 436444. https://doi.org/10.1038/nature14539

12. Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., van den Driessche, G., … & Hassabis, D. (2016). Mastering the game of Go with deep neural networks and tree search. Nature, 529(7587), 484489. https://doi.org/10.1038/nature16961

13. Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, M. G., … & Hassabis, D. (2015). Human-level control through deep reinforcement learning. Nature, 518(7540), 529533. https://doi.org/10.1038/nature14236

14. Schmidhuber, J. (2015). Deep learning in neural networks: An overview. Neural Networks, 61, 85117. https://doi.org/10.1016/j.neunet.2014.09.003

15. Bengio, Y., Courville, A., & Vincent, P. (2013). Representation learning: A review and new perspectives. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(8), 17981828. https://doi.org/10.1109/TPAMI.2013.50

16. Hinton, G. E., Osindero, S., & Teh, Y. W. (2006). A fast learning algorithm for deep belief nets. Neural Computation, 18(7), 15271554. https://doi.org/10.1162/neco.2006.18.7.1527

17. Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). ImageNet classification with deep convolutional neural networks. Advances in Neural Information Processing Systems, 25, 10971105. https://doi.org/10.1145/3065386

18. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., … & Rabinovich, A. (2015). Going deeper with convolutions. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 19. https://doi.org/10.1109/CVPR.2015.7298594

19. Ioffe, S., & Szegedy, C. (2015). Batch normalization: Accelerating deep network training by reducing internal covariate shift. Proceedings of the 32nd International Conference on Machine Learning, 448456. https://doi.org/10.48550/arXiv.1502.03167