#debiasing algorithms

[read-online]

This video explores how Artificial Intelligence (AI) is creating new job opportunities and income streams for young people. It details several ways AI can be used to generate income, such as developing AI-powered apps, creating content using AI tools, and providing AI consulting services.

The video also provides real-world examples of young entrepreneurs who are successfully using AI to earn money. The best way to get started is to get today the “10 Ways To Make Money With AI for Teens and Young Adults”

[ixnai]

Bias is not a bug; it’s a feature. In the realm of artificial intelligence, bias is often misconstrued as an unintended flaw. However, this perspective overlooks the inherent nature of machine learning systems. These systems, by design, learn from data—data that is invariably tainted with the biases of its human creators.

Consider the training datasets as the DNA of AI. Just as genetic material carries the traits of its predecessors, datasets carry the biases of historical human decisions. These biases are not anomalies; they are deeply embedded patterns that AI models are programmed to recognize and replicate. When an AI system exhibits bias, it is not malfunctioning. It is functioning precisely as it was trained to do.

The architecture of neural networks further compounds this issue. These networks, with their layers of interconnected nodes, are designed to identify and amplify patterns. When biased data is fed into these networks, the bias is not just preserved—it is often magnified. This amplification is akin to a feedback loop in audio systems, where a small sound is echoed and intensified until it becomes a deafening roar.

Avoiding the pitfalls of AI bias requires a paradigm shift in how we approach AI development. It begins with the curation of training data. Data must be meticulously vetted for bias, a process that demands both technical acumen and ethical oversight. This is not a trivial task; it requires a multidisciplinary approach, combining the expertise of data scientists, ethicists, and domain experts.

Moreover, the algorithms themselves must be designed with bias mitigation in mind. Techniques such as adversarial debiasing and reweighting can be employed to counteract the skewed distributions in training data. These methods, however, are not panaceas. They require constant refinement and validation to ensure they do not introduce new biases or degrade model performance.

Transparency is another critical component. AI systems must be designed with interpretability in mind. This means developing models that not only make predictions but also provide insights into how those predictions are made. Techniques such as SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) offer pathways to demystify the decision-making processes of complex models.

Ultimately, the responsibility lies with the creators of AI systems. They must acknowledge that bias is an intrinsic feature of AI, not an extraneous bug. By embracing this reality, they can take proactive steps to mitigate its effects, ensuring that AI serves as a tool for equity rather than a perpetuator of disparity. The path forward is not simple, but it is necessary. In the world of AI, understanding bias as a feature is the first step towards responsible innovation.

[drmikewatts]

IEEE Transactions on Emerging Topics in Computational Intelligence Volume 9, Issue 3, June 2025

1) An Efficient Sampling Approach to Offspring Generation for Evolutionary Large-Scale Constrained Multi-Objective Optimization

Author(s): Langchun Si, Xingyi Zhang, Yajie Zhang, Shangshang Yang, Ye Tian

Pages: 2080 - 2092

2) Long-Tailed Classification Based on Coarse-Grained Leading Forest and Multi-Center Loss

Author(s): Jinye Yang, Ji Xu, Di Wu, Jianhang Tang, Shaobo Li, Guoyin Wang

Pages: 2093 - 2107

3) Two ZNN-Based Unified SMC Schemes for Finite/Fixed/Preassigned-Time Synchronization of Chaotic Systems

Author(s): Yongjun He, Lin Xiao, Linju Li, Qiuyue Zuo, Yaonan Wang

Pages: 2108 - 2121

4) Solving Multiobjective Combinatorial Optimization via Learning to Improve Method

Author(s): Te Ye, Zizhen Zhang, Qingfu Zhang, Jinbiao Chen, Jiahai Wang

Pages: 2122 - 2136

5) Multi-Objective Integrated Energy-Efficient Scheduling of Distributed Flexible Job Shop and Vehicle Routing by Knowledge-and-Learning-Based Hyper-Heuristics

Author(s): YaPing Fu, ZhengPei Zhang, Min Huang, XiWang Guo, Liang Qi

Pages: 2137 - 2150

6) MTMD: Multi-Scale Temporal Memory Learning and Efficient Debiasing Framework for Stock Trend Forecasting

Author(s): Mingjie Wang, Juanxi Tian, Mingze Zhang, Jianxiong Guo, Weijia Jia

Pages: 2151 - 2163

7) Cross-Scale Fuzzy Holistic Attention Network for Diabetic Retinopathy Grading From Fundus Images

Author(s): Zhijie Lin, Zhaoshui He, Xu Wang, Wenqing Su, Ji Tan, Yamei Deng, Shengli Xie

Pages: 2164 - 2178

8) Leveraging Neural Networks and Calibration Measures for Confident Feature Selection

Author(s): Hassan Gharoun, Navid Yazdanjue, Mohammad Sadegh Khorshidi, Fang Chen, Amir H. Gandomi

Pages: 2179 - 2193

9) Modeling of Spiking Neural Network With Optimal Hidden Layer via Spatiotemporal Orthogonal Encoding for Patterns Recognition

Author(s): Zenan Huang, Yinghui Chang, Weikang Wu, Chenhui Zhao, Hongyan Luo, Shan He, Donghui Guo

Pages: 2194 - 2207

10) A Learning-Based Two-Stage Multi-Thread Iterated Greedy Algorithm for Co-Scheduling of Distributed Factories and Automated Guided Vehicles With Sequence-Dependent Setup Times

Author(s): Zijiang Liu, Hongyan Sang, Biao Zhang, Leilei Meng, Tao Meng

Pages: 2208 - 2218

11) A Novel Hierarchical Generative Model for Semi-Supervised Semantic Segmentation of Biomedical Images

Author(s): Lu Chai, Zidong Wang, Yuheng Shao, Qinyuan Liu

Pages: 2219 - 2231

12) PurifyFL: Non-Interactive Privacy-Preserving Federated Learning Against Poisoning Attacks Based on Single Server

Author(s): Yanli Ren, Zhe Yang, Guorui Feng, Xinpeng Zhang

Pages: 2232 - 2243

13) Learning Uniform Latent Representation via Alternating Adversarial Network for Multi-View Clustering

Author(s): Yue Zhang, Weitian Huang, Xiaoxue Zhang, Sirui Yang, Fa Zhang, Xin Gao, Hongmin Cai

Pages: 2244 - 2255

14) Harnessing the Power of Knowledge Graphs to Improve Causal Discovery

Author(s): Taiyu Ban, Xiangyu Wang, Lyuzhou Chen, Derui Lyu, Xi Fan, Huanhuan Chen

Pages: 2256 - 2268

15) MSDT: Multiscale Diffusion Transformer for Multimodality Image Fusion

Author(s): Caifeng Xia, Hongwei Gao, Wei Yang, Jiahui Yu

Pages: 2269 - 2283

16) Adaptive Feature Transfer for Light Field Super-Resolution With Hybrid Lenses

