The paper presents new features and performance improvements in Theano, a linear algebra compiler for symbolic mathematical computations. Key highlights include: 1. **Symbolic Mathematical Expressions**: Theano supports symbolic differentiation and optimization of mathematical expressions, enabling efficient gradient computation and model prototyping. 2. **Fast Execution**: Theano leverages NumPy and SciPy for easy implementation of mathematical operations, and uses CUDA for GPU acceleration. 3. **Stability and Community Support**: Theano has a robust test suite and a growing user community, ensuring code quality and correctness. 4. **New Features**: - **Scan Operator**: Facilitates symbolic loops and recurrent models, improving efficiency and flexibility. - **R-Operator**: Supports Hessian-Free optimization, enhancing second-order methods. - **Lazy Evaluation (CVM)**: Enables lazy evaluation of operations, reducing overhead and improving performance. - **More C Implementations**: Additional C implementations for existing operations to avoid Python context switches. - **Better Sparse Matrix Support**: Enhanced support for sparse matrices and their derivatives. - **CPU Parallelism**: Support for multi-core CPU parallelization using OpenMP. - **Asynchronous GPU Execution**: Asynchronous function calls on GPUs, allowing concurrent CPU operations. 5. **Benchmarks**: - **Neural Networks**: Theano outperforms Torch7 on CPU and GPU for various neural network tasks, including multi-layer perceptrons and deep neural networks. - **Recurrent Neural Networks**: Theano performs well on recurrent neural networks, showing competitive performance with RNNLM. The paper concludes by highlighting Theano's strengths and its ability to be faster than competing software in most cases, making it a powerful tool for machine learning software development.The paper presents new features and performance improvements in Theano, a linear algebra compiler for symbolic mathematical computations. Key highlights include: 1. **Symbolic Mathematical Expressions**: Theano supports symbolic differentiation and optimization of mathematical expressions, enabling efficient gradient computation and model prototyping. 2. **Fast Execution**: Theano leverages NumPy and SciPy for easy implementation of mathematical operations, and uses CUDA for GPU acceleration. 3. **Stability and Community Support**: Theano has a robust test suite and a growing user community, ensuring code quality and correctness. 4. **New Features**: - **Scan Operator**: Facilitates symbolic loops and recurrent models, improving efficiency and flexibility. - **R-Operator**: Supports Hessian-Free optimization, enhancing second-order methods. - **Lazy Evaluation (CVM)**: Enables lazy evaluation of operations, reducing overhead and improving performance. - **More C Implementations**: Additional C implementations for existing operations to avoid Python context switches. - **Better Sparse Matrix Support**: Enhanced support for sparse matrices and their derivatives. - **CPU Parallelism**: Support for multi-core CPU parallelization using OpenMP. - **Asynchronous GPU Execution**: Asynchronous function calls on GPUs, allowing concurrent CPU operations. 5. **Benchmarks**: - **Neural Networks**: Theano outperforms Torch7 on CPU and GPU for various neural network tasks, including multi-layer perceptrons and deep neural networks. - **Recurrent Neural Networks**: Theano performs well on recurrent neural networks, showing competitive performance with RNNLM. The paper concludes by highlighting Theano's strengths and its ability to be faster than competing software in most cases, making it a powerful tool for machine learning software development.