Integrating the Human: Look Beyond the Algorithm's Plate — Get to Context

AI systems are departing from the past. We first hard-coded expert knowledge, then we jumped into data-driven approaches, and now we are departing from the past again. AI systems are no longer confined to rigidly predefined knowledge or only the patterns in the training data, but they can adapt and grow in tandem with human input.

Users are giving their input and engage in a collaborative dialogue with the AI. This collaborative dynamic finds its roots in the concept of Language Models (LMs), where users interact with the AI.

Rather than simply acting as bystanders or stepping in only when issues arise, users now play an integral role in shaping AI outcomes from the outset. By tapping into the collective knowledge and intuition of users, AI systems can potentially unlock new levels of accuracy, responsiveness, and adaptability. The synergy between human input and AI capabilities amplifies the potential for innovation, pushing the boundaries of what can be achieved. But beyond that: we can foster more adoption of AI by bringing the human into the loop.

A Quick Sneak-Peek into LLMs

Quantum physics transformed into a Shakespearean sonnet. Trade theory unraveled by a pirate on the high seas. A delightful children's tale starring a dinosaur venturing into outer space. It's no secret that people have taken immense pleasure in challenging modern chatbots to produce text that is truly out of the ordinary.

But here's the twist: these playful experiments have yielded more than just amusement. In the real world, they are already starting to prove their usefulness. Need a meticulously crafted travel itinerary? They've got you covered. Struggling with that daunting school essay? Let them lend a helping hand. Seeking a snippet of computer code? They'll provide it.

Yet, a word of caution: these extraordinary language models, known as LLMs, may occasionally stumble upon a factual error or indulge in what their creators amusingly term "hallucinations." So, while they are indeed capable of wondrous feats, it's wise to approach their output with a discerning eye.

What makes these LLMs so remarkable is their versatility. They possess the uncanny ability to generate an astonishing array of text, catering to a multitude of needs and whims. They can handle nearly any type of natural language input, which makes them particularly handy for users. But it doesn't stop there. They also thrive on human interaction, continually learning and adapting from our inputs.

They perform best if the user gives the machine his knowledge about the task, and the research he has done, and if the user clearly defines what he wants as his output. The key lies in allowing users to contribute their own input to the LLM.

Going Beyond LLMs

Designing interactions that empower users to contribute input is more challenging when we go outside of LLMs. Most inputs to machine learning models are fixed, constrained to a set of features or dimensions.

But by creating intuitive interfaces, we can enable user input that actually sets the stage for AI's operations. We can enhance the effectiveness of AI systems and enriche the user experience at the same time. By embracing the user's input and allowing them to shape the AI's trajectory, we unlock a world of untapped potential, where human intelligence and machine capabilities converge to drive groundbreaking outcomes.

1. Natural Language Input: Enable users to interact with ML algorithms using natural language inputs. By allowing users to express their requests or provide context in their own words, the algorithms can better understand their intentions and incorporate that information into their decision-making process. We can use them as inputs for LLMs, which have been adopted for Recommendation Systems and classification, but we can also use the input and embed it and train the deep learning algorithm on the different embeddings.

2. Contextual Prompts: Present users with prompts or questions that encourage them to provide additional context or specific inputs. For example, in image recognition tasks, the system could ask users to annotate or label certain parts of an image to refine the model's understanding.

3. Interactive Visualization: Provide interactive visualizations that allow users to explore and manipulate data inputs. By interacting directly with the data, users can gain insights, identify patterns, and refine their inputs to improve the model's outcomes. Users can drag to increase a variable and see its actual effects on the output.

4. Feature Selection: Provide users with the ability to select or prioritize specific features that they consider important or relevant. This can be particularly useful in regression models where users can highlight the variables they believe should have more influence on the predictions. We can easily fit different models that are custom to the user and see whether they actually achieve a worse, better or equal performance compared to the baseline model that we have fitted during training.

5. Active Learning: Incorporate active learning techniques where the ML algorithm actively seeks input from the user to clarify ambiguous or uncertain cases. This iterative process allows the algorithm to acquire additional labeled data and improve its performance over time.

6. User-Driven Constraints: Allow users to set constraints or define boundaries within which the ML algorithm operates. This could involve specifying limits, thresholds, or rules that guide the system's decision-making process according to the user's preferences. This can be in the form of filters in recommendation algorithms that restrict the recommendation space to certain genres or categories.

So, what lies on the horizon? Picture a world where AI systems dance gracefully with human ingenuity, orchestrating a symphony of innovation. It's a tango of intelligence, where each partner brings their unique moves to the dance floor. We should focus on what unique things the user can bring to the table and what unique capabilities our AI systems can contribute. It's a tag team match where human wit and machine capabilities join forces.

Thank you for reading AI Adoption. This post is public so feel free to share it.