In AI research, as in life, sometimes the most effective path to a desired outcome isn't a direct one. This is the essence of obliquity, a concept popularized by John Kay in his book "Obliquity: Why Our Goals Are Best Achieved Indirectly." It suggests that when faced with complex, nonlinear problems, pursuing our goals head-on can often be counterproductive, while approaching them indirectly, with a broader perspective and focusing on related activities, can yield surprisingly successful results. Obliquity arises from the inherent complexity and unpredictability of many systems. Direct, linear approaches often fail because they:
Oversimplify the problem: Reducing a complex problem to its most basic components can strip it of crucial context and dependencies.
Create unintended consequences: Actions taken with a direct goal in mind can trigger unforeseen repercussions that negate the initial progress.
Fall victim to emergent behavior: Complex systems often exhibit emergent behavior, where the whole is more than the sum of its parts. Direct interventions can disrupt these delicate balances.
In contrast, an oblique approach:
Embraces complexity: Acknowledges the intricate interconnectedness of the system and focuses on understanding the underlying mechanisms.
Exploits feedback loops: Uses indirect interventions and observes the resulting changes, iteratively refining the approach.
Encourages exploration: Explores a wider range of possibilities, fostering creativity and serendipitous discoveries.
This article will explore how the principle of obliquity can be applied in various areas of AI research, illustrating its potential through concrete examples.
Reinforcement Learning (RL): Rewards and Intrinsic Motivation
Direct Approach: A common approach in RL is to define a precise reward function that directly reflects the desired behavior. For example, if we want a robot to navigate a maze, we might reward it for each step closer to the goal and penalize it for collisions.
Oblique Approach: Sometimes, the direct reward approach can lead to reward hacking, where the agent learns to exploit the reward function in unintended ways, achieving the letter of the goal but not the spirit. Consider a simulated world where an agent is rewarded for collecting coins. A direct reward might lead the agent to simply spin in circles collecting coins from a single respawn point, rather than exploring the world as intended.
Instead, an oblique approach focuses on intrinsic motivation. This involves rewarding the agent for curiosity, exploration, and learning novel skills, even if these activities don't immediately lead to the main goal. Examples include:
Curiosity-driven learning: Rewarding the agent for visiting new states or encountering unexpected outcomes.
Information gain: Rewarding the agent for reducing its uncertainty about the environment.
Self-supervised learning: Training the agent to predict its own actions and sensations, fostering a deeper understanding of the world.
By prioritizing exploration and understanding over immediate reward, these oblique methods can lead to more robust and generalizable agents that are better equipped to solve complex tasks and adapt to changing environments.
Example: The "Go Explore" algorithm exemplifies obliquity. Instead of focusing on directly winning the game of Go, the algorithm first encourages the agent to explore the entire state space by rewarding it for reaching novel states. This allows the agent to discover advantageous game states that it might not have encountered through traditional reward-based learning, ultimately leading to superior performance.
Natural Language Processing (NLP): Language Models and Understanding
Direct Approach: A direct approach to building a language model might focus solely on predicting the next word in a sequence, optimizing for perplexity (a measure of how well the model predicts the next word).
Oblique Approach: While perplexity is a useful metric, it doesn't necessarily equate to true language understanding. A model can be highly proficient at predicting the next word without possessing any meaningful comprehension of the underlying semantics or pragmatics.
Oblique approaches in NLP focus on:
Pre-training on diverse tasks: Training language models on a variety of tasks, such as masked language modeling, next sentence prediction, and question answering, forces them to learn a broader range of linguistic skills and knowledge. This is the foundation of models like BERT, GPT, and their successors.
Fine-tuning on downstream tasks: After pre-training, the model is fine-tuned on specific downstream tasks, such as text classification, sentiment analysis, and machine translation. This allows the model to apply its general language understanding skills to solve real-world problems.
Adversarial training: Training models to withstand adversarial examples (inputs designed to fool the model) forces them to learn more robust and generalizable representations.
These methods indirectly improve language understanding by exposing the model to a richer set of data and tasks, encouraging it to learn deeper representations of meaning and context.
Example: GPT-series, while trained primarily on predicting the next word, demonstrates impressive capabilities in a wide range of tasks, including writing code, summarizing text, and even generating creative content. This emergent behavior arises from the model's vast scale and the diverse dataset it was trained on, illustrating how an indirect approach to language modeling can lead to unexpected and powerful results.
Computer Vision: Object Recognition and Scene Understanding
Direct Approach: Directly training a model to identify specific objects in images can lead to overfitting on the training data and poor generalization to new images.
Oblique Approach: Oblique approaches in computer vision focus on:
Self-supervised learning: Training models to learn representations from unlabeled data, such as by predicting the relative position of patches in an image or by solving jigsaw puzzles. This allows the model to learn general features of images without requiring explicit object labels.
Generative models: Training models to generate images from noise allows them to learn a deeper understanding of the underlying structure and variations in the data. This can then be used to improve object recognition and scene understanding.
Attention mechanisms: Allowing the model to focus on the most relevant parts of an image, rather than processing the entire image uniformly, can improve accuracy and efficiency.
By learning general representations and focusing on relevant features, these oblique methods can improve the robustness and generalizability of computer vision models.
Example: Contrastive Learning, a self-supervised technique, exemplifies obliquity in computer vision. Models are trained to identify which images are similar based on their content, without explicit labels. This encourages them to learn robust visual representations that are useful for a variety of downstream tasks, such as image classification and object detection.
Robotics: Task Planning and Motion Control
Direct Approach: Traditionally, robotics relies heavily on pre-programmed behaviors and precisely defined control algorithms. For complex tasks, this often leads to rigid and brittle systems that struggle to adapt to unexpected situations.
Oblique Approach: Oblique approaches in robotics focus on:
Learning from demonstration: Instead of explicitly programming robot behaviors, learning from demonstration allows robots to learn from human examples, mimicking their actions and adapting them to new situations.
Embodied AI: Embodied AI focuses on developing AI agents that interact with the physical world through sensors and actuators. This allows them to learn from experience and adapt to their environment in real-time.
Compositional robotics: Developing robots that can decompose complex tasks into simpler sub-tasks and combine them in novel ways.
By learning from experience and interacting with the environment, these oblique methods can lead to more flexible and adaptable robots that are better equipped to handle complex and unpredictable tasks.
Example: A robot trained to grasp objects through reinforcement learning might struggle to generalize to new objects with different shapes and textures. However, if the robot is trained to explore its environment and interact with objects in a variety of ways (e.g., pushing, pulling, tapping), it may develop a more general understanding of object properties and be better able to adapt to new grasping scenarios.
Challenges and Considerations:
While obliquity can be a powerful approach, it's important to acknowledge its challenges:
Defining the Indirect Path: Identifying the right indirect approach can be challenging and requires a deep understanding of the underlying system.
Evaluating Progress: Measuring progress towards the indirect goal can be difficult, as the connection to the ultimate goal may not be immediately apparent.
Computational Cost: Oblique methods often require more data and computational resources than direct methods.
The principle of obliquity offers a valuable perspective for tackling complex problems in AI research. By embracing complexity, focusing on related activities, and exploring a wider range of possibilities, we can unlock new insights and develop more robust and adaptable AI systems. While direct approaches have their place, especially for well-defined problems, embracing obliquity allows us to navigate the uncharted territories of AI research and uncover solutions that might otherwise remain hidden in plain sight. By looking sideways, we can often find the most direct path to success.
Comments