What is Human in the Loop?

The power of machine learning (ML) is quite amazing. It allows us to overcome seemingly hopeless obstacles. It can do tedious work rapidly, and find unexpected solutions. ML often has an uncanny ability to predict and recognize important features. However, at other times it may tell you that a dragonfly is actually definitely a manhole cover.

What is Human in the Loop?

ML models make mistakes, some of which may seem profoundly naïve to a human. Human in the Loop (HitL) is a strategy used in ML to deal with this. It combines human intelligence with the power and speed of ML. These models interact with humans to constantly improve.

HitL can work in various ways, depending on the type of model and its applications. It can be as simple as developers retraining models when they identify problems. However, more advanced systems can integrate this feedback into consumer products. This allows crowdsourcing feedback to improve the accuracy of models.

Other “active learning” systems identify areas of uncertainty. These systems will provide examples where tagging would provide the greatest improvements. This feedback system allows the rapid training of artificial intelligence (AI) models.

The Pareto Principle

Vilfredo Federico Damaso Pareto was an Italian philosopher and economist who lived from 1848 to 1923. He made significant contributions to economic theory, particularly in studying income distribution. One observation he made was that in many countries, 20% of the population holds 80% of the wealth. This inequality appears in other settings as well. In studying other systems he observed that often 80% of effects come from 20% of causes. This has become a popular principle in economics.

People have applied this principle in many different areas. Computer scientists and software developers have applied this concept to software development projects. In large programming projects, the hardest 20% of the codebase generally takes 80% of the time.

The Pareto principle has also been used to explain the way HitL models balance machine and human intelligence. The idea is that ML models may have trouble getting above 80% accuracy. The hardest 20% of examples are responsible for 80% of the errors made. By combining human and machine intelligence, humans can address the difficult few. This still allows for an 80% reduction in human work, with greater improvements as the model learns from feedback.

The Benefits of Human in the Loop

Human in the Loop strategies can greatly improve ML applications. They leverage both the efficiency of ML and the accuracy of human input.

Rapid Iterative Development

Forms of HitL are common in modern ML. Most models are not perfect on the first try. As a result, retraining models with new labeled data and improved examples is common. This is often done in a way that requires developers to rerun workflows from scratch. These systems do not provide the flexibility needed to do iterative design in an efficient way.

Some models can provide feedback on the examples they need in order to improve their accuracy. This is a process called active learning. Developers can also add active learning to production products, enabling end-users to contribute to training. This allows the iterative development process to automatically integrate user feedback.

Developing systems that expedite the process of iterative design and retraining is essential. Platforms that streamline this process have great potential to improve the effectiveness of ML in real-world applications. Though HitL approaches are being used today, there are few general-purpose platforms for this process. However, streamlined solutions for natural language processing (NLP) data tagging are available from swivl.

Greater Accuracy Than ML or Humans Alone

The key advantage of HitL strategies is leveraging the advantages of both human intelligence and ML. These strategies allow algorithms to continually improve based on human input. By gaining new inputs, the accuracy of these models continually approaches human-level accuracy.

When to Use Human in the Loop

HitL is an important strategy for pushing ML past a critical point of accuracy. ML applications perform very well for many tasks, including extracting broad trends and estimating outcomes from large datasets. In many of these uses, it is unimportant that the results are always precise.

Critical Systems

However, other use cases such as autonomous vehicles and medical applications require far more precision. In these cases, uncertainty has truly dire consequences. In cases like this, HitL may provide the needed boost to make ML practical. Requiring a driver to interact with an autonomous vehicle may seem less sleek, but may provide the needed edge.

Lacking Data

HitL is not only useful for “mission-critical” applications. It can greatly improve models that have little available data or few examples for specific categories. The effectiveness of ML is limited by the quality of data on which it is trained. If you put low-quality data in, you will get low-quality predictions.

Getting high-quality data is no easy task. Data cleaning and munging are generally the most time-consuming steps in all data science workflows. This is why it is essential to find efficient ways to tag and classify your datasets. Getting human input from users to tag new data is a great way to improve available datasets.

In addition, in some applications of ML, rare cases may appear. This variance is not easy for ML models to respond to. Humans are far better at adapting to this sort of change. In cases like this HitL is a key strategy for allowing ML to perform well.

HitL is often a part of the training and tuning processes. ML algorithms are rarely perfect on the first try and require tweaking and re-training by humans. This has generally been done by data scientists, who may need to manually re-run complex workflows. However, providing no-code solutions and streamlining this process to allow end-user input to directly improve models has great potential.

Allowing customer support reps and other non-technical employees to contribute to this process is essential. This saves valuable time for developers and allows those who will use the model to do the training.

Human in the Loop is the Future

As ML models become more common and more specialized, the need for training will grow. Being able to rapidly and efficiently incorporate feedback into models will be hugely beneficial. As tools and techniques to do this become more common, model accuracy will improve. The largest barrier to this now is simply having tools to efficiently incorporate this. This barrier will continue to shrink.

Businesses often envision ML and AI as cure-alls that can replace all human work. The reality is that, while ML has impressive results, humans are still essential. The areas where ML falls short require humans to bridge the gap. The combined power of humans and AI has enormous potential. Streamlining AI-human interactions is essential to this process.

Examples

• Facebook uses HitL for facial recognition in photos. Their DeepFace algorithm achieves near human-level accuracy. By asking users to confirm or reject a suggested guess, users improve the model. This allows Facebook to achieve a success rate of 97.35%!
• Coca-Cola uses a HitL strategy for its rewards program MyCokeRewards. In order to simplify user input, they built an optical character recognition (OCR) system using a convolutional neural network (CNN). This algorithm had to identify codes printed on multiple surfaces. They initially trained their model on simulated images. However, to improve accuracy, real-world data was needed.
  The app has users take pictures of their codes. The app makes guesses with confidence intervals for each character. If the code fails, the application highlights low-confidence characters for users to correct. The user’s input trains the model, improving accuracy. This has created a model with over 96% accuracy.
• Customer support teams using the swivlStudio take advantage of the power of Human-in-the-Loop. The Intelligent Agent is the front-line of customer interactions. It provides autonomous responses to most users’ questions. However, it also identifies the cases where the the algorithm has less than a 70% confidence level to a question or request.  The customer needs a human interaction. swivl can then push to a human, who can easily take over the conversation, and access customer data from integrated CRM to deliver information.
• The swivl team believes the future of AI is a partnership between machines and humans –  and deliver needed the human touch.

Please visit swivl,  and meet our Intelligent Agent, Hoover.  We invite you to learn more about our solutions to use automation,  personalization, and HitL to improve customer experience.

Please join our newsletter to keep up with all the latest trends on Artificial Intelligence powered Chatbots.

‍