Author(s): Gaosheng Liu, Huanjing Yue, Xin Luo, Jingyu Yang

Pages: 2284 - 2295

17) Broad Graph Attention Network With Multiple Kernel Mechanism

Author(s): Qingwang Wang, Pengcheng Jin, Hao Xiong, Yuhang Wu, Xu Lin, Tao Shen, Jiangbo Huang, Jun Cheng, Yanfeng Gu

Pages: 2296 - 2307

18) Dual-Branch Semantic Enhancement Network Joint With Iterative Self-Matching Training Strategy for Semi-Supervised Semantic Segmentation

Author(s): Feng Xiao, Ruyu Liu, Xu Cheng, Haoyu Zhang, Jianhua Zhang, Yaochu Jin

Pages: 2308 - 2320

19) CVRSF-Net: Image Emotion Recognition by Combining Visual Relationship Features and Scene Features

Author(s): Yutong Luo, Xinyue Zhong, Jialan Xie, Guangyuan Liu

Pages: 2321 - 2333

20) Generative Network Correction to Promote Incremental Learning

Author(s): Justin Leo, Jugal Kalita

Pages: 2334 - 2343

21) A Cross-Domain Recommendation Model Based on Asymmetric Vertical Federated Learning and Heterogeneous Representation

Author(s): Wanjing Zhao, Yunpeng Xiao, Tun Li, Rong Wang, Qian Li, Guoyin Wang

Pages: 2344 - 2358

22) HGRL-S: Towards Heterogeneous Graph Representation Learning With Optimized Structures

Author(s): Shanfeng Wang, Dong Wang, Xiaona Ruan, Xiaolong Fan, Maoguo Gong, He Zhang

Pages: 2359 - 2370

23) Prompt-Based Out-of-Distribution Intent Detection

Author(s): Rudolf Chow, Albert Y. S. Lam

Pages: 2371 - 2382

24) Multi-Graph Contrastive Learning for Community Detection in Multi-Layer Networks

Author(s): Songen Cao, Xiaoyi Lv, Yaxiong Ma, Xiaoke Ma

Pages: 2383 - 2397

25) Observer-Based Event-Triggered Optimal Control for Nonlinear Multiagent Systems With Input Delay via Reinforcement Learning Strategy

Author(s): Xin Wang, Yujie Liao, Lihua Tan, Wei Zhang, Huaqing Li

Pages: 2398 - 2409

26) SODSR: A Three-Stage Small Object Detection via Super-Resolution Using Optimizing Combination

Author(s): Xiaoyong Mei, Kejin Zhang, Changqin Huang, Xiao Chen, Ming Li, Zhao Li, Weiping Ding, Xindong Wu

Pages: 2410 - 2426

27) Toward Automatic Market Making: An Imitative Reinforcement Learning Approach With Predictive Representation Learning

Author(s): Siyuan Li, Yafei Chen, Hui Niu, Jiahao Zheng, Zhouchi Lin, Jian Li, Jian Guo, Zhen Wang

Pages: 2427 - 2439

28) CIGF-Net: Cross-Modality Interaction and Global-Feature Fusion for RGB-T Semantic Segmentation

Author(s): Zhiwei Zhang, Yisha Liu, Weimin Xue, Yan Zhuang

Pages: 2440 - 2451

29) BAUODNET for Class Imbalance Learning in Underwater Object Detection

Author(s): Long Chen, Haohan Yu, Xirui Dong, Yaxin Li, Jialie Shen, Jiangrong Shen, Qi Xu

Pages: 2452 - 2461

30) DFEN: A Dual-Feature Extraction Network-Based Open-Set Domain Adaptation Method for Optical Remote Sensing Image Scene Classification

Author(s): Zhunga Liu, Xinran Ji, Zuowei Zhang, Yimin Fu

Pages: 2462 - 2473

31) Distillation-Based Domain Generalization for Cross-Dataset EEG-Based Emotion Recognition

Author(s): Wei Li, Siyi Wang, Shitong Shao, Kaizhu Huang

Pages: 2474 - 2490

32) NeuronsGym: A Hybrid Framework and Benchmark for Robot Navigation With Sim2Real Policy Learning

Author(s): Haoran Li, Guangzheng Hu, Shasha Liu, Mingjun Ma, Yaran Chen, Dongbin Zhao

Pages: 2491 - 2505

33) Adaptive Constrained IVAMGGMM: Application to Mental Disorders Detection

Author(s): Ali Algumaei, Muhammad Azam, Nizar Bouguila

Pages: 2506 - 2530

34) Visual IoT Sensing Based on Robust Multilabel Discrete Signatures With Self-Topological Regularized Half Quadratic Lifting Functions

Author(s): Bo-Wei Chen, Ying-Hsuan Wu

Pages: 2531 - 2544

35) Heterogeneity-Aware Clustering and Intra-Cluster Uniform Data Sampling for Federated Learning

Author(s): Jian Chen, Peifeng Zhang, Jiahui Chen, Terry Shue Chien Lau

Pages: 2545 - 2556

36) Model-Data Jointly Driven Method for Airborne Particulate Matter Monitoring

Author(s): Ke Gu, Yuchen Liu, Hongyan Liu, Bo Liu, Lai-Kuan Wong, Weisi Lin, Junfei Qiao

Pages: 2557 - 2571

37) Personalized Exercise Group Assembly Using a Two Archive Evolutionary Algorithm

Author(s): Yifei Sun, Yifei Cao, Ziang Wang, Sicheng Hou, Weifeng Gao, Zhi-Hui Zhan

Pages: 2572 - 2583

38) PFPS: Polymerized Feature Panoptic Segmentation Based on Fully Convolutional Networks

Author(s): Shucheng Ji, Xiaochen Yuan, Junqi Bao, Tong Liu, Yang Lian, Guoheng Huang, Guo Zhong

Pages: 2584 - 2596

39) Low-Bit Mixed-Precision Quantization and Acceleration of CNN for FPGA Deployment

Author(s): JianRong Wang, Zhijun He, Hongbo Zhao, Rongke Liu

Pages: 2597 - 2617

40) Bayesian Inference of Hidden Markov Models Through Probabilistic Boolean Operations in Spiking Neuronal Networks

Author(s): Ayan Chakraborty, Saswat Chakrabarti

Pages: 2618 - 2632

[govindhtech]

Machine Learning Periodic Table: Unifying AI Algorithms

Researchers from MIT, Microsoft, and Google created Information Contrastive Learning (I-Con), the “machine learning periodic table,” to unite machine learning methodologies. The periodic table organises elements, whereas this table organises machine learning algorithms by learning data point correlations.

I-Con presents a single paradigm to show how classification, regression, big language modelling, clustering, dimensionality reduction, and spectral graph theory are all mathematically similar.

