Bart Kosko argues that machines do not “think” in the traditional sense but rather approximate functions, emphasizing that they are essentially pattern recognition devices. Key points from his perspective include:
1. Machine Function Approximation: Kosko asserts that machines, including AI systems, excel at approximating complex functions, particularly those related to patterns, such as images, speech, or other signals. They convert inputs into outputs using vast computational power.
2. Exponential Computing Power: He highlights the exponential growth in computing power, driven by Moore’s Law, as the primary factor enabling machines to handle increasingly complex pattern recognition tasks.
3. Algorithms and AI: Kosko discusses popular AI algorithms, such as k-means clustering and backpropagation (neural networks), and points out that they are essentially special cases of the expectation-maximization (EM) algorithm, a standard statistical technique. He emphasizes that these algorithms involve no more “thinking” than calculating sums and picking the highest.
4. Hill-Climbing and Pattern Recognition: He likens much of machine thinking to “hill-climbing” in probability spaces, where algorithms seek optimal solutions. As computers get faster, these equilibria appear more intelligent, but they fundamentally involve calculations rather than conscious thought.
5. Algorithmic Continuity: Kosko notes that many AI algorithms have remained fundamentally unchanged since the 1960s, highlighting the continuity in machine learning methods.
6. Future of Thinking Machines: He predicts that future “thinking” machines will continue to rely on existing algorithms running on faster computers, with improvements coming from increased computational speed.
In summary, Bart Kosko emphasizes that machines excel at approximating complex functions and recognizing patterns but do not engage in true thought or consciousness. He underscores the critical role of computing power in advancing AI capabilities.