For years, even before ChatGPT put artificial intelligence (AI) firmly at the forefront of the public imagination, AI slowly took hold in every industry, from medicine to aerospace. However, the technology has not lived up to its potential. far cry.
a Recent research We found that only 11% of companies using AI are seeing financial benefits. Even tech giants are struggling. IBM’s $20 billion diagnostic AI system, Watson Health, diagnosed cancer more accurately than doctors in clinical trials, but failed in the field.have become Commercial and reputational damage For Famous American Multinational Corporations.
IBM commissioned a number of engineers to develop Watson Health, so the failure cannot be attributed to lack of technical expertise. Our extensive research into AI development challenges in a variety of commercial environments shows some surprising causes. Watson Health was developed and brought to market in a way that works well for traditional IT but not for AI. This is due to the fundamental difference between traditional software and AI. The former processes data, but AI continuously learns from data, improves over time, and even exceeds its intended capabilities if properly nurtured.
We argue that best parenting styles and similar practices can accelerate the development of AI. We prescribe an AI development approach based on nurture and learn. We see this as a key factor in the success of our 200+ AI projects for industrial and other customers.
1. Adopt early and learn from mistakes
Children learn how to ride a bike by stepping on the pedals instead of watching an educational video while riding. Valuable lessons learned with every painful fall – Then the magic happens.
The same logic applies to AI. Many companies, such as IBM, believe they need to collect vast amounts of data to perfect their algorithms before deployment. This is irrelevant. By making AI work in the real world, rather than isolating it in a controlled environment, it can generate more data and feed it back into the development process.
While early adoption is inherently risky, it also sets in motion an ongoing feedback loop in which the algorithm is enhanced with new data. In addition, it is important that the data arise from both standard and difficult or atypical situations, and collectively support holistic AI development.
ChatGPT is a good example. The chatbot had many reasons to get ahead of its competitors, but with his ears still wet, he was released to the public by OpenAI in November 2022. Either way, this bet paid off. ChatGPT became a global phenomenon that not only spooked companies like Google’s Bard, but also gained millions of users in its early release, generated a huge amount of data OpenAI released an improved version of the bot, GPT-4, just a few months later.
Another example is grammatically. Refining the writing support system with user feedback has proven its continuous writing ability. AI improvement and adaptationEspecially in areas of language that are complex and context-dependent.
ApodigiAs a frontrunner in the digitalization of the pharmacy business, of An AI-supported pharmacy app launched in June 2020 that allows you to learn while working. The app, called Treet, suggests medication based on a doctor’s prescription, which is reviewed and adjusted by a pharmacist. The pharmacist’s response becomes a continuous feedback stream, fine-tuning the algorithm and contributing to better recommendations that address the complex needs and preferences of each individual patient.
In response, IBM developed and tested Watson Health extensively in the lab and brought diagnostic tools to market without incorporating continuous learning from real-world data. This traditional build, test, and deploy process proved inadequate for AI training.
2. Development of safety mechanism
AI development requires safety mechanisms to protect consumers and protect reputations. Simulator environments such as AILiveSim allow you to safely and comprehensively test full-scale AI systems before deploying them in the real world.
Tesla, meanwhile, runs a new version of its self-driving software in the background while a human drives the car. Decisions made by the software: Turning the steering wheel is compared to the driver’s judgment. Any significant deviations or anomalous decisions are analyzed and the AI is retrained if necessary.
AI developed for creative applications probably needs stronger guardrails. In the same way that children interact with bad company and learn undesirable habits, AI can be exposed to training data full of bias and discriminatory content.
To get ahead of this, OpenAI takes an approach called e.g. hostile training This is to train AI models to not be fooled by malicious inputs from attackers. This method exposes the chatbot to hostile content that threatens to overcome the bot’s standard constraints, allowing the chatbot to recognize malicious content and avoid falling for it in the future. .