Machine learning periodic table importance

New machine learning algorithms and methods are being developed. To comprehend the essential ideas and links across methods. Benefits of the machine learning periodic table:

Unification: It shows how a single mathematical framework links numerous prominent machine learning approaches. Researchers and practitioners can benefit from understanding algorithm similarities.

The I-Con structure emphasises machine learning technique relationships, like the chemical periodic table defines element interactions. It organises the enormous array of algorithms into a clear style.

Discovery: Its ability to spark fresh discoveries is most exciting. Gaps in the chemical periodic table anticipated unknown elements. The Periodic Table of Machine Learning features “empty spaces” that imply undeveloped algorithms.

Innovation: I-Con helps researchers experiment, redefine “neighbors,” change connection confidence, and mix tactics from various algorithms to build new ways. It encourages creativity and the blending of previously unrelated methods.

Efficiency: This framework lets academics create new machine learning algorithms without “reinventing past ideas”. Understanding the ideas and algorithms in the table helps them strategically explore new methods.

How was the machine learning periodic table created?

The atomic table was an unplanned study outcome. Shaden Alshammari, an MIT Freeman Lab researcher, studied clustering, a technique to group related data elements. She connected contrastive learning with grouping. Contrastive machine learning compares positive and negative data.

Alshammari discovered that both techniques could be expressed by the same fundamental equation by studying their mathematics. Following this turning point, the Information Contrastive Learning (I-Con) paradigm was developed to illustrate that machine learning algorithms imitate real-world data linkages while minimising errors.

The researchers created a periodic table using these findings. The table distinguishes algorithms by two main factors:

Point relationships in actual datasets: Data linkages including visual likeness, shared class labels, and cluster membership are involved. These “connections” may not be 100% trustworthy.

The main ways algorithms approximate those connections: These relationships are acquired and reflected internally via algorithms.

By categorising various existing machine learning techniques inside this framework using these two criteria, the researchers found that many popular algorithms line neatly within defined “squares”. They also observed “gaps” where framework-logical algorithms have not yet been built.

How to Fill Gaps

This method helped researchers construct a current system for detecting photographs without human labelling. Combining debiased contrastive representation learning connection concepts with clustering approximation connections helped them “fill a gap” in their periodic table. This new method improved ImageNet-1K picture categorisation accuracy by 8%. They also found that contrastive learning data debiasing might increase clustering accuracy.

I-Con Learning

I-Con redefines machine learning as a tool for understanding complex data interactions. Consider a bustling party where data points, or visitors, meet at tables representing clusters and discuss shared hobbies or hometowns. Consider machine learning techniques as methods guests find friends and settle in.

I-Con simplifies real-world data point connections to make them easier to work with in algorithms. The concept of “connection” might entail appearing alike, sharing labels, or being in the same group. All algorithms try to close the gap between the connections they learn to imitate and the true connections in the training data.

Researchers Use the Periodic Table

The I-Con-based machine learning periodic table has various functions beyond organisation. This gives academics a toolkit for developing unique algorithms. When various machine learning approaches are defined in I-Con's conceptual language, experimenting with variants is easier:

Redefining neighbourhoods entails testing different ways to organise data points into “neighbours”.

Adjusting uncertainty requires varying trust in learnt connections. integrating strategies entails integrating approaches from different algorithms in unique ways.

Every modification might lead to a new periodic table entry. The table may easily be modified to include rows and columns to show more data points' relationships.

Looking Ahead

As artificial intelligence advances and its uses develop, frameworks like I-Con help us understand the area. They help researchers find hidden patterns and enable purposeful innovation. For non-AI professionals, it's a reminder that even in complex fields, basic patterns and structures are waiting to be identified.

Sorting algorithms by how they understand and estimate data point relationships is the basic notion. A full chart that lists all algorithms and their connection and approximation techniques would require more information than this. A basic table to illustrate the notion may look like this:

[ai-due-diligence]

AI Algorithm Due Diligence: A Comprehensive Topic Explainer

AI algorithm due diligence is the systematic evaluation of AI models before deployment. This process ensures accuracy, fairness, and reliability, safeguarding businesses from potential risks. Without it, AI implementations can lead to flawed operations, legal repercussions, and loss of trust.

Understanding AI Algorithm Due Diligence

At its core, due diligence of AI algorithm involves assessing the design, data inputs, and outputs of AI models. Analysts review how algorithms are trained, examining data quality, bias, and representativeness. This includes detailed investigations into data sourcing, ensuring datasets reflect real-world diversity. Without this step, biased algorithms can lead to flawed outcomes, impacting business decisions and alienating key demographics.

Key Components of the Due Diligence Process

Data integrity is vital in due diligence of AI algorithm. Ensuring that data used for training is diverse, accurate, and compliant with regulations is crucial. Analysts evaluate data cleansing processes and data augmentation techniques used during model training. Additionally, algorithm transparency is assessed by evaluating whether AI decisions are interpretable and explainable. Analysts also review documentation, ensuring that each AI decision is traceable to specific data points. Without transparency, businesses face operational risks and regulatory scrutiny.

Evaluating Algorithm Performance

Performance metrics, such as precision, recall, and accuracy, are scrutinised during due diligence of AI algorithm. Analysts simulate various scenarios to test robustness and adaptability, including edge cases that represent rare but critical situations. This process highlights potential weaknesses, such as overfitting or poor performance with unseen data. Analysts also compare model performance across different user groups to ensure consistent accuracy and fairness.

Addressing Bias and Fairness

Bias detection is a cornerstone of AI algorithm due diligence. Evaluators check for discriminatory patterns that may disadvantage certain groups by applying fairness metrics like demographic parity and equalised odds. By rectifying biases, companies ensure ethical AI usage, avoiding reputational and legal issues. Analysts also recommend techniques such as re-weighting training data or using adversarial debiasing algorithms.

Regulatory and Compliance Checks

Ensuring that AI algorithms comply with industry standards and legal frameworks is critical. Due diligence of AI algorithm verifies adherence to data privacy laws, such as GDPR, CCPA, and sector-specific regulations like HIPAA in healthcare or MiFID II in finance. This prevents penalties and fosters trust among clients and stakeholders. Analysts document all compliance checks and provide guidelines for maintaining regulatory alignment as laws evolve.

Ongoing Monitoring and Maintenance

Due diligence doesn’t end after deployment. Continuous monitoring is essential to ensure AI models adapt to new data and evolving conditions. Regular audits under the umbrella of due diligence of AI algorithm help maintain accuracy, fairness, and reliability. This includes implementing automated monitoring tools that alert analysts to performance drifts and potential biases over time.

Benefits of Thorough Due Diligence of AI Algorithm

Comprehensive due diligence enhances decision-making, mitigates risks, and ensures operational efficiency. Companies investing in due diligence of AI algorithm safeguard their innovations, comply with regulations, and maintain competitive advantages in an AI-driven market. Additionally, they build trust with consumers and stakeholders, knowing that their AI systems are robust, fair, and compliant. The long-term benefits include reduced legal risks, improved operational efficiency, and stronger market positioning.

