The chapter discusses the limitations of current machine learning algorithms and proposes deep architectures as a solution to achieve artificial intelligence (AI). The authors argue that while kernel methods and shallow architectures are flexible, they are inefficient and limited in their ability to learn complex, high-dimensional functions. Shallow architectures, such as kernel machines, are shallow in structure, consisting of a single layer of template matchers followed by a linear combination of coefficients. This structure is inefficient in terms of computational resources and examples required for good generalization. The curse of dimensionality further exacerbates these limitations, especially with local kernels. Deep architectures, on the other hand, are composed of multiple layers of parameterized non-linear modules, allowing for more efficient representation of complex functions. They can handle a wide range of functions with fewer computational resources and examples compared to shallow architectures. The authors provide examples and mathematical analyses to support their claims, demonstrating that deep architectures can generalize non-locally, which is crucial for complex tasks like perception and reasoning. The chapter also discusses the trade-offs between convexity and non-convexity in optimization, suggesting that non-convex optimization can sometimes be more efficient for learning complex functions from weak prior knowledge. The authors conclude that deep architectures offer a promising avenue for scaling machine learning towards AI, emphasizing the importance of flexible prior knowledge specification, deep architectures, and efficient training methods.The chapter discusses the limitations of current machine learning algorithms and proposes deep architectures as a solution to achieve artificial intelligence (AI). The authors argue that while kernel methods and shallow architectures are flexible, they are inefficient and limited in their ability to learn complex, high-dimensional functions. Shallow architectures, such as kernel machines, are shallow in structure, consisting of a single layer of template matchers followed by a linear combination of coefficients. This structure is inefficient in terms of computational resources and examples required for good generalization. The curse of dimensionality further exacerbates these limitations, especially with local kernels. Deep architectures, on the other hand, are composed of multiple layers of parameterized non-linear modules, allowing for more efficient representation of complex functions. They can handle a wide range of functions with fewer computational resources and examples compared to shallow architectures. The authors provide examples and mathematical analyses to support their claims, demonstrating that deep architectures can generalize non-locally, which is crucial for complex tasks like perception and reasoning. The chapter also discusses the trade-offs between convexity and non-convexity in optimization, suggesting that non-convex optimization can sometimes be more efficient for learning complex functions from weak prior knowledge. The authors conclude that deep architectures offer a promising avenue for scaling machine learning towards AI, emphasizing the importance of flexible prior knowledge specification, deep architectures, and efficient training methods.