Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Predictable Artificial Intelligence

Lexin Zhou, Pablo A. Moreno-Casares, Fernando Martínez-Plumed, John Burden, Ryan Burnell, Lucy Cheke, Cèsar Ferri, Alexandru Marcoci, Behzad Mehrbakhsh, Yael Moros-Daval, Seán Ó hÉigeartaigh, Danaja Rutar, Wout Schellaert, Konstantinos Voudouris, José Hernández-Orallo

TL;DR

It is argued that achieving predictability is crucial for fostering trust, liability, control, alignment and safety of AI ecosystems, and thus should be prioritised over performance.

Abstract

We introduce the fundamental ideas and challenges of Predictable AI, a nascent research area that explores the ways in which we can anticipate key validity indicators (e.g., performance, safety) of present and future AI ecosystems. We argue that achieving predictability is crucial for fostering trust, liability, control, alignment and safety of AI ecosystems, and thus should be prioritised over performance. We formally characterise predictability, explore its most relevant components, illustrate what can be predicted, describe alternative candidates for predictors, as well as the trade-offs between maximising validity and predictability. To illustrate these concepts, we bring an array of illustrative examples covering diverse ecosystem configurations. Predictable AI is related to other areas of technical and non-technical AI research, but have distinctive questions, hypotheses, techniques and challenges. This paper aims to elucidate them, calls for identifying paths towards a landscape of predictably valid AI systems and outlines the potential impact of this emergent field.

Predictable Artificial Intelligence

TL;DR

It is argued that achieving predictability is crucial for fostering trust, liability, control, alignment and safety of AI ecosystems, and thus should be prioritised over performance.

Abstract

We introduce the fundamental ideas and challenges of Predictable AI, a nascent research area that explores the ways in which we can anticipate key validity indicators (e.g., performance, safety) of present and future AI ecosystems. We argue that achieving predictability is crucial for fostering trust, liability, control, alignment and safety of AI ecosystems, and thus should be prioritised over performance. We formally characterise predictability, explore its most relevant components, illustrate what can be predicted, describe alternative candidates for predictors, as well as the trade-offs between maximising validity and predictability. To illustrate these concepts, we bring an array of illustrative examples covering diverse ecosystem configurations. Predictable AI is related to other areas of technical and non-technical AI research, but have distinctive questions, hypotheses, techniques and challenges. This paper aims to elucidate them, calls for identifying paths towards a landscape of predictably valid AI systems and outlines the potential impact of this emergent field.
Paper Structure (11 sections, 4 equations, 3 figures, 5 tables)

This paper contains 11 sections, 4 equations, 3 figures, 5 tables.

Table of Contents

  1. What is Predictable AI?
  2. The centrality and importance of Predictable AI
  3. AI ecosystems and predictability
  4. Predictability
  5. What can be predicted and where to operate?
  6. Framing predictability
  7. Who predicts and how?
  8. Scenarios
  9. The trade-offs
  10. Challenges and opportunities
  11. Impact and vision

Figures (3)

  • Figure 1: A driving scenario where six AI systems (A,B,C,D,E,F) control self-driving cars, with all systems having the same expected validity (62.5%); the grids of colour green, yellow and red represent fully valid, partially valid, and invalid cases, respectively. The distributions of validity for the six systems, $p_A, p_B, p_C, p_D, p_E, p_F$, differ across windingness and fogginess. Which one is the most predictable and hence better? Illustration adapted from burnell2022not.
  • Figure 2: Performance of a selection of GPT and LLaMA models over human difficulty on the ReliabilityBench benchmark zhou2024llmrel. The values are split by correct, avoidant and incorrect results. The $x$-axis is split into 30 equal-sized bins, whose ranges must be taken as indicative of different distributions of perceived human difficulty across benchmarks.
  • Figure 3: Scaling laws of neural models kaplan2020scaling. The test loss is predictable from the compute used during training, the training dataset size and the number of parameters of the model.