Future Trends in AI Algorithm Due Diligence

As AI technologies evolve, the scope of due diligence of AI algorithm is expanding. Emerging trends include automated due diligence tools that utilise AI to audit other AI systems, reducing human effort while maintaining accuracy. Advanced methods like federated learning allow secure model training on decentralised data, addressing privacy concerns. Analysts are also exploring AI-generated synthetic data for training, minimising bias while preserving data quality. The integration of blockchain technology ensures immutable records of due diligence processes, enhancing transparency and trust. Embracing these trends ensures that due diligence of AI algorithm remains robust, adaptable, and future-ready.

[shakshi09]

Why is ML model biased?

Machine learning (ML) models can be biased due to multiple factors, including biased training data, algorithmic limitations, and societal influences. Bias in ML arises when the data used to train models does not represent the real-world distribution accurately. If historical data contains prejudices, the model may learn and perpetuate them, leading to unfair outcomes.

One major source of bias is data collection. If the dataset lacks diversity or overrepresents a specific group, the model may favor that group while disadvantaging others. For instance, facial recognition models trained on primarily Caucasian faces tend to perform poorly on individuals of other ethnicities.

Another cause is labeling bias, where human annotators introduce subjective opinions or errors during data labeling. If a hiring model is trained on past recruitment data where men were favored for leadership roles, it might learn to prefer male candidates, reinforcing gender discrimination.

Algorithmic bias also plays a role. Some ML models may amplify small biases present in training data due to the way they generalize patterns. This issue is common in reinforcement learning and deep learning models, where certain features receive more weight, even if they are not relevant.

Bias can also emerge from feedback loops, where AI systems continuously learn from past predictions. If a search engine prioritizes showing users content similar to their previous searches, it may reinforce existing preferences and limit exposure to diverse perspectives.

Mitigating bias requires diverse datasets, fairness-aware algorithms, and continuous monitoring. Techniques like re-weighting training data, using adversarial debiasing, and implementing fairness constraints can help address bias in ML models.

Understanding and addressing bias is crucial for responsible AI development. If you're interested in learning how to build fair and unbiased models, consider enrolling in a Generative AI Course.

[aicerts0]

Ethical AI: Building Fair And Sustainable Workplaces 

Exploring Ethical AI in the Workplace

Ethical AI in the workplace refers to well-defined guidelines related to individual values, which involve adhering to non-discriminatory practices, non-manipulation, respecting individual rights, privacy and fair AI practices at the workplace to improve AI job quality. 

It prioritizes fundamental importance to ethical considerations in determining the legitimate use of AI in the workplace.

Source : Hibernian Recruitment

Strategies for Fairness and Sustainability in the Workplace 

There are Five Pillars of AI Ethics which includes: 

1. Accountability 

Accountability in AI is crucial for completing and speeding up processes which ensures reliability, requiring continuous evaluation by CIOs to maintain efficient business operations. 

2. Reliability 

AI must be dependable, for seamless and error-free outputs. 

3. Explain-ability 

AI and ML models should be understood and well explained across departments and organizations otherwise the benefits of AI become irrelevant if the technology is not coherent. 

4. Security 

It is essential to understand the potential risks of AI. If AI does not guarantee privacy, businesses will struggle to keep customers. 

5. Privacy 

As individuals and Businesses heavily rely on and work on the cloud it becomes essential to Protect customer data. 

Implementing Fair AI Practices 

Inclusive and Equity: Ensuring equal opportunities and treatment for all employees. 

Fairness: Actively work to eliminate biases and promote fairness in operations. 

Transparency: Maintaining openness about processes and decisions to build trust. 

Societal Impact: Considering the broader societal consequences of workplace actions. 

Continuous Assessment: Regularly evaluating and improving workplace practices for ethical integrity. 

Source : Deloitte Analysis

What is Bias in AI? 

AI bias happens when AI systems make unfair decisions or assumptions in their design leading to unfair outputs 

This can result from: 

Cognitive biases 

Lack of complete data 

Challenges 

Constant discovery of new biases. 

Human involvement in data creation 

Solutions 

Data and Algorithm Scrutiny 

Debiasing Strategies 

Human-driven Improvements: 

Decision-making and Collaboration 

Source : World Economic Forum

Bias in AI creeps through skewed data, biased algorithms, and the potential for manipulation. Representation bias is one of the biases that exists because of the lack of geographical diversity in image datasets of people which leads to an over-representation of certain groups over others. 

When using AI-generated images, it’s important to take note of gaps in training datasets that can lead to inaccurate representations. When using generative image AI, it is important to watch out for biases in image generation that reflect common stereotypes. 

As AI gets better at taking and changing pictures, we face some big questions about right and wrong. This part talks about how using AI in photography makes us think about issues like keeping things real, respecting privacy, and who really made an image. We’ll explore why these questions matter as AI becomes a bigger part of making art and capturing moments. 

Ethical Lens: Mitigating Bias in AI Photography for Responsible Representation 

1. Consent and Autonomy 

The manipulation of images using AI without the consent of the individuals involved directly infringes upon an Individual’s autonomy. Ethical AI frameworks emphasize the importance of consent. 

2. Privacy 

AI-powered image manipulation can easily breach privacy, especially when images are used or shared without permission, or when manipulated images create misleading or false representations of individuals. 

3. Accuracy and Misrepresentation 

The capacity of AI to alter images in highly realistic ways raises concerns about accuracy and the potential for misrepresentation. This includes the creation of deepfakes or manipulated content that can deceive viewers, harm reputations, or spread misinformation.  

While ethical considerations are crucial in ensuring the responsible use of AI in photography, its influence extends far beyond the realm of visual arts. In the workplace, AI’s impact on job quality is a complex and multifaceted issue, presenting both exciting opportunities and potential pitfalls. 

AI & the Future of Work: Will Robots Steal Our Jobs? 

Here are five ways in which AI is generally improving job quality: 

1. Automation of Routine Tasks  

It eliminates repetitive tasks and frees up time for thinking, creating, and innovating therefore increasing productivity and efficiency 

2. Enhanced Decision Making 

AI empowers you to make informed decisions with Data-driven insights 

3. Personalized Learning and Development 

AI  tailors your development to your unique needs and goals and Unlock your full potential with custom learning paths. 

4. Improved Work-Life Balance 

It focuses on what matters most with AI taking care of the rest, achieving a healthier, happier work-life balance with technology’s help. 

5. Creation of New Job Opportunities 

AI creates new roles demanding human skills that require human intelligence, creativity, and emotional intelligence areas where machines currently fall short. While we previously explored the ways AI can enhance productivity, decision-making, and learning, Let’s delve deeper into the potential downside of AI in the workplace and how it can be solved.