3. Capture user behavior
In an ideal AI development cycle, The developer records all user reactions and actions to help further develop the algorithm without questioning the accuracy or value of its recommendations and predictions. For example, the Netflix AI Content Recommender simply records whether a user has watched recommended content and for how long.
Kemira is a company specializing in sustainable chemical solutions for water intensive industries such as paper and pulp production. machine learning based system Detect production problems before they occur and provide recommendations to prevent breakdowns before they occur. The AI system learns from responses and makes better recommendations in the future.
Had the Watson Health developers followed this principle, they would have achieved better results. R.Rather than programming an algorithm that asks doctors to rate their AI-generated recommendations, the system could have been trained to simply record doctors’ prescriptions. Integrating Watson Health into patient information systems also incorporates feedback loops for ongoing training based on real-world cases and patient outcomes.
User feedback provides excellent training data for specific focused vertical applications. Jasper is one such example. AI-powered content generator that learns from what users say Suggested text corrections.
User behavior can be translated into feedback at every stage of the AI’s learning process. His learning process consists of three parts. Creation of teaching materials. teaching. Collect performance feedback. First, the developer collects labeled data to train her AI. AI performance is then compared to desired outcomes or performance metrics. Finally, feedback is collected, fed back into the training process, and iterate.
Data, especially labeled data, has become a key asset for AI companies. Instead of hiring humans to label data, developers should think about how to automate the process as much as possible. For example, connecting a vehicle’s front camera feed to the steering wheel can automatically label winding roads and feed the feed into an AI model that learns to drive the car on complex routes.
In fact, developers have to deploy many automatic data collectors. Design explicit feedback loops for learning at scale. In the driving assistance development example above, Many vehicles are capable of a wider range of situations than few. When a vehicle cuts in front of Tesla, it triggers the upload of video seconds before the event. The system feeds the video into Tesla’s deep neural network, which learns various signals, such as gradual movement toward a lane divider, to anticipate interruptions and take appropriate action, such as slowing down.
In contrast, traditional auto companies are often stereotyped, developing and deploying driver assistance software with little automated feedback collection or data updates.
4. DESIGNED FOR CONTINUOUS LEARNING AT LARGE
Just like children don’t stay in kindergarten forever, AI training methods should be ongoing. Upgraded. However, AI developers too often focus on the latest developments in AI algorithms and individual use cases rather than designing systems to cover a large number of use cases and data streams.
For example, consider Kemira, a company that specializes in sustainable chemical solutions for water-intensive industries such as paper and pulp production. that is, machine learning based solution Uncover causal relationships between variables in the papermaking process and identify process conditions and phenomena that can lead to production failures and quality issues. As the system accumulates In addition to gaining insight into the root cause of potential instability, Generate actionable risk mitigation recommendations For paper machine operators. To ensure scalability, the system is cloud-based and uses MLOps for model retraining governance, enabling expansion to more use cases.
A key element in designing learning at scale is a system architecture that automatically collects feedback, updates a large number of AI models frequently, and generates simulated training data. Finnish energy company Suur-Savon Sähkö has developed an AI forecasting method that: Learn from historical and real-time consumption data to Improve the efficiency of energy production and the accuracy of heating supply temperature prediction by more than 50%..
Going one step further, companies can develop simulation environments that generate synthetic data and enable faster development cycles. For example, Tesla takes data from its vehicles and feeds it into simulators that simulate complex traffic. This will generate new synthetic training data.
AILiveSim has developed a parametric and domain independent system. Simulation environment Supports mechanical applications in automobiles, autonomous ships, and autonomous mining. Simulators allow companies to build prototypes and validate concepts. Create a synthetic data set for training an AI system. Algorithm debugging and optimization. We test and validate our products. Accelerate the development of machine learning systems by acquiring data and testing rare, real-world cases.
In summary, ohOrganizations that switch to a growth mindset and adopt the continuous learning method described above are more likely to create AI solutions that are well-suited for a rapidly evolving world. By using a continuous stream of data and feedback to nurture algorithms, organizations can ensure their products and services are agile, secure and relevant.
