Perceptual Evaluations: Evals for Aesthetics — Diego Rodriguez, Krea.ai

Key Concepts

• Generative AI Evaluation: Assessing AI-generated content considering human perception, aesthetics, and opinion.
• Information Theory & Communication: Applying principles of information theory, particularly compression, to understand AI limitations and evaluation.
• Perceptual Compression (JPEG, MP3, MP4): Exploiting human sensory limitations (brightness vs. color, audible frequencies) to reduce data size.
• FID Score: A common metric for evaluating generative models, but sensitive to artifacts like JPEG compression.
• Relativity of Metrics: Recognizing that metrics can be misleading and may not capture artistic intent or human understanding.
• Traffic (Unforeseen Consequences): Identifying the "traffic" or unexpected challenges arising from AI advancements, beyond the core technology.
• Translation & Communication: The impact of AI-powered translation on global communication and understanding of diverse perspectives.
• Subjective Evaluation: Incorporating individual opinions and preferences into AI evaluation metrics.
• Perceptually Aware Metrics: Metrics that account for how humans perceive data, including artifacts and subjective qualities.

Detailed Summary

Introduction: The Problem of Human Perception in AI Evaluation

Diego Rodriguez, co-founder of Korea, an AI startup, introduces the core problem: current AI models struggle to understand human perception and aesthetics. He illustrates this with an example of an AI model (GPT-3) failing to recognize the unnatural appearance of an AI-generated hand image. This highlights the disconnect between AI's ability to process data and its understanding of human preferences and preconceived notions. The talk focuses on improving AI evaluation by considering human perception and asking critical questions about current practices.

Information Theory and Compression: Parallels to AI Limitations

Drawing a parallel to Claude Shannon's information theory, Rodriguez emphasizes the importance of communication. He connects classic information theory to modern AI, particularly variational autoencoders and neural networks. He then discusses perceptual compression techniques like JPEG, MP3, and MP4, which exploit human sensory limitations to reduce data size. JPEG, for example, leverages the human eye's lower sensitivity to color compared to brightness. This leads to the question: are AI models being limited by the compressed data they are trained on, inheriting our perceptual flaws?

Example: JPEG compression downsamples color channels because humans are less sensitive to color variations than brightness variations. This reduces file size without significantly impacting perceived image quality.

The Flaws of Current Metrics: FID Score and JPEG Artifacts

Rodriguez critiques the use of standard metrics like FID (Fréchet Inception Distance) score. He points out that FID scores are highly sensitive to JPEG artifacts, even when the perceptual difference is negligible. This raises concerns about relying on such metrics to evaluate generative AI models. He argues that current evaluations often focus on easily measurable aspects (object recognition, color accuracy) while neglecting subjective qualities and artistic intent.

Example: An image with JPEG artifacts can receive a low FID score, even if it is perceptually similar to the original, uncompressed image.

The "Traffic" Problem: Beyond Core Technology

Referring to a conversation with Changloo from Midjourney, Rodriguez introduces the concept of "traffic." Predicting the car's invention when horses were the norm was relatively straightforward. The real challenge lies in predicting the unforeseen consequences or "traffic" that arise from technological advancements. He urges the audience to consider the broader implications and potential pitfalls of AI development.

Quote: "Predicting the car back when everything was horses... what's hard to predict? Traffic." - Changloo, Midjourney (paraphrased)

Translation and Communication: Bridging the Language Barrier

Rodriguez discusses the myth of the Tower of Babel and its relevance to the current state of AI development. He notes that AI-powered translation is breaking down language barriers, enabling communication and understanding across diverse perspectives. This has implications for subjective evaluation, as it allows for better expression and understanding of opinions.

Example: Rodriguez uses AI translation to provide customer support in Japanese, despite not speaking the language fluently.

Evolving Evaluations: Incorporating Subjectivity and Perception

The core argument is that AI evaluations need to evolve to incorporate subjectivity and human perception. Rodriguez suggests training AI models to understand human preferences and perceptual biases. This involves considering factors like visual learning styles and the nature of the training data (e.g., the prevalence of JPEG artifacts). He advocates for developing perceptually aware metrics that align with human judgment.

Conclusion: The Future of AI Evaluation

The talk concludes with a call to action to rethink AI evaluation methods. It emphasizes the need to move beyond simple, easily measurable metrics and incorporate human perception, aesthetics, and subjective opinions. This requires developing new metrics and training AI models to understand the nuances of human judgment.

Key Takeaways

• Current AI evaluation methods often fail to capture human perception and aesthetics.
• Perceptual compression techniques highlight the limitations of human senses and the potential for AI to inherit these limitations.
• Standard metrics like FID score can be misleading due to their sensitivity to artifacts.
• It is crucial to consider the "traffic" or unforeseen consequences of AI advancements.
• AI-powered translation is breaking down language barriers and enabling better communication.
• AI evaluations need to evolve to incorporate subjectivity and human perception.
• Developing perceptually aware metrics is essential for creating AI models that align with human values and preferences.