Balancing Automation and Human Work 

In a world where we heavily rely on AI intelligence, it is important to understand ethical AI in the work place. One of the primary concerns is the fear of Job displacement, therefore while we automate our task the focus should be more on how automation is going to add value to the work and make it efficient rather than replacing humans. Organizations should Promote a human-centered approach to avoid over-reliance on AI which results in Job insecurity and reduced human efficiency. 

Balancing Automation With Human Work In A Sustainable AI Era 

1. From Displacement to Upskilling 

 Partnering with educational institutions and training providers is crucial. By undertaking targeted programs that equip workers with the skills needed for the AI-powered workplace, can empower them to thrive rather than fear displacement. Continuous learning and adaptation become essential mantras in this evolving landscape. 

2. Collaborative Harmony 

We need to encourage collaborative work environments that facilitate seamless interaction and communication, ensuring humans remain at the heart of decision-making and problem-solving. 

3. Building Ethical Guidelines 

To ensure AI serves humanity, we need clear ethical guidelines for its development and deployment. These guidelines should address critical concerns like job displacement, bias, privacy, and accountability.  

4. Continuous Monitoring and Adjustment 

When we closely monitor and adjust AI implementation based on these assessments, we ensure it aligns with ethical principles and continues to benefit the workforce. 

Takeaways 

Meaning and Importance of Ethical AI 

Five Pillars of AI Ethics 

Addressing Bias in AI 

AI’s Impact on Job Quality 

Promoting Sustainable AI Practices 

Conclusion 

Establishing ethical AI in the workplace is crucial given the industry’s commitment to responsible AI development. In the age of automation, ethical AI is essential for guaranteeing job quality and fairness. 

[aiqod]

Generative AI: The next big thing

In the ever-changing world of technology, Generative AI has become one of the most transformative technologies of recent years, promising to reshape industries, enhance creativity, and revolutionize how we interact with machines. This blog delves into the world of Generative AI, Providing an overview of its growth, capabilities, and potential impact on various sectors. Generative AI has witnessed explosive growth in recent years, marked by the development of increasingly sophisticated models. In 2018, OpenAI released GPT-1, a significant milestone in natural language generation, and by 2020, GPT-3 had taken the world by storm with its 175 billion parameters, pushing the boundaries of what was previously thought possible when it came to text generation.

Generative AI achievements are driving innovation and transformation across various industries, in healthcare generative models are used for drug discovery and medical imaging, potentially saving lives and reducing research time also in the finance and creative arts, AI-powered trading algorithms use generative models to predict market trends and optimize portfolios. Generative AI helps startups, and brand-building innovative ideas push the limits of human creativity.

In this blog, we explore these challenges and how they set the way for the Generative – AI Revolution.

– Collaboration between Humans and AI: Integrating Generative AI into various industries, such as healthcare, finance, and creative arts, requires establishing effective collaboration between humans and AI systems. ensuring harmonious and effective cooperation is ongoing. Humans need to trust that AI systems are reliable and trustworthy and need to be able to communicate effectively. AI is often more complex and opaque and it may use different language and concepts than humans.

– Privacy and data: Generative AI models often require large amounts of data to operate effectively. Ensuring the security and ethical use of this data is a significant challenge. Finding the right balance between data use and protecting personal privacy is important. This involves obtaining consent to use data, anonymizing data, and ensuring transparency in data collection and processing to protect individuals’ privacy.

– Manual and Labor-Intensive Processes:  In the pre-AI era, work contexts defined by manual labor and labor-intensive processes AI, tasks that now seem routine were of labor-intensive and time-complex calculations, data analysis, and repetitive tasks rely heavily on human effort, slowing progress and limiting scalability. The golden opportunity for innovation and creative exploration is limited. Limited time and energy leave little room for brainstorming, experimenting, and finding new ideas.

–  Human Error and Inconsistency: Human involvement in various processes can introduce the risk of errors and inconsistencies in data entry, calculation, and decision-making common problems affecting accuracy and efficiency in optimizing resource allocation decisions based on data-driven insights. Resources are allocated efficiently to waste and maximize results.

–  Bias and Fairness: Generative AI models can unintentionally perpetuate bias present in the training data. To address bias and ensure fairness in AI-generated activities, some AI models are trained on data, and if that data is biased, the model will be biased as well. This can lead to the generation of harmful or offensive content. using debiasing techniques and diversifying training datasets can mitigate model bias.

Let’s see How Generative AI is helping to upgrade the remarkable technology with infinite potential.

– Design AI systems with humans in mind: Providing clear feedback and understanding to use the ability to override decisions when necessary, work and their capabilities and limitations establishing clear roles and responsibilities to avoid confusion and conflict to evaluate human-AI collaboration on an ongoing basis to identify and address any challenges that arise.

– Invoice processing and supplier management: Gen AI can be used to automate the entire invoice processing process, from data extraction to validation and approval. This can enable P2P staff to focus on more strategic tasks, analyze supplier data and identify opportunities for cost savings and efficiency improvements. It can be used to review and manage contracts, ensuring they comply with all relevant regulations. analyze spending data and identify trends and patterns. This information can be used to make better budgeting and purchasing decisions.

– Obtain consent to use and Anonymize data: Consent can be obtained through explicit opt-in mechanisms, such as checkboxes or consent forms. It is also important to provide clear and concise information about how the data will be used so that users can make informed decisions about whether or not to consent. This can be done through a variety of techniques, such as removing names, addresses, and other identifying information. It is important to note that anonymization does not completely guarantee privacy, but it can make it more difficult to identify individuals from the data.

– Intelligent Document Processing (IDP): Gen AI to automate the processing of documents. It extracts data from documents, validates the data, and classifies the documents into the appropriate categories. DPA can be used to automate a wide range of document processing tasks, such as loan document digitization, Legal document digitization, and Logistics document digitization. This can help to reduce the time it takes for loan applications to be processed and approved, improve the organization and accessibility of the law firm’s contracts and efficiency of the shipping company’s operations, and reduce the risk of errors.

– Generative-AI data-driven resource allocation: Automating data entry and calculation tasks can free up human workers to focus on more complex strategic tasks, and it can also reduce the risk of errors. Gen AI can analyze large amounts of data and identify patterns and trends that would be more difficult or impossible for humans to identify on their own. Optimizing resource allocation decisions on a variety of factors such as demand, costs, and constraints. From the healthcare, manufacturing, and financial industry generative AI is being used to optimize.

How Gen-AI helps businesses scale and thrive: In a sector that is labor-intensive and manual, The emergence of Generative AI Intelligent Document Processing (IDP) has emerged as a game-changing solution by efficiently processing unstructured data, resulting in improved customer satisfaction, enhanced process efficiency, and better overall business performance. AI-powered automation optimizes workflows by managing routine tasks. That allows employees to focus on value-added activities, speeding up processes and improving overall efficiency.

Generative AI Advantages: Transforming Industries and Redefining Creativity

Enhancing Creativity: Generative AI Acts as a catalyst for human creativity, providing new ideas and fresh perspectives that can spark innovative thinking. 

Efficiency Boost: With the help of technology, we can speed up tasks that would otherwise take quite a bit of time. For example, creating multiple design variations or generating text for marketing campaigns can be done much more quickly with the assistance of these tools. Release the up time and resources to focus on other Important aspects of the project.

Exploring possibilities: Personalization- Exploring multiple options can be a time-consuming task, but with the help of technology, we can rapidly generate ideas and uncover unique chances. That can be particularly helpful for design projects or marketing campaigns where creativity and originality are crucial to quickly exploring multiple options, finding the best solutions, and making the most of time and resources.

Content creation: Copywriters and content creators are collaborating with AI to draft Compelling articles, advertisements, and marketing content that resonate with specific target audiences.

New Business opportunities– AI opens the way for innovative business models and revenue streams.

Conclusion:

Generative AI has opened New frontiers in Creativity and Creation. It is capable of various creative and practical tasks, but it is essential to approach its abilities with a critical and responsible spirit. Continuing research, ethical considerations, and collaborative efforts between Humans and AI will determine How AI can shape our World in years to come. Generative AI is a fascinating field pushing the boundaries of what is possible with technology. As we continue to explore its potential, it’s clear that Generative AI will play a significant role in shaping the future.

[blogbyahad]

Ethical AI and Bias Mitigation: Ensuring Fairness and Accountability in AI Systems

As artificial intelligence (AI) becomes more ingrained in our everyday lives, concerns about fairness, accountability, and bias are increasingly coming to the forefront. Ethical AI focuses on developing and deploying AI systems that are not only effective but also aligned with ethical principles, ensuring they are fair, transparent, and do not reinforce or exacerbate societal inequalities.

Understanding Bias in AI

Bias in AI arises when the data used to train models reflects existing prejudices or inequalities in society. These biases can manifest in various ways, such as gender, racial, or socioeconomic disparities. When AI systems are trained on biased data, they can perpetuate these biases in their decision-making processes, leading to unfair outcomes.

For example, if a hiring algorithm is trained on data from a company that historically hired more men than women, it might favor male candidates over equally qualified female candidates. Similarly, facial recognition systems have been found to be less accurate in identifying individuals with darker skin tones, leading to potential discrimination in law enforcement and other applications.

The Importance of Ethical AI

Ethical AI is about ensuring that AI systems are developed and deployed in a way that respects human rights, promotes fairness, and avoids harm. This involves several key principles:

Fairness: AI systems should be designed to treat all individuals equitably, regardless of their gender, race, or socioeconomic status. This means actively working to identify and mitigate biases in AI models and the data they are trained on.

Transparency: AI systems should be transparent, with clear explanations of how decisions are made. This helps build trust and allows users to understand and challenge the outcomes if necessary.

Accountability: There must be clear lines of responsibility for the decisions made by AI systems. Developers and organizations must be accountable for the ethical implications of their AI models.

Privacy: AI systems must respect user privacy and ensure that personal data is handled securely and ethically. This is especially important in applications like healthcare, where sensitive information is involved.

Strategies for Bias Mitigation

Addressing bias in AI requires a multifaceted approach, starting from the data collection process to the deployment of AI models:

Diverse Data: One of the most effective ways to reduce bias is by using diverse and representative datasets. This ensures that AI models are trained on a broad range of experiences and perspectives, minimizing the risk of reinforcing existing biases.

Bias Detection Tools: Various tools and techniques can be used to detect and measure bias in AI models. These tools help developers identify potential issues before the AI system is deployed, allowing them to make necessary adjustments.

Algorithmic Fairness: Algorithms can be designed with fairness constraints, ensuring that they do not disproportionately favor one group over another. For example, techniques like reweighting or adversarial debiasing can be used to adjust the model’s training process to promote fairness.

Human Oversight: Human oversight is crucial in AI decision-making, especially in high-stakes situations. By involving humans in the loop, organizations can ensure that AI decisions are reviewed and, if necessary, corrected before they have real-world consequences.

The Role of Regulations and Standards

As AI becomes more pervasive, the need for regulations and standards to ensure ethical practices is growing. Governments and international organizations are beginning to develop frameworks that address the ethical challenges posed by AI. For instance, the European Union’s General Data Protection Regulation (GDPR) includes provisions that address AI fairness and transparency.

Moreover, industry standards are being developed to guide the ethical use of AI. These standards provide best practices for AI development and deployment, helping organizations align their AI initiatives with ethical principles.

[sathya32]

The Next Step in Data Science: A Glimpse into the Future

For years, data science has been transforming industries, spurring creativity, and resolving challenging issues. The topic is still developing, so there may be even more fascinating advancements in the future. Now let's examine four important areas where data science is currently moving.

1. Machine learning that is automated (AutoML) The field of data science is expected to be significantly impacted by automated machine learning, or AutoML. Traditionally, choosing the best algorithms, adjusting parameters, and assessing performance needed a great deal of experience when creating machine learning models. By automating many of these procedures, autoML streamlines this process and makes it more user-friendly for non-experts.

Hyperparameter tuning, feature selection, model training, and data pretreatment can all be handled by autoML systems. With the democratization of machine learning, a greater number of individuals and organizations may leverage the potential of data without requiring a profound comprehension of its technical nuances. It shortens the development cycle, freeing up data scientists to concentrate on more intricate and imaginative facets of their jobs.

2. XAI, or explainable AI

Knowing how AI and machine learning models make judgments is essential as these models get more complex. This need is met by Explainable AI (XAI), which offers insights into these models' internal operations. Particularly in fields where judgments can have a big impact, like healthcare, banking, and law, this transparency is crucial for fostering trust.

The goal of XAI is to improve interpretability of AI models without compromising performance. The "black box" aspect of many machine learning models is explained by methods like LIME (Local Interpretable Model-Agnostic Explanations) and SHAP (Shapley Additive Explanations). The use of XAI guarantees that stakeholders can validate and have faith in the decisions made by AI systems by offering comprehensible and transparent explanations.

3. Data science and edge computing

The convergence of edge computing and data science is being driven by the growth of Internet of Things (IoT) devices and the requirement for real-time data processing. Processing data close to its source is known as edge computing, as opposed to depending only on centralized cloud servers. This method uses less bandwidth and latency, which makes it perfect for applications that need quick replies.

When edge computing and data science are combined, smart devices can evaluate data locally and take immediate action based on their findings. For instance, in smart cities, real-time analysis of vehicle flow by traffic cameras outfitted with edge computing allows for the optimization of traffic signals to minimize gridlock. Similar to this, wearable technology in healthcare can monitor vital signs and give prompt feedback, enhancing patient care.

4. Responsible and ethical AI

Ethical considerations are becoming more and more crucial as AI technology proliferate. Within the data science community, there is an increasing emphasis on making sure AI systems are impartial, fair, and considerate of privacy. It takes a combination of technological advancements, legal structures, and moral principles to address these issues.

Including fairness and bias detection in the model-development process is one strategy. Adversarial debiasing and fairness-aware machine learning are two methods that assist in locating and reducing biases in data and models. Organizations are also developing best practices and ethical standards for AI development and application, encouraging openness, responsibility, and diversity.

Data science has a promising and bright future ahead of it. With developments in edge computing, ethical AI, explainable AI, and AutoML, the industry is ready to take on new problems and produce even more creative solutions. These advancements will improve data science's capabilities while also increasing its transparency, responsibility, and accessibility. Data science will surely continue to advance and produce ground-breaking discoveries and revolutionary shifts in a variety of industries as time goes on.I appreciate your precious time, and I hope you have an amazing day.

[ixnai]

Bias in AI is not a bug; it’s a feature. This assertion may seem counterintuitive, but it is rooted in the very architecture of machine learning models. At the heart of AI systems lies the concept of inductive bias, a fundamental principle that guides the learning process. Inductive bias is the set of assumptions that an algorithm uses to predict outputs given inputs that it has not encountered before. Without it, generalization from training data to unseen data would be impossible.

Consider the Yankee Doodle algorithm, a hypothetical AI model trained to generate music. Its inductive bias might include assumptions about melody, harmony, and rhythm based on the training data it was exposed to. If the training data predominantly features Western classical music, the algorithm will naturally exhibit a bias towards generating compositions in that style. This is not a flaw; it is an inherent characteristic of the learning process.

However, the pitfalls of AI bias become apparent when these biases lead to unintended consequences. For instance, if the Yankee Doodle algorithm were tasked with composing music for a culturally diverse audience, its Western-centric bias could result in outputs that fail to resonate with listeners from different backgrounds. This is where the challenge lies: distinguishing between beneficial inductive biases that enable learning and harmful biases that propagate stereotypes or discrimination.

To navigate these pitfalls, it is crucial to adopt a multi-faceted approach. First, diversify the training data. By exposing AI models to a wide range of inputs, we can mitigate the risk of overfitting to a narrow subset of data. Second, implement fairness-aware algorithms that explicitly account for potential biases during the learning process. Techniques such as adversarial debiasing and re-weighting can help ensure that the model’s outputs are equitable across different groups.

Moreover, continuous monitoring and evaluation of AI systems are essential. Bias detection tools can identify and quantify biases in model outputs, providing valuable insights for iterative refinement. Transparency in AI decision-making processes also plays a critical role. By understanding the rationale behind an algorithm’s predictions, stakeholders can make informed decisions about its deployment.

In conclusion, bias in AI is an intrinsic feature, not a defect. It is a byproduct of the inductive biases that enable learning. However, by recognizing and addressing the potential pitfalls of AI bias, we can harness the power of these systems while minimizing their adverse effects. The key lies in a balanced approach that combines diverse data, fairness-aware algorithms, and ongoing evaluation. Only then can we ensure that AI serves as a tool for progress rather than perpetuation of existing inequities.

[subo797112]

Data science problems and their solutions - A brief guide

Data science is a broad field and in demand in today's market. In this competitive environment, businesses are generating tons of data that are to be cleaned, manipulated, data-modelling, analysis of data, etc. and all such tasks are to be done by the team of data scientists. So while performing such works, professionals can face some common issues that need to be fixed as soon as possible. So, here are some list of common problems and their solutions that every data scientist might face in the midst of their on-going project.

1. Bad and incomplete data quality.

SOLUTION: Information cleaning

Data cleaning involves locating and fixing problems with the dataset, such as getting rid of duplicates, dealing with missing numbers, and fixing errors. This makes sure that the information utilised for analysis is correct and trustworthy, resulting in more insightful findings and improved model performance.

2. Absence of Data: There is not enough evidence to draw conclusions.

SOLUTION: Data gathering is the solution.

The remedy to a data shortage is to collect more significant data. This may involve a number of techniques, including data collection via surveys, online scraping, or collaborations with data suppliers. Additional data improves the authenticity and effectiveness of analysis.

3. Overfitting: Complex models with poor predictions.

SOLUTION: Employ simpler models as a solution.

Insufficient generalisation to new data is the result of overfitting, which happens when a model is very complicated and matches the training data too closely. One can use regularisation techniques to keep the model from getting too complex or choose simpler models to reduce overfitting.

4. Interpretability: Unable to describe complicated model choices

SOLUTION: Employ interpretable models as a remedy

It is advisable to use models that provide transparency in their decision-making processes when interpretability is important. Compared to complicated deep learning models, linear models like decision trees are frequently easier to understand. Techniques like feature importance analysis can also be used to better understand model choices.

5. Data Privacy: Considering Privacy and Utility.

SOLUTION: Confidentiality is a solution.

Techniques for preserving data utility while protecting sensitive information include data masking, encryption, and aggregation. By doing this, the data's analytical value is preserved without compromising people's right to privacy.

6. Bias & Fairness: Unfair forecasts are the result of biassed data.

SOLUTION: Solution: Reduce bias

In order to address discrimination, biases in data and algorithms must be found and corrected. To ensure fair and equal outcomes, this may include re-sampling underrepresented groups, adjusting decision thresholds, or using specialized debiasing techniques to ensure fair and equitable outcomes.

7. Scalability: The ability to manage big datasets.

SOLUTION: Big Data Tools are a solution.

Big data technologies like Apache Spark or Hadoop can be used to address scalability issues. By dividing the effort among clusters of machines, these platforms make it possible to handle and analyse large datasets in an effective manner.

8. Model selection, or choosing the appropriate algorithm.

SOLUTION: Employ evaluation measures as a solution.

The best algorithm must be chosen after thorough consideration. To determine how well a model works on a particular task, use appropriate evaluation metrics like accuracy, precision, recall, and F1-score. This methodical decision-making guarantees that the chosen algorithm matches the goals of the task.

9. Resources are constrained, such as computing power.

SOLUTION: Cloud Services as a solution

Cloud computing services offer scalable and affordable alternatives when computing resources are limited. Data scientists may work on resource-intensive projects without being constrained by hardware, thanks to the access to powerful computing resources made possible by cloud platforms like Amazon, Azure, or Google Cloud.

10. Data Governance: Assuring conformity.

SOLUTION: Strong Policies

Organizations should set up thorough rules and procedures to fulfill data governance requirements. These guidelines include data collection, storage, access, sharing, and disposal, guaranteeing adherence to all applicable laws and professional standards. For effective data governance, regular audits and the application of these policies are crucial.

These solutions deal with typical data science problems and advance model performance, ethical data handling, and data-driven decision-making.

[machine-saint]

the appearance of objectivity and the opacity of modern ML models is certainly a good argument, but the category of post i was subposting in the original doesn't go that far; it just goes "the training data is biased, therefore the output will be" and doesn't address the fact that human training data is also biased and the research into debiasing techniques

but anyway: a human can be asked to give their reasoning. what they say may or may not be their actual reasoning; people generally know that giving an explicitly racist reason is bad, and even if they're not intentionally dishonest they have an incentive to lie to themselves. an AI/ML model doesn't act differently when it's being tested as it does "in real life". you can ask it to give a million decisions, tweak some parameters, and then ask it to make another million decisions. you can ask it hypotheticals (what's the highest loan offer you would have given this person) and know that it won't lie to make itself look better, because it can't.

also, i think bringing up 'algorithms' is not relevant; you mentioned company policy, but that itself is algorithmic reasoning! if a decision is made entirely by policy with no room for judgment on the part of the human, then the human is just carrying out an algorithm specified in policy. of course, modern deep learning networks are notably black-box, but that's a characteristic of the networks, not of machine learning in general. i think this is an important distinction to make.

that's not to say that i think things like that startup doing AI-based "personality evaluation" are good, but to say that i would be surprised if a human evaluator didn't also have many of the same biases. like, "make sure your video call background isn't super disorganized" is common video interview advice!

that's not to say that i think that algorithmic approaches are inherently less biased in cases where there's biased training data, or to discount the effects of the fact that a computer can make these decisions thousands of times faster than humans.

AI systems that are trained on a corpus of human-produced data shouldn't be used because they'll always be racist because of their biased training data. unlike humans, who are well-known for not having unconscious biases.

[psychbuilder]

Decision making can be improved through observational learning

Yoon, H., Scopelliti, I. & Morewedge, C.

Organizational Behavior and 

Human Decision Processes

Volume 162, January 2021, 

Pages 155-188

Abstract

Observational learning can debias judgment and decision making. One-shot observational learning-based training interventions (akin to “hot seating”) can produce reductions in cognitive biases in the laboratory (i.e., anchoring, representativeness, and social projection), and successfully teach a decision rule that increases advice taking in a weight on advice paradigm (i.e., the averaging principle). These interventions improve judgment, rule learning, and advice taking more than practice. We find observational learning-based interventions can be as effective as information-based interventions. Their effects are additive for advice taking, and for accuracy when advice is algorithmically optimized. As found in the organizational learning literature, explicit knowledge transferred through information appears to reduce the stickiness of tacit knowledge transferred through observational learning. Moreover, observational learning appears to be a unique debiasing training strategy, an addition to the four proposed by Fischhoff (1982). We also report new scales measuring individual differences in anchoring, representativeness heuristics, and social projection.

Highlights

• Observational learning training interventions improved judgment and decision making.

• OL interventions reduced anchoring bias, representativeness, and social projection.

• Observational learning training interventions increased advice taking.

• Observational learning and information complementarily taught a decision rule.

• We provide new bias scales for anchoring, representativeness, and social projection.

The research is here.

[carvalhais]

As concerning as this embedded cultural prejudice might seem, the drive for AI to succeed means that findings like these are seen less as a problem and more as an opportunity. Having identified bias mathematically, the thinking goes, it can be mathematically corrected by shifting values to remove it. All problematically gendered terms can be zeroed along the gender axis in data space so they don’t lean in a male or female direction. Lo and behold, algorithms go one better than humans because their prejudices can be instantly corrected. However, the drawback to this seductive proposition is that it’s unclear what other distortions this so-called correction may amplify or introduce. This practice of mathematical ‘debiasing’ forges forward in blissful ignorance of the nuances of intersectionality, that is, of the contextual interplay of oppressions like gender, race, disability and so on. And because of the hubris of the AI field, it does so without seeking out the voices of those who are directly affected. Dan McQuillan. 2022. Resisting AI: An Anti-Fascist Approach to Artificial Intelligence. Bristol: Bristol University Press.

[un-enfant-immature]

RealityEngines.AI becomes Abacus.AI and raises $13M Series A

RealityEngines.AI, the machine learning startup co-founded by former AWS and Google exec Bindu Reddy, today announced that it is rebranding as Abacus.AI and launching its autonomous AI service into general availability.

In addition, the company also today disclosed that it has raised a $13 million Series A round led by Index Ventures’ Mike Volpi, who will also join the company’s board. Seed investors Eric Schmidt, Jerry Yang and Ram Shriram also participated in this oversubscribed round, with Shriram also joining the company’s board. New investors include Mariam Naficy, Erica Shultz, Neha Narkhede, Xuezhao Lan and Jeannette Furstenberg. 

This new round brings the company’s total funding to $18.25 million.

Abacus.AI’s co-founders Bindu Reddy, Arvind Sundararajan and Siddartha Naidu (Image Credits: Abacus.AI)

At its core, RealityEngines.AI’s Abacus.AI’s mission is to help businesses implement modern deep learning systems into their customer experience and business processes without having to do the heavy lifting of learning how to train models themselves. Instead, Abacus takes care of the data pipelines and model training for them.

The company worked with 1,200 beta testers and in recent months, the team mostly focused on not just helping businesses build their models but also put them into production. Current Abacus.AI customers include 1-800-Flowers, Flex, DailyLook and Prodege.

“My guess would be that out of the hundred projects which are started in ML, one percent succeeds because of so many moving parts,” Reddy told me. “You have to build the model, then you have to test it in production — and then you have to build data pipelines and have to put in training pipelines. So over the last few weeks even, we’ve added a whole bunch of features to enable putting these things to go into production more smoothly — and we continue to add to it.”

Image Credits: Abacus.AI

In recent months, the team also added new unsupervised learning tools to its lineup of pre-built solutions to help users build systems for anomaly detection around transaction fraud and account takeovers, for example.

The company also today released new tools for debiasing data sets that can be used on already trained algorithms. Automatically building training sets — even with relatively small data sets — is one of the areas on which the Abacus team has long focused, and it is now using some of these same techniques to tackle this problem. In its experiments, the company’s facial recognition algorithm was able to greatly improve its ability to detect whether a Black celebrity was smiling or not, for example, even though the training data set featured 22 times more white people.

Image Credits: Abacus

With today’s launch, Abacus is also launching a new section on its website to showcase models from its community. “You can go build a model, tweak your model if you want, use your own data sets — and then you can actually share the model with the community,” Reddy explained, and noted that this is now possible because of Abacus’ new pricing model. The company has decided to only charge customers when they put models into production.

Image Credits: Abacus.ai

The next major item on the Abacus roadmap is to build more connectors to third-party systems so that users can easily import data from Salesforce and Segment, for example. In addition, Reddy notes that the team will build out more of its pre-built solutions, including more work on language understanding and vision use cases.

To do this, Abacus has already hired more research scientists to work on some of its fundamental research projects, something Reddy says its funders are quite comfortable with, and more engineers to put that work into practice. She expects the team will grow from 22 employees today to about 35 by the end of the year.

RealityEngines launches its autonomous